Meta’s multi-hundred million dollar comp offers and Google’s multi-billion dollar Character AI and Windsurf deals signal that we are in a crazy AI talent bubble.

The talent mania could fizzle out as the winners and losers of the AI war emerge, but it represents a new normal for the foreseeable future. If the top 1% of companies drive the majority of VC returns, why shouldn’t the same apply to talent? Our natural egalitarian bias makes this unpalatable to accept, but the 10x engineer meme doesn’t go far enough – there are clearly people that are 1,000x the baseline impact.

This inequality certainly manifests at the founder level (Founders Fund exists for a reason), but applies to employees too. Key people have driven billions of dollars in value – look at Jony Ive’s contribution to the iPhone, or Jeff Dean’s implementation of distributed systems at Google, or Andy Jassy’s incubation of AWS.

The tech industry gradually scaled capital deployment, compounding for decades to reach trillions in market cap. The impact on the labor force has been inflationary, but predictable. But in the two and a half years post-ChatGPT, AI catch-up investment has gone parabolic, initially towards GPUs and mega training runs. As some labs learned that GPUs alone don't guarantee good models, the capital cannon is shifting towards talent.

Silicon Valley built up decades of trust – a combination of social contracts and faith in the mission. But the step-up in the capital deployment is what Deleuze would call a deterritorializing force, for both companies and talent pools. It breaks down the existing rules of engagement, from the social contract of company formation, to the loyalty of labor, to the duty to sustain an already-working product, to the conflict rules that investors used to follow.

Trust can no longer be assumed as an industry baseline. The social contracts between employees, startups, and investors must be rewritten. In the age-old tension between mission and money, missionary founders must prepare themselves for the step-function increase in mercenary firepower.

Hypercapitalist AI talent wars will rewrite employment contracts and investment norms, concentrate returns, and raise the bar for mission and capital required to create great new companies.

Talent

As a thought exercise, how much should Google have paid for Deepmind? In 2014, a $400M acquisition for a pre-revenue company seemed nonsensical. But with the leverage that comes with Google scale, the DCF value could be quite high – a few percentage points in net savings from their datacenter costs could make it a 100x+ return over a decade, and that’s in a pre-LLM world! In the context of Google paying $3B for Noam, they’ve probably already earned back that investment in his help getting Gemini training runs unstuck; the deal even looks modest with a year of hindsight.

From the Big Tech point of view, if AI is a $10T+ revenue opportunity, and your research team sized scales sublinear to revenue with a cap of a few hundred researchers, is the difference between spending $5M/year/researcher and $10M/year and $20M/year enough to stop you? $10B per year in researcher comp is less than a quarter of Meta’s annual capex. No matter the odds of ultimate product-market fit, the sunk cost is too large to turn back now.

Even in 2014, the AI talent wars existed: Meta was reportedly the other bidder in the Google-Deepmind deal. But why didn’t pricing for top talent run up sooner? The confluence of compute leverage, urgency, and supply constraint means the labor share of value is higher than prior technological waves:

• Compute leverage: Large labs spent tens of billions of dollars on compute clusters, and are ramping to hundreds of billions. If the utility of compute is a function of compute × research efficiency, and well-used compute drives value across a massive revenue base, the willingness to pay for top researchers grows exponentially.
  

• Demand urgency: AI products distribute faster than software or internet products – chatbots and codegen being two early examples – so the urgency to get ahead is extreme as the market pecking order is established.
  
  We’re still in an indefinite R&D phase of AI, where research quality determines frontier product status. If the codegen market is representative of frontier pricing power, the delta in enterprise value capture for being 2 months ahead vs 2 months behind has never been larger.
  

• Supply constraint: If you think the key end markets will be decided in the next 1-2 years across the product categories that matter, there’s little time to train up new talent, and only a few hundred people currently have the skills to harness current frontier model capabilities.

No single analogy is perfect, but we can learn a lot from athletes, actors, and traders, where the best are worth 10x or 100x the average. All three categories have tremendous capital magnification – a superstar needs expensive infrastructure (compute clusters, risk systems, studio marketing, training facilities). Managing these superstars has a few unique properties:

• Discovery: For non-urgent hires, athlete scouting could translate nicely to AI labor to spot young talent before it runs up in pricing. In Hollywood, spotting early talent is similar to AI researchers in that it’s more art than science; until they have their “break” performance, actors are hard to value.
  

• Negotiation / pricing: CAA emerged in the 1970s to build up acting talent, package players together as teams, and increase the protections for stars. Like in sports and acting, researchers are already starting to use talent agents to represent themselves. For lower-priced talent, Hollywood guilds emerged to set day-rates as a collective bargaining mechanism to protect the labor.
  
  In sports, moneyball shifted talent pricing from intuition to statistics. But unlike in sports, researcher impact is hard to quantify, especially ahead of time. Even in hindsight, the corporate history gets rewritten by the victors, and it is often challenging to attribute research breakthroughs singularly.
  

• Retention: On Wall Street, losing a star trader is costly in alpha leakage, and financial institutions fight hard to lock people in with non-competes and garden leave policies. In tech, it’s harder to prevent information diffusion when researchers are in the same social circles, but the labs are becoming more locked down from an infosec perspective. Most secrets get held across many labs, but a small fraction are proprietary; and for many submarkets, marginal model edge determines pricing power.

Hypercapitalism erodes Silicon Valley’s trust culture. Industry-level trust alone no longer guarantees loyalty between companies and talent. With trade secret leakage risk and money big enough to tear teams apart, vanilla at-will employment contracts don’t protect either side.

The industry needs a SAFE equivalent for tech talent. New employment contracts must satisfy demands from both companies and talent:

• Company side: I expect companies to push for more aggressive trade secret protection, stricter NDAs (SSI-esque secrecy), more exclusivity (named-competitor noncompetes), and garden leaves to safeguard against flighty talent. I’m not an employment lawyer, and non-competes get tricky in California, but I wouldn’t be surprised to see heavier measures to approximate exclusive contracts, like deferred compensation or equity clawback agreements for defection to a named list of competitors.

• Talent side: AI researchers are one of the only super-earning groups outside of finance that doesn’t have negotiating help, and seem poised for professional representation or collective bargaining. To match public company alternatives, startups may need to offer liquidity guarantees or early IPOs. The rank and file employees also need new mechanisms to ensure that founders and execs won’t walk away without them.

We’re in the early days of labor repricing – big tech’s AI capex investments are so large that they already have the sunk cost, and the labor as a percentage of total investment is still low. Companies must re-think their recruiting and retention strategies.

Companies

The talent war is a net-consolidating force on the AI research frontier. At the research labs, big dollars for researchers makes it nearly impossible for new entrants to play. For the same reasons, it’s nearly impossible to start a new quant fund – you can’t get the same leverage out of the talent that big players can.

In the tradeoff between money and mission, the money has gone parabolic. Founders use both to magnetize top talent to their companies, but as the capital opportunity cost increases, only the strongest missions can justify the economic sacrifice that candidates make. To the credit of both OpenAI and Anthropic, money alone has not been enough for the best researchers to defect – their cult status effectively creates a multiplier on their R&D budgets.

The labs feel the talent war most directly, but all startups now require extreme resource aggregation to make AI R&D bets. When the opportunity cost of top talent is higher (both founders and engineers), it becomes harder to coordinate top talent around an early-stage bet. SSI, Thinking Machines, and Physical Intelligence all required massive funding rounds for a shot on goal. A single research hire can cost the entirety of a Series A fundraise, making AI R&D far more expensive for startups, pushing most to live above the APIs.

A startup industrial complex around the Seed → Series A → Series B progression emerged in the 2010s to support the growth of software companies. Some companies still follow this pattern successfully: Harvey, Abridge, Glean, and others. But I believe that going forward, an increasing share of startup successes will have a “fat pitch” founding story: incubations with stacked founding teams, high institutional credibility on day one, and uniquely powerful missions.

Modern successes like SpaceX, Anduril, and OpenAI could not be built as lean startups. They are too long-horizon and capital intensive to work through the traditional apparatus. The most promising tech frontiers often have high activation energy – foundation models, robotics, biology – where mega-rounds are the only way to bridge to the future. The AI capital influx means that mega-projects no longer seem outlandishly expensive. This is good for the world!

On the big tech side, the talent wars thin the playing field to companies with 1) tens of billions in net income that they can cut into, and 2) leaders with founder-like agency that will heavily sacrifice earnings for a seat at the AI table. A sharper power law will create a new “giga-cap” class, with multiple $10T companies by 2035.

Some subset of startup winners will have a similar formula to what worked in the 2010s: small, scrappy team, building iteratively until they crack product-market fit. An increasing percentage of the new winners will have large war chests and strong missions from day one. AI talent wars will be a net-consolidating force.

Subscribe now

Investors

Being a rigid seed or Series A-only investor in 2025 is anachronistic. Should you simply ignore the most important tech companies of this generation?

Investors must be more flexible than prior generations. The best companies won’t map to the predictable fundraise sequence of the past 20 years. Rapid product adoption will require investors to swallow their pride and admit misses much more quickly. For some companies investors passed on 6 months ago, the right decision is to invest today at 2-3x the valuation.

At the early stage, a new deal consideration has emerged: investors evaluate companies with the team quality constituting their downside case. Character and other “talent deals” make investors think that they can’t lose money investing in top tier research teams – it’s almost like investing in an AI researcher labor union. As long as the company exits for more than the cumulative fundraising amount, investors get paid back first. Investors then justify putting more dollars in at an earlier stage than they would otherwise.

SSI and Thinking Machines both follow this pattern. Investors don’t even need to scrutinize the exact technical approach, because the upside of AGI is infinity (1% chance of breakthrough → $10T+ company). If you believe you can’t lose money given the team quality, the upside case is almost like a “free” call option.

But if VCs assess the talent wrong – over-estimate the talent, or over-estimate the talent’s commitment to the company – they can get nuked on big slugs of capital. Even if a team gets to a technical breakthrough, capturing the value isn’t guaranteed. Research teams that achieve technical breakthroughs are not necessarily the ones that get product and sales right.

Historically, the social contract of starting a company meant the founders would see it through to an exit. But how big do the numbers need to get before that breaks? People didn’t used to leave companies they founded, especially when they’re nascent and/or highly valued, but the AI talent war is deterritorializing. This fragility enables a CEO or key execs to leave their company with minimal recourse.

Like the founder <> researcher social contract, investors also need to reconfigure the founder investor social contracts for this new world, particularly for research-heavy teams:

• Key man clauses: Like VC funds themselves often have key man clauses for key investors to protect LPs, investors need protection against extreme instances of talent departure, particularly when it involves the founders. One instantiation could be that founders taking new jobs would qualify as an M&A event, or allow investors to redeem their capital.

• Dilution: For any company participating in the AI talent war, investors need to factor in more option pool dilution than they would otherwise. The share of value captured by employees vs. the capital layer may look fundamentally different for this type of company. Some may need fewer employees to win, but for many categories, AI companies will be structurally less profitable and more dilutive than internet companies.

Only lead checkwriters of large rounds have enough leverage to command these terms. This makes it harder than ever to be a pure early-stage investor in AI research-driven companies.

As an investor, the founders you invest in need an answer to the talent war. Either they have a cult-like missionary following, or need a clear path to winning the mercenary game with higher stakes than ever.

Conclusion

In the 2010s software bull market, success in the startup world was broad, and it felt like everyone could win (anyone could start a software company, or at least join / invest in one). That’s still somewhat true; lots of people have built multi-$m ARR AI businesses in short order. The one-person unicorn concept suggests that anyone can start a big company using AI.

But in the new world, the concentration of outcomes will be different, both at the talent and company levels. Fewer companies getting more funding and revenue, fewer employees getting paid more. Only the fiercest founders and strongest missions can offset the inflection in mercenary market forces.

High earners usually avoid attention, but splashy 9-figure researcher offers draw significant public interest. There is a human bias against accepting singular winners (of talent, of companies); it doesn’t feel fair to have a few people run away with big markets. There is something uniquely unstable about a more uneven distribution of success. The French had a uniquely high Gini coefficient before the Revolution.

The M&A talent war is just beginning, raising compensation baselines and labor promiscuity. To protect against the deterritorialization, I expect new labor dynamics to emerge on both sides of the table: agents, unions, aggressive non-compete tactics. As the numbers get bigger for talent and companies, all sides need to reimagine the social contract. As the glue holding teams together, company mission matters more than ever.

The AI talent wars will rewire Silicon Valley.

Thanks to Axel Ericsson, Philip Clark, Melisa Tokmak, Joey Krug, Cat Wu, Will Manidis, Robert Windesheim, Lachy Groom, and Will Depue for their thoughts and feedback on this article.

Two seemingly contradictory but equally popular narratives about the future of foundation models have taken hold. In one future, AI centralizes: scaling laws will hold, and value accrues primarily to scaled, closed-source players. In the other future, AI decentralizes: foundation models have no moat, open-source has caught up to closed-source, and we’ll have many competing models.

Today, both narratives seem true. We have powerful closed models, and a thriving ecosystem of sacrosanct open-source models. Llama-3 recently put open-source on the map of GPT-4 class models. Meanwhile, an unusual open-source alliance has formed among developers who want handouts, academics who embrace publishing culture, libertarians who fear centralized speech control and regulatory capture, Elon who doesn’t want his nemesis to win AI, and Zuck who doesn’t want to be beholden to yet another tech platform.

As an accelerant of modern software, open-source maintains a cherished place in tech. Who can argue with free stuff, decentralized control, and free speech? But open and closed-source AI cannot both dominate in the limit: if centralizing forces hold, scale advantages will compound and leave open-source alternatives behind.

Despite recent progress and endless cheerleading, open-source AI will become a financial drain for model builders, an inferior option for developers and consumers, and a risk to national security. Closed-source models will create far more economic and consumer value over the next decade.

Model builders

Open-source software started as an act of charity – the world owes the likes of Linus Torvalds and Fabrice Bellard for endowing humanity with Linux, Git, and FFmpeg. Because free stuff is popular, open-source became a great freemium marketing strategy (think Databricks or Mistral), and sometimes a market equilibrium in itself (Android e.g. is a cheap smartphone option and reinforces Google’s search monopoly).

Companies that earned free marketing from open-source eventually succumbed to business physics: Red Hat hid CentOS behind a subscription service, ElasticSearch changed their licensing after accidentally seeding competition, and Databricks owns the IP that accelerates Apache Spark.

Unlike the charity work of open-source in the early software era, today it is subsidized by businesses with their own goals. Given Meta is the primary deep-pocketed large open-source model builder, open-source AI has become synonymous with Meta AI. So the operative question for open-source AI is, what game is Meta playing? In a recent podcast, Zuckerberg explains Meta’s open-source strategy:

1. He was burned by Apple’s closedness for the past 2 decades, and doesn’t want to suffer the same fate with the next platform shift. It’s a safer bet to commoditize your complements.

2. He likes building cool products, and cheap, performant AI enhances Facebook and Instagram. There is some call option value if AI assistants become the next platform.

3. He bought hundreds of thousands of H100s for improving social feed algorithms across products, and this seems like a good way to use the extras.

That all makes sense, and Llama has been great developer marketing for Facebook. But Zuck also suggests several times that there’s some point at which open-source AI no longer makes sense, either from a cost or safety perspective. When asked whether Meta will open-source the future $10b model, the answer is “as long as it’s helping us”. At some point, they’ll shift their focus towards profit.

Unlike other model providers, Meta is not in the business of selling model access via API. So while they'll open-source as long as it is convenient for them, developers are on their own for model improvements thereafter.

That begs the question: if Meta is only pursuing open-source insofar as it benefits themselves, what is the tipping point at which Meta stops open-sourcing their AI? Sooner than you think:

• Exponential data: Frontier models were trained on the corpus of the internet, but that data source is a commodity – model differentiation over the next decade will come from proprietary data, both via model usage and private data sources.
  
  Open-source models have no feedback loop between production usage and model training, so they foot the bill for all incremental training data, whereas closed-source models drive compounding value with data from incremental usage. If Meta differentiates their model based on their social graph or user feedback, they’ll want to capture that value via their closed products, and not share it with the world.

• Exponential capex: A lagging-edge model that requires just a few percent of Meta’s $40b in capex is easy to open-source, and nobody will ask questions. But once you reach ten billion dollars or more in capex spend for model training, shareholders will want clear ROI on that spend (the Metaverse raised some question marks at a certain scale, too).

• Diminishing returns on model quality within Meta: There is a large upfront benefit for Meta building an open-sourced AI model, even if it’s worse than the frontier closed-source counterpart. There are lots of small AI workloads (think feed algorithms, recommendations, and image generation) where Meta doesn’t want to rely on a third party provider like they had to rely on Apple.
  
  But it’s unclear whether Facebook products benefit much from models approaching AGI quality. It’s equally possible that Meta’s model improvements will be very particular to their own internal use cases. And this is where things aren’t aligned with users: if the ROI of generalized, frontier models doesn’t make sense for Meta’s products, they certainly won’t build them for the open-source community.

Zuck is not running a charity, he’s a savvy capitalist. While Meta can justify scaling capex on incremental models for their own ends, their open-source strategy will only make less sense over time.

Developers and consumers

As a developer choosing an open-source model, what do you get in terms of cost, model quality, and data security?

Cost: Open-source models have the illusion of being free. But developers bear the inference costs, which are often more expensive than comparable LLM API calls: either pay a middleman to manage GPUs and host models, or pay the direct costs of GPU depreciation, electricity, and downtime. Only large enterprise scale can amortize these fixed costs; in other categories like cloud infrastructure, even the largest F500 companies use third party cloud hosting like AWS and Azure. Unoptimized GPU spend punishes you for diseconomies of subscale (the inverse of economies of scale).

A certain type of cost-conscious enterprise or consumer will put up with weaker products and paywalls; they want pure cost optimization. But the closed-source cost curve is still coming down radically, so it’s not even clear that open-source can be cheaper in the medium term. Dot-com companies used to spend half of their budget buying server racks, until AWS fixed the cloud capex problem; closed-source model providers do the same for AI.

In capitalist America, free is never really free, so you should wonder how you’ll ultimately be monetized. This isn’t Linux where a single developer built the product as a gift to humanity; these are cash-incinerating endeavors whose only way out is to eventually monetize you. You’re probably committing to a closed-source complement in time. Every open-source company rolls out a paid tier eventually; even Android eventually monetized via Google Play and search.

Even if self-hosted open-source models are marginally cheaper above a certain breakeven point, marginal cost optimization is the wrong focus at this stage in the cycle: for most applications, capabilities are holding back adoption, not price.

Model quality: Like housing, healthcare, and education, the paid version is generally better than the free version. Even within software, the open-source winner is rarely the best product: Android is worse than iOS, OpenOffice is worse than Office or Google Docs, Godot is worse than Unity, FreeCAD is worse than SolidWorks. A corollary is that engineers focused on the best platforms make more money; they’re more likely building cutting-edge products.

Everyone is celebrating that Llama-3 is on par with GPT-4, a year later. The product quality gap between iOS and Android, or MacOS and consumer Linux, has stayed wide for a long time, because the best software creators are aligned with paying customers. When you choose closed-source models, you’re not making a point-in-time decision on model quality; you’re paying for future model improvements, where the roadmap is aligned with paying customers.

Most people focus on the last war (GPT-4) and not the next. So while open-source models are a healthy part of the ecosystem, they’re largely backwards looking. I don’t expect a capabilities plateau until the capex spend on GPUs and data reaches the tens of billions, on par with semiconductor manufacturing. Will the key open-source model builders find enough revenue to justify spending that much?

Data security: Some enterprises need the utmost data security: financial services, healthcare, legal. But I’m not sure using open-source models on-prem or via third-party cloud hosting is actually safer than using third party LLMs in the cloud; this is a legacy belief from the early internet era where an on-premise data center was the Fort Knox of data security.

As a customer, I’d trust Microsoft with healthcare data security more than my IT department’s self-managed data center. And that bridge has already been crossed: when 65% of the risk-averse Fortune 500 already uses Azure OpenAI, it makes you wonder who is dealing with data that is too sensitive for cloud-based LLMs.

National security

Even if it makes eventual economic sense for model builders to build open-source, should they? Advocates like Yann LeCun claim that open-sourced AI is safer than closed. It makes me wonder if he really believes in Meta’s AI capabilities. Any reasonable extrapolation of capabilities with more compute, data, and autonomous tool use is self-evidently dangerous.

Appealing to American security may seem overwrought, but the past five years of geopolitics has confirmed that not everyone is on the same team. Every country outside America has an interest in undermining our closed-source model providers: Europe doesn’t want the US winning yet another big tech wave, China wants free model weights to train their own frontier models, rogue states want to use unfiltered and untraceable AI to fuel their militaristic and economic interests.

AI is a technology of hegemony. Even though open-source models are lagging behind the frontier, we shouldn’t export our technological secrets to the world for free. We’ve already recognized the national security risk in other parts of the supply chain via export bans in lithography and semiconductors. When open-source model builders release the model weights, they arm our military adversaries and economic competitors. If I were the CCP or a terrorist group, I’d be generously funding the open-source AI campaign.

A common retort is that the CCP has their own frontier AI models that are comparable to the US. But they’re still behind, and if these technologies follow a Moore’s Law pattern where capabilities compound, it’s increasingly difficult to get to the frontier if you’re not already there.

Looking at semiconductors as a historical analogy, nobody has been able to catch up to the frontier of semiconductors outside of TSMC + Nvidia, despite decades of trying. Russia is decades behind the semiconductor frontier, and even China’s SMIC is half a decade behind TSMC.

Language models are in their infancy, but already instructive towards cyberattack generation, bioweapons research, and bomb assembly. Yes, Google indexes dangerous information too, but the automation of LLMs is what makes it dangerous: a rogue model doesn’t simply explain a cyberattack concept like in a Google result, but instead can write the code, test it, and deploy it at scale – short-circuiting an otherwise arduous criminal activity.

A common open-source claim is that decentralized control of the model is better than trusting a central party. But that’s a luxury US belief where the perceived downside is limited to content moderation and corporate greed. This is a much more consequential technology: LLMs will be used to attack critical infrastructure in the West, fuel disinformation campaigns during critical elections, power cyber attacks during war time, and commit fraud. These bad actors are, and will be, empowered by open-source models.

New technologies have a long history of upsetting the balance of power in the world. As a technology with militaristic ramifications, AI should not be treated lightly.

Conclusion

There are varying degrees of open-source AI optimism: many simply argue that open-source is good for humanity, others claim that building open-source models is a good business strategy, and on the extreme end, and some believe that leading model providers should be forced to open-source their technology. But all three camps are dubious: open-source models empower our adversaries, and will reach increasingly negative ROI for model builders and developers.

I admit I have an allergic reaction when many open-source advocates expose their socialist tendencies from Europe, academia, or both. While early AI undoubtedly benefited from research that was openly published before its commercial value was fully recognized, academia seems ill-suited to drive frontier research going forward. Stanford's NLP lab has only 64 GPUs, and even Fei-Fei Li admits that academia is "falling off a cliff" relative to industry.

America's tech success is subject to endless criticism from those who missed out, but we handily won the last tech wave because American capitalism aligns users and companies for the long term. The software frontier doesn't move forward in the long run without relentless and sustained obsession from correctly incentivized companies. This is even more true in the context of capital intensive foundation models.

Open-source will have a home wherever smaller, less capable, and configurable models are needed – enterprise workloads, for example – but the bulk of the value creation and capture in AI will happen using frontier capabilities. The impulse to release open-source models makes sense as a free marketing strategy and a path to commoditize your complements. But open-source model providers will lose the capital expenditure war as open-source ROI continues to decline.

For companies on the verge of spending tens of billions without a clear business model, and the developers betting on that ecosystem, I wish them the best of luck. But the winning models of the next decade should and will be closed-source.

Note: I work at Founders Fund, which has invested in OpenAI and Scale.

Thanks to Divyahans Gupta, Axel Ericsson, Harry Elliott, Phil Clark, Cat Wu, Michael Solana, Joey Krug, Mohit Agarwal, Yash Patil, Koren Gilbai, Theodor Marcu, and Gustavs Zilgalvis for their thoughts and feedback on this article.

With internet, SaaS, and mobile tailwinds in the rear-view mirror, the tech industry has become cash-rich but growth-poor. 2023 marked an unexpected return to dot-com era growth across ChatGPT, Midjourney, Character, and Copilot. Even more viral was the growth in the number of tech companies wanting a part of the newfound tailwinds.

Multiple factors drove the launch success of the 2022 AI products: magical user experiences, large-scale data acquisition and refinement, engineering team quality, viral press distribution, and scaled compute. Most of those advantages are hard, if not impossible, to replicate.

For the thousands of cash-rich and growth-poor tech companies that were caught off guard, the easiest thing to copy was the large-scale GPU purchasing, and figure out the rest later. This made Nvidia the AI kingmaker, as they could determine who was worthy of purchasing their frontier H100 GPUs at scale. Dozens of companies cropped up to resell GPUs and profit off of the ensuing shortage.

Will the exponential growth of Nvidia demand continue to outpace supply? Answering the trillion dollar shortage question is challenging – with FUD and propaganda driving the GPU conversation, it is hard to see through the noise to develop an intuition for the supply and demand dynamics at play.

A few factors amplified the GPU shortage, and understanding them should help us understand how the 2020s will unfold:

• The incentives to undermine Nvidia’s GPU monopoly have grown to 4-comma proportions. With several players each buying hundreds of millions (some purchasing billions) of frontier Nvidia chips at software-level margins, the incentives are stronger than ever to undermine their pricing power.

• There is a fat long tail of GPU purchasers. The enormous backlog of GPU demand is more than just big tech. In fact, the core hyperscalers (Google, Microsoft, Amazon) only constitute about half of Nvidia H100 sales. There is a long tail ranging from seed-stage venture-backed startups to nation-states.

• There are many FOMO-driven GPU cluster builds with unclear ROI. The long tail of demand can’t avoid the Nvidia tax by building GPUs themselves, but its longer-term demand is also much shakier than stable big tech companies. The number of foundation models with product-market fit can be measured on one hand, but the number of GPU purchasers is in the thousands.

• Nvidia operates in boom-bust cycles. Big tech companies are historical anomalies in terms of their steady compounding. Unlike the other big tech companies, Nvidia has a far more volatile growth story.

It is impossible to precisely predict the future of GPU supply and demand, but I hope to bring tech-savvy GPU-laypeople up to speed on the Nvidia craze, LLM training mega-runs, and the resulting shortage.

Why GPU scaling matters

GPUs, or graphical processing units, were first developed in the 1970s to perform graphical operations very quickly for video games. They are the closest thing we have to an antidote to the breakdown of Moore’s Law.

While CPUs can perform any workload, GPUs are optimized at the hardware-level for parallel operations. To oversimplify somewhat, that means you can perform thousands of operations at the same time, instead of waiting for operation A to complete before starting B. As it turned out, matrix multiplication is useful for more than rendering polygons on a screen – it is also used in training and running inference on neural networks.

Moore’s Law enabled the technological frontier for the past half-century. Unfortunately, the law broke down in the 2010s for CPUs. While the compute per dollar isn’t scaling quite as quickly as Moore’s Law, the absolute compute performance gains are still quite impressive:

It’s hard to know whether there’s some efficiency frontier at which we should stop investing in incremental compute performance. Laying internet fiber, for example, became a lot less profitable in the last decade.

But I’m consistently surprised by the emergent phenomenon from pushing the compute frontier. In 2010, a model like GPT-4 was not technically feasible – the largest supercomputer had computing power equivalent to ~1,000 H100s. If each order of magnitude increase in data and compute drives a jump in LLM IQ points, what might happen at the next order of magnitude?

Most exponentials in technology are really just S-curves – internet adoption, ad dollars, SaaS market sizes – but compute may be one of the few that endures as a true exponential, more similar to GDP or energy consumption (though even energy’s exponential growth eventually stalled). In 2043, it seems clear that the 2023 market for compute will look small in the rear-view mirror.

When I was an undergrad at Stanford, few people paid attention to electrical engineering – it was perceived as a solved problem, and instead everyone moved up the stack to software. But ignoring the hardware layer was a mistake – we have a duty to understand computing hardware as one of the closest things to an enduring exponential that we have.

The Nvidia empire

Many big tech players have plans to wean themselves off of Nvidia, but If you want to compete at the frontier of large-scale foundation models today, Nvidia is the only practical choice.

My friend Ben Gilbert did a great series on Nvidia’s corporate history here, but in short: the company was founded in 1993 to bring 3D graphics to gaming, and released its first GPU in 1999. They invented CUDA in 2006, which allows for parallel processing across multiple GPUs, setting the stage for the development of GPU clusters used to train larger workloads.

They have a monopoly on supercomputers, powering 74% of the top 500 supercomputers and nearly all large-scale AI training clusters. Startups pitch their relationship with the company as a core differentiator (e.g. being first in line to buy chips). While Nvidia started as a consumer GPU company, the heart of Nvidia’s banner 2023 is the data center business.

Data Center business

Nvidia’s data center business sells GPUs for AI + high-performance computing workloads. It hit a $17b revenue run rate (56% of revenue) in Q1, and more than doubled (!) to a $41b run rate in Q2 (76% of revenue). Wall Street analysts are anticipating another banner year in 2024 (FY25), with $60-100b in data center revenue.

Nvidia lineup overview

The devices below are Nvidia’s high-performance computing line, intended for large-scale workloads and often assembled in clusters (many GPUs networked together into one supercomputer):

As of Q2, Nvidia’s data center revenue is roughly half A100s and half H100s, but shifting rapidly to H100s as A100 production winds down. For training large-scale LLMs (think GPT-4 level), H100s are the best performers in terms of both speed (time to train) and cost (i.e. flops per dollar). For smaller language models and image models (think GPT-3 or Stable Diffusion), A100s (and even the cheaper A10s) are a cost-effective GPU for both training and inference.

To train a model to GPT-3 level performance, the compute required is relatively low – this explains why there were so many demos within just a few months of ChatGPT’s launch that approximated GPT-3 performance. There are dozens of clusters with 1,000+ A100s.

For GPT-4 level performance, the bar is significantly higher. Only a handful of companies including Microsoft, Amazon, Google, and Facebook have access to tens of thousands of frontier GPUs in a single cluster, not to mention the hurdles around high-quality data acquisition and training techniques. Scaling data acquisition and compute is hard and expensive – this is why nobody has surpassed general GPT-4 performance, even 8 months after launch.

This chart should give you a sense for the scale of compute required to train various models:

The GPU demand curve: growth or glut?

In the summer of 2023, it became clear that there was a run on Nvidia GPUs. As of summer 2023, H100s were sold out until Q1 2024.

There is an unstoppable force of GPU efficiency gains and supply chain unlocks against an immovable wall of $100m+ LLM training runs. It is impossible to know when the GPU shortage ends, but we can triangulate the duration of the shortage by better understanding the current demand dynamics and where they’re heading.

Present day: who exactly is buying GPUs?

There are a handful of downstream players buying the Nvidia GPUs: the hyperscalers (Google, Microsoft, Amazon), the large foundation models, and a surprisingly large number of mid-sized AI teams at startups and enterprises.

The big tech companies that need large-scale GPU clusters for internal purposes are best positioned to train their own LLMs, as they already have the ML expertise and internal demand for large H100 clusters. The core hyperscalers (AWS, Azure, GCP) are each spending $20b+ in capex per year, a figure still scaling by 25%+ annually. The capex race is causing smaller cloud providers to play catchup: Oracle has doubled its capex twice, from $2.1b in 2021 to $4.5b in 2022 to $8.7b in 2023.

The hyperscalers don’t report exact GPU spend, but industry insiders estimate that each player spends ~20-35% of data center capex on GPU build-outs (GPU purchase orders, surrounding data centers / power supplies, etc.), with Microsoft being the most aggressive. The percentage of revenue that GPUs are driving for the cloud services is still small, but growing quickly. Rough estimates are in the 5-15% range, with Azure on the high end.

The hyperscalers each have 150k+ A100s, with individual clusters up to ~25k units. But they’ve largely paused on A100 purchases in favor of the next-gen H100. The largest H100 clusters are 20-30,000 units (Microsoft, Meta, Amazon, and Google), and are used for a combination of internal products, external rental via Azure / AWS / GCP, and their respective AI partners (OpenAI, Inflection, Anthropic). There’s a mid-tail of purchasers with a few thousand A100s and H100s, like Stability, Tesla, Lambda Labs, Hugging Face, Aleph Alpha, and Andromeda.

H100s serve the LLM training use case uniquely well given their high bandwidth and memory. But H100 performance is overkill for seemingly adjacent applications like stable diffusion (image generation models), smaller model training (even GPT-3 scale), and LLM inference. It is unclear whether the other use cases like stable diffusion and lagging-edge LLM inference will scale up their memory and bandwidth needs into requiring H100s.

Medium-term outlook: expect volatility

Nvidia plans to ramp its production to 2m+ H100s next year (some say as high as 3-4 million). Can the demand match it? To make an educated guess on what 2024 looks like, we can look at the demand across the two groups of purchasers: the hyperscalers and the long tail.

Hyperscalers

The hyperscalers are receiving tens of thousands of H100s each this year – anywhere from 20k on the low end, to 100k+ on the high end. Microsoft is one of the largest purchasers of H100s (if not the largest), and while Google wants to support its internal TPU efforts, is also ramping H100 purchases to support their 26,000 H100 A3 cluster.

In their Q2 earnings report, Nvidia disclosed that one cloud service provider accounted for 29% of quarterly data center revenue, totaling a $3.9b purchase (!) in the first 6 months of 2023, equating to ~130k H100s. I suspect this is Microsoft purchasing ahead of demand to support upcoming Office Copilot launches, OpenAI compute, and Azure services. They even reserved excess capacity via CoreWeave and Oracle – a great move to flex up GPU capacity if Office Copilot takes off, while minimizing risk of overspending on GPU capex if demand comes in light.

Hyperscalers constitute roughly 50% of all H100 demand today. And of those hyperscaler purchases, a large fraction is to support upcoming product launches like Office Copilot and Google Gemini. The attach rate for big tech’s LLM-enabled products will be the earliest indicator for what the medium-term Nvidia H100 demand looks like. 

This set of products will undoubtedly see immense adoption in absolute dollar terms, but I suspect won’t be enough to justify a much larger H100 investment until these companies work out the product kinks, vertical-specific use cases, and sales motions.

As one example: GitHub Copilot is one of the fastest-growing AI products at $100m+ in ARR after a year, but relatively small compared to 1) all VSCode users (14m users implies sub-5% Copilot penetration), and 2) Microsoft’s overall AI investment in the tens of billions. For the GPU investment to pay off, they’ll need the broader Office Copilot offering to increase Office’s $50b revenue base by at least 10%, or $5b+ in revenue.

The long tail

If big tech is only part of the H100 demand story, who else is buying in volume? Outside of the hyperscalers, there is a gigantic long tail of buyers. Even at the big tech companies, a large portion of the H100s (~30-50%) will be utilized by their customers as part of the cloud offerings (AWS, GCP, Azure). This means that ~⅔ of total H100 utilization is the long tail.

Seemingly everyone ordered anywhere from a few hundred to a few thousand H100s – Cloudflare, Palantir, HP, Voltage Park, Cohere, and even nation-states like Saudi Arabia and the UAE. And those that haven’t bought chips are spending aggressively on reserving GPU-time from the GPU landlords.

For those that are training or fine-tuning models on their own GPUs, cluster utilization is often low. ChatGPT was anomalous in that the product-market fit came essentially as soon as the product was launched. This means that internal demand for inference-focused GPUs ramped up quickly, keeping GPU utilization high.

Most companies training foundation models won’t achieve product-market fit as quickly – either the model isn’t good enough, or the product isn’t good enough, or the go-to-market isn’t good enough. This could create an internal GPU demand curve could look very roughly like this:

If you multiply that across the entire industry, you could see a large temporary gap in GPU demand for the long tail. The medium-term question is whether the inference ramp-up of the winners happens before the low-ROI training runs die out.

Cyclicality of demand

One challenge in forecasting GPU demand is that each release is highly cyclical and front-weighted – just 3 years after the A100 launch in 2020, for example, Nvidia is ramping down production.

While Nvidia’s GPUs were primarily used for gaming through the early 2010s, the 2017 and 2020-era crypto booms massively accelerated their growth. And with nearly perfect timing, the AI boom kicked in right as crypto started cooling off:

The AI-driven demand should be far more durable than the crypto mining boosts, given the higher quality of the customer base and ROI that should flow directly into productivity benefits. But because GPUs are typically bought as irregular capital expenditures, the semiconductor industry has historically been a boom-and-bust business.

Despite seemingly perfect exponential growth in the rearview mirror, Nvidia’s unpredictable growth spurts create one of the most volatile revenue bases of the big tech companies:

Geographical scaling

20% of the demand is Chinese tech: given bans on high-performance GPU sales to China, many Chinese tech companies have been stockpiling an unusually high number of chips anticipating increased US export controls.

Chinese tech giants Baidu, ByteDance, Tencent, and Alibaba are spending $1b on chips (primarily A800s, a watered-down A100 per US-CN AI regulations) that will be delivered this year, and $4b on chips for 2024. With China at 20% of Nvidia’s revenue (and potentially higher if you consider the potential pass-through from Middle Eastern GPU purchases), increased export controls could create a meaningful headwind over the medium term.

Subscribe now

Long-term outlook: does the ROI come?

Over a multi-decade time horizon, we’ll certainly scale global compute exponentially, so it is easy to make the long-term bull case on GPUs. Trying to predict 5-year demand is much harder. Understanding the key levers – model size, model quantity, model ROI, and inference – will help us make some guesses as to what direction the aggregate demand is heading.

Scaling the number and size of models

To forecast long-term training GPU demand, you need some perspective on 1) how big foundation models get, and 2) how many models will win.

On #1, if large-scale models like OpenAI, Bard, and Anthropic determine that scaling laws hold (that is, more compute → more intelligence), they could easily increase their frontier GPU demand by an order of magnitude in the coming years, and others would follow suit. On the flip side, many B2B AI products are minimizing underlying model sizes by fine-tuning domain-specific, constrained-output models – these are far less GPU intensive.

Regarding #2 on the consumer side, early indicators suggest there will be a power law distribution of model success. Of the top 50 consumer AI apps, ChatGPT is 60% of all web traffic (an imperfect measure, to be sure), and running on less than 5% of global H100 capacity:

Regarding #2 on the enterprise side, the API market is similarly lopsided, with a short list of runners-up like Llama-2, Claude, and Cohere. More infra providers will come online, but like cloud infrastructure providers in the 2010s, you only need a small handful of providers to cover 80% of the market needs.

This leads to a puzzling scenario: seemingly infinite GPU demand from thousands of tech companies, but only a few players seeing ROI. Industry estimates suggest that the Microsoft cluster for OpenAI’s GPT-4 was around 20-25k H100s, and there are 3+ million H100s coming online in the next 18 months. How many GPT-4 level models do we really need?

I understand that there’s a believable optimistic case – OpenAI could scale GPT-5 to train on much larger clusters, others would follow suit, and there could be tens of thousands of companies fine-tuning models. But the hyperscalers will likely need some time to digest their large H100 orders from 2023. Demand into the millions of H100s per year is dependent on a many-models world that I’m not sure is sustainable.

The ROI gap

By the end of 2024, Nvidia will have sold $100-120b in H100s. The key AI-powered applications (OpenAI, Midjourney, Copilot, and a few others) sum to less than $5b in revenue today. When accounting for the ancillary costs like energy and labor, this means that AI application revenue needs to grow at least 20-50x to justify the capital investment, via home-run startups and $10b+ smash hit LLM products across the big tech players. As David Cahn put it clearly: there is a $100b+ hole to be filled for each year of capex at current spend levels.

The first cohort of ChatGPT reactionaries should demonstrate their ROI over the next few quarters. The returns will be highly variable – while the compute spend is surprisingly widely distributed, the revenue and user growth is not:

The early results from the second wave of fast-follower foundation models are already not showing the same ROI as the first – in spite of dozens of foundation model product launches in 2023, none have outshined the first crop of products that launched in 2022. Without the novelty X-factor and free global PR that made the early winners work, I suspect that the vast majority of the reactionary foundation models will be in for a world of hurt.

Inference

Today, there are relatively few scaled AI applications – it takes time to build out interaction paradigms that are sufficiently accurate, reliable, and trusted. Even Microsoft’s relatively productionized GPU clusters are running primarily training workloads, with the minority on inference. Most big tech companies estimate that by 2025, their GPU workloads will flip to 30% training and 70% inference (of course, this could underestimate the training ramp-up from some of the major AI labs).

The demand for inference could dramatically exceed the available GPU supply on short notice, though. For example, Midjourney has ~10,000+ GPUs for inference, and ~1.5 million active users. If they reach 30 million active users at the same efficiency, they could single-handedly demand hundreds of thousands of GPUs. There are dozens of software companies with millions of users that are releasing LLM-powered products in the coming months. If a fraction of those cross-sell effectively, inference demand could easily soak up all available GPU capacity.

The GPU market has been heavily weighted towards training, where Nvidia shines, and not inference. For inference, which is an unusually big expense item relative to the “zero marginal cost” world of 2010s software, cost structure really matters.

The H100 is often overkill for inference-style workloads, given the GPU is idle for the time spent transmitting to the CPU, which is why Nvidia created the lower-end L4 and A10 devices. Inference is where other companies have a fighting chance against Nvidia: Google’s v5e TPU is already a popular inference option (but requires GCP lock-in), and AMD’s MI300 is already proving it is well-suited for inference as a CPU-GPU hybrid.

As inference ramps up where price performance dominates, AMD has a shot at becoming a key provider by undercutting Nvidia. AMD’s AI revenue base is small relative to Nvidia, but enough to be a credible alternative, with $2b in AI chip sales anticipated for 2024. This should ensure they are a must-evaluate in any large-scale GPU purchase. Given the product requirements differ for training and inference, we could see increased GPU market segmentation.

The supply side: Nvidia’s monopoly on GPUs

Nvidia is tripling its output of H100s to ship 2m+ units in 2024, bottlenecked primarily by TSMC. You can read a detailed analysis of the upstream component supply chain bottlenecks from Dylan Patel’s SemiAnalysis here.

Nvidia has monopoly market share over GPU supply (80%+ of revenue), which was a somewhat reasonable equilibrium when there was less attention on the market. If you’re spending $10 million, or even $50 million, per year of your compute bill on GPUs, and it’s a tiny fraction of your cost structure, it’s easy to outsource to Nvidia.

But if you’re spending $500m/yr on GPUs and they become a material part of your variable cost structure, the calculus is different. Now, the set of companies with $100m+ in GPU spend that are hungry for margin expansion is much larger. The incentives to break Nvidia’s market dominance have grown tenfold in the past year.

This means that Nvidia’s customer base is riskier than meets the eye: the long tail of customers may not prove any ROI from their nascent foundation model GPU clusters, but the largest customers are trying very hard to undermine their Nvidia dependence.

To borrow a phrase from Bezos, “your margin is my opportunity” – and Nvidia has created tens of billions in opportunity this year.

Hyperscalers

Google started TPU development in 2015, as one of the largest consumers of accelerated computing at the time (search, ads, translate). TPU performance isn’t as good as Nvidia’s H100 series (they don’t even try comparing it), but it doesn’t need to be – they get a massive cost reduction from internal chip development. Nvidia sells the H100 unit for $30,000, and their COGS are ~$3,320, creating software-level gross margins, far higher than any other link of the GPU supply chain.

Microsoft is working on its own AI chip with AMD’s help, called the Athena project, which The Information reported on in April 2023. There are 300 people staffed on the project, which started in 2019. There are some rumors that Microsoft will prioritize support for AMD’s upcoming MI300 chip.

Even Amazon is ramping up their own chip development efforts, though fighting from behind. They still seem to be figuring out their AI strategy: they’re developing their own chips, trying to train their own foundation models, investing in Anthropic, and hosting Llama-2 instances. Typically, a higher number of strategies corresponds to a lower level of conviction.

AMD

Players like AMD are trying their best to catch up, or at least drive down the A100/H100 monopoly pricing power. Their frontier MI300 chip won’t beat the H100, but in the words of an industry insider, “everyone is looking to AMD for cost reduction” – in other words, you don’t need to believe that the MI300 dominates the H100 in absolute terms to threaten Nvidia’s pricing power; you just need to undermine their price performance.

MosaicML has gotten AMD’s MI250s to work for training runs, but given the AMD software stack is distinct from Nvidia’s proprietary and dominant CUDA language, it will be hard to use the AMD stack in practice, unless OpenAI’s Triton or Modular can close the software support gap. In the world of a private cloud exec, “Who is going to take the risk of deploying 10,000 AMD GPUs or 10,000 silicon chips from a startup? That’s almost a $300 million investment.”

Startups

With $720m+ raised and a chip on the market, Cerebras is the most credible semiconductor startup. The architecture is radically different from Nvidia, with a 56x larger chip size to reduce the need for cross-GPU interconnect. While they seemingly have competitive performance to Nvidia, they’ll need to build developer trust to capture meaningful market share. Even if they can’t build commercial-scale production fast enough, the technology has immense M&A value for a second or third-place chip provider.

As for newer programs, it takes 5 years minimum from whiteboard to tapeout of a new chip. Even if AMD’s efforts go as planned, it’ll be 2025 at least before the Nvidia monopoly over GPU supply is broken. But the target on their back is only getting bigger, and while I think Nvidia ultimately retains its GPU dominance, its pricing power could get squeezed in the medium-term.

The post-Nvidia supply chain: is renting GPUs a good business?

Nvidia is unusual in that it typically doesn’t sell its products directly to end customers. After the GPU is designed and produced, there is a post-production supply chain of resellers and systems integrators:

If cloud providers have GPUs for rent, why do you need the upstarts? Hyperscalers like Azure, AWS, and GCP are only doling out GPUs at “reasonable” prices via long-term fixed contracts – e.g. AWS will rent you 8 H100s for 3 years for $1.1m (twice as much if you don’t reserve it up front!). Smaller companies have more sporadic demand and are looking for e.g. 1,000 GPUs for 1 month, then 200 GPUs for 3 months, then 500 GPUs for 5 months, etc.

Hyperscaler upstarts CoreWeave, Lambda, ML Foundry, and Together are working to plug that gap by providing easy, variable, rented (“serverless”) access to GPUs. These businesses had near-infinite demand this year, and very fast paybacks on their hardware setups. The businesses have a similar risk profile as the crypto miners from the last cycle: you’re necessarily making a levered bet on the GPU trade, in the same way that crypto miners made a levered bet on crypto.

On the demand side of the equation, there’s every reason to believe that demand for variable GPUs will sustain. If you're a small or medium-sized enterprise (say $0-100m raised) and want to train your own cutting-edge foundation model, you may need 5,000 Nvidia H100s – you can’t buy and install them yourself at $30k+ per device, and even if you could, you wouldn’t run them often enough to break even. It makes sense to use a third party reseller.

AWS announced a 20,000 H100 GPU cluster this summer, which should satisfy the needs of larger ML training runs once completed. And they’re taking advantage of the supply shortage, pricing to an implied 68% operating income margin. Long-term, though, AWS and Azure do not want to be GPU rental businesses – they want to be selling their broader services at high margins. In this sense, CoreWeave and Lambda can be durable as a Rackspace or DigitalOcean-style companies, if they can survive the volatility of supply and demand. 

Nvidia rewards these resellers with high capacity and good pricing, since they’re not building competing chips like Microsoft and Google. This is a double-edged sword for Nvidia, since supporting the resellers gives the hyperscalers an even greater incentive to compete.

The resellers are benefitting from bringing highly sought after GPU supply online before anyone else did. Can they build durable businesses? The key axes of differentiation are scale, price performance, supply scalability, and the robustness of the software offering:

1. Scale: You need a critical scale for two key purposes: 1) to support larger language model training runs on cutting-edge models (e.g. 5,000+ 2020 vintage Nvidia A100s or cutting-edge H100s for weeks at a time, and 2) to keep utilization high across highly variable customer demand.

2. Price performance: If you view these companies as GPU-time resellers, then the supply cost really matters. Having preferential pricing as an Nvidia partner (i.e. getting the same pricing as Azure, AWS, et al) can save 30% of your cluster buildout cost. Players like Crusoe are lowering supply cost via cheaper energy. CoreWeave claims their A100s are 30% more price performant than A100s at a hyperscaler due to cluster architecture, virtualization methodology, cooling techniques, interconnect quality, etc. Some companies like Salad are even trying to aggregate cheaper supply by finding latent GPUs owned by consumers. 

3. Supply scalability: Companies like Together and ML Foundry have “acquired” (rented) latent GPU supply from underutilized private clusters at mid-stage startups. While still in the technical research phase, synthetically aggregating GPU clusters could massively unlock the supply side of the GPU equation (similar to what Airbnb did for homes).

4. Software: Some companies are going “up the stack”, i.e. improving the developer experience on top of the GPUs. Modal, Lambda and RunPod each claim to provide a faster onboarding and smoother developer experience via software.

Given the primary revenue stream for these businesses is GPU-hours, they’re essentially a levered bet on GPU demand scaling, and a bet on the longevity of the H100 (if the GPUs become obsolete in 3 years, the depreciation hit will be painful). To make an investment in these companies, you need a strong directional view on medium-term GPU consumption scaling, or believe in the long-term leverage of building out a broader software suite.

Conclusion

A myriad of foundation model copycats resemble modern-day cargo cults. They are trying their best to reproduce the magic of the first wave of products, but don’t have all the required ingredients.

Given increased compute spend seems logarithmically tied to intelligence, it’s hard to envision a future where compute doesn’t grow exponentially. But unusually, the 20-year outlook is clearer to me than the 5-year. In the medium-term, the risk of overbuilding GPU clusters is high.

There are a few things worth watching to better understand the medium-term GPU outlook:

• Large foundation model demand ramp: How much will OpenAI and other large-scale LLM providers ramp up GPU spend over the next 3-5 years? Many model providers are copying whatever OpenAI does, so their GPU consumption should be highly predictive of overall GPU demand growth. Most importantly, will the big GPU spenders ramp their cluster sizes faster than the long tail of demand withers?

• Inference vs. training mix: Seeing a shift in GPU usage for inference would clarify whether the industry will remain H100-dependent over the coming years, or diversify to chips that are more inference-specific.

• Price performance from alternatives: It’ll take years for others to catch up to the absolute performance of frontier Nvidia chips, but I would pay attention to the price performance of new semiconductor efforts. The bar here isn’t parity with Nvidia’s relatively low COGS, but with their retail price – this is when their pricing power could come under attack.

Publics

Nvidia’s exceptional growth comes with a curse of two challenging customer bases:

1. A concentrated set of buyers with strong incentives to develop competing chip programs.

2. A long tail of customers that are running speculative “build / fine-tune your own foundation model” experiments over the next year, where GPU demand predictability is impossible.

Given the cyclicality of GPU spend based on release timelines, and the valuation embedding indefinite exponential growth, I expect Nvidia to peak in the next 6-9 months, and cool off over a 2-3 year time horizon. I would change my mind if I see large H100 cluster scale-up plans from large foundation models, or high attach rates for upcoming AI product launches like Office Copilot.

I think buying AMD is a good way to bet on the long-term growth of GPUs without crowding into the Nvidia trade. They’ve lost the early race for accelerated compute, but there are lots of companies with multi-billion dollar incentives to see them succeed, and their upcoming MI chip series is promising. If anyone can chip away at Nvidia dominance, it’ll be AMD.

Don’t sue me if you lose money trading based off of my blog post (disclaimer: I’m long NVDA and AMD as of publishing).

Startups

• There will be immense value in building software to reduce Nvidia dependence, particularly if it comes with clear improvements on price performance by leveraging other GPUs. If new ML frameworks like Modular can make workloads agnostic to the underlying GPU hardware, this would unlock cheaper supply, and change the cost-performance curve.

• Be careful of investing into the GPU trade at the wrong time (and/or over-extending the risk of your “long GPU” trade), like investing in crypto mining rigs in 2021. In the last crypto cycle, there was big money to be made early in mining, until there was a supply glut and margins got crushed.

• If AMD chips end up playing a meaningful role in the inference (deployment) phase of LLMs, there will be large new data centers built, given the current data centers focus on H100s / A100s, which are best for training. Inference-focused clusters may look radically different, requiring a mix of frontier Nvidia and AMD chips.

Thanks to Divyahans Gupta, Akshat Bubna, Melisa Tokmak Luttig, Harry Elliott, Philip Clark, Axel Ericsson, Cat Wu, Lachy Groom, David Cahn, Ben Gilbert, and Matt Kraning for their thoughts and feedback on this article.

Since the 1950s, AI has been a story of the boy who cried wolf – prophecies were over-promised and under-delivered. Early AI pioneers Herbert Simon and Marvin Minsky anticipated AGI before 2000. Even in 2021, AI products were generally niche (Siri, Alexa), protracted (autonomous vehicles, robotics), or domain-specific (recommender systems, Face ID, Google Translate, personalized ads).

AI seemed stuck in the realm of big tech and academia. Most people instead focused on more profitable endeavors elsewhere for the past decade: earning paychecks from the big tech profit pool, buying tokens in the crypto run-up, building SaaS companies that’d trade at 30x ARR, collecting management fees on ever-bigger VC funds.

When ChatGPT came online six months ago, even those working in the industry were caught off guard by its rapid growth. While tech as an industry is bigger than ever, the AI frontier is being driven by a much smaller fraction of the ecosystem than former technological shifts like the internet and software. How many people had a hand in building ChatGPT, Claude, and Bard? The transformer architecture was published in 2017, but very few people worked to implement it at scale. OpenAI and Anthropic combined employ fewer than 800 people.

With a sudden phase change driven by very few people, many technologists fear that the new world we’re entering will leave them behind. This fear creates hallucinations including hope (the LLM market deployment phase will unfold in a way that benefits me), cope (my market position isn’t so bad), and mope (the game is already over, and I lost).

Today, these hallucinations propel hyperbolic narratives around foundation model FUD, open source outperformance, incumbent invincibility, investor infatuation, and doomer declarations. It’s hard to know what to believe and who to trust. Every public narrative should be read with skepticism. You should read this article with a grain of salt, too (FF invested in Scale, and a large foundation model provider).

I’ll address a few of these narratives that are always shared with a high degree of certainty, when the trajectory of the industry is anything but.

Foundation models have already plateaued

If you were in AI, but not one of the few hundred researchers who contributed to releasing a frontier LLM in the past 6 months, there is a strong incentive to believe those models are not the terminal winners.

This is why you see a lot of companies raising $50m+ seed rounds to build “better” foundation models, where the story is something like “OpenAI and Anthropic are great, but we need [cheaper pricing, more openness, less filtering, better architecture, more localization, better domain-specificity].”

But starting a new language model company in 2023 requires coping with your non-winning market position. You might tell yourself a story that the current models are the proverbial end of history, there will be many winning models, or that the winners have no moat.

A common refrain I hear is that GPT-5 won’t be that much better than GPT-4, implying we’ve reached the end of the LLM capability S-curve. I think this is a form of cope. If you believe GPT-4 is the end of what’s possible to achieve with current LLM architecture, then two conclusions follow: 1) there will be many comparable LLM providers as others catch up, and 2) LLMs will be minimally disruptive to the current state of business affairs.

With conservative extrapolation, it’s hard to envision we won't get much more sophisticated LLMs within five years. But let’s pretend the skeptics are right, and GPT-5 is a smaller leap than GPT-4 was. Even then, GPT-4 has immense latent productive value that has yet to be mined. Training on proprietary data, reducing inference costs, and building the “last mile” of UX will drive a step-function in model capabilities, unlocking enormous economic surplus.

Humans are slow to adopt new tools. It took boomers a decade to use Google en masse. I am still figuring out how to best use GPT-4. But between new model capabilities and software products built around current LLM technology, we won’t look back on 2023 as a plateau.

Open source models will dominate

Open source has made incredible strides over the past few months, from Whisper forks to LLaMA. That said, many people have strong reasons to hope open source models win: 

• Developers who don’t like relying on Big Companies

• VCs and AI infra startups that profit in a “many models” world

• Founders and researchers who feel like side characters of the AI story written in 2023

• Big tech companies without a hyperscaler to capture infra spend (if I can’t win, I at least want you to lose)

• People who want an ungovernable product (some benign, some malevolent)

• Culture warriors who want a politically unbiased product

I’m not sure if the sum of this demand means that open source wins. Debating the merits of OSS publicly is delicate: open source software is a protected class. It is unbecoming to say mean things about it.

But I have a lingering feeling that open source will not be the dominant modality for frontier, productionized models. Clayton Christensen famously thought the iPhone’s closed approach would never work:

“The transition from proprietary architecture to open modular architecture just happens over and over again. It happened in the personal computer… You also see modularity organized around the Android operating system that is growing much faster than the iPhone. So I worry that modularity will do its work on Apple.”

The current open source AI discussion has echoes of the early iPhone vs Android debate. Modular, open source approaches undoubtedly democratize technologies (Android dominates in emerging markets). But integrated products almost always capture the value. The iPhone has just 21% market share in terms of shipments, but 50% of revenue and 82% of profits.

Look at the list of operating systems. There are hundreds. How many ended up mattering? The power law distribution of operating system adoption was even more extreme than smartphones. Even within open source operating systems, there was only room for one power law winner in Linux. Almost every technology market has a power law distribution of outcomes: smartphones, social media networks, cloud infrastructure, even SaaS.

Regulators may prevent open source frontier models for safety purposes, but even in terms of market forces, open source has a weak position for most applications. Open source provides healthy competitive pressure: models from Facebook, Databricks, Hugging Face and others will keep the pricing power and neutrality of closed models in check. But the pitch for OSS models would make more sense if the large foundational model companies were abusing their power with extortionary pricing or extreme censorship. So far, I haven’t seen much of either.

There are some contexts in which open source makes a lot of sense:

1. For larger enterprises, on-prem open source models could become a critical modality: enterprises want to control their destiny and keep data private. In the data warehouse context, for example, some companies pick Databricks over Snowflake for its data interoperability, vs. Snowflake’s more locked-in product.

2. For certain applications, local processing with near-zero latency is more important than model quality (think real-time speech-to-text). These applications can use open-source models locally to trade off accuracy and breadth for speed.

But in the context of consumer-facing LLM apps and SaaS, it is hard to beat the quality of closed-source, well-maintained, integrated products. On rigorous eval sets, frontier models still dominate open source. As expectations rise for LLM reliability, robustness, and flexibility, I anticipate that open source approaches won’t be the primary modality for applications, and an even smaller percentage of the value captured.

Subscribe now

Only incumbents will win in AI

The AI moper has convinced himself the game is already over, and only incumbents will capture value. It is an easy argument to make: the biggest wins of the past six months clearly favored incumbents. Microsoft’s GitHub Copilot and OpenAI are driving shareholder excitement, NVIDIA is on a tear via a GPU demand spike, and Midjourney benefits Google’s cloud platform.

Transitioning to cloud was hard: very few 20th century software companies did it successfully, leaving room for new entrants like Salesforce, Atlassian, and NetSuite. In contrast, LLMs are surprisingly easy for incumbents to integrate. Across many software categories, it seems like incumbents scooped up the obvious opportunities: Notion rolled out an AI document editor, Epic Health rolled out an AI-powered EHR, Intercom and Zendesk rolled out AI customer support.

The LLM API form factor means quick integration, and incumbents’ wide distribution means they show the fastest wins. In the short term, AI seems to be a sustaining innovation, not a disruptive one. But I think there are failures of imagination at the product ideation level: it is much easier to modify what’s already working than to think from scratch about what new opportunities exist.

Even with GPT-4 capabilities (not to mention subsequent models), there will be entirely new product form factors that won’t have the formula of incumbent software workflow + AI. They will require radically different user experiences to win customers. This is white space for new companies with embedded counter-positioning against incumbents, who can’t re-architect their entire product UX.

In just a few months of production-quality LLMs, promising subcategories emerged: enterprise search (Glean), legal automation (Harvey, Casetext), marketing automation (Typeface, Jasper), among others. Many of these apps will succeed. New categories take time: in the productivity wave of the 2010s, Figma and Notion took years to build products ready for prime time.

To say that only incumbents will win is lazy thinking: it justifies a mopey attitude of doing nothing since the game is over. The counterfactual seems preposterous: there will be a 20-year platform shift and no new startups win? Unlikely.

There are many VC-investable AI opportunities

On the opposite side of the gloomers are the VC cheerleaders who hope for investable startups. For a long time, venture capital was synonymous with technological progress. In periods of rapid tech adoption, VCs would make a lot of money.

But VCs aren’t the primary shareholders in the current leading AI companies. How many VCs hold equity in NVIDIA or Microsoft? Even among the startups with scale, Midjourney was bootstrapped, and OpenAI has only raised a small fraction of its capital from VCs.

Venture capitalists primarily invest in early stage companies, so they need to tell narratives about the opportunities for young startups: new foundation models, vector databases, LLM-driven applications. Some of these narratives could prove correct! But VCs have a strong predisposition to hallucinate market structures in the quest for AI returns.

Many firms have been quick to develop LLM infra market maps (1, 2, 3), which shoehorn AI into a 2010s SaaS framework: lots of point solutions that constitute an “AI stack”, each of which captures some value.

This seems shaky. Much of the infra today is built to complement the limitations of LLMs: complementary modalities, increased monitoring, chained commands, increased memory. But foundation models are built via API to abstract away infrastructure from developers. Incremental model updates cannibalize the complementary infra stack that has been built around them. If only a handful of LLMs dominate, then a robust infrastructure ecosystem matters less.

The app layer should play more favorably for VCs, where many new entrants require venture capital to get off the ground. LLMs will unlock new software categories in stodgy industries, from legal to medical to services. But because software incumbents can take advantage of LLM APIs themselves, the app-level opportunities for startups will be limited to new product paradigms.

If there aren’t many VC-investable opportunities, then VCs will have no economics in the most important platform shift of the last several decades. At the infra layer, VCs deeply want the “fragmented AI stack” thesis to be true, both as a coping mechanism and to demonstrate to LPs and founders that they’re on top of it.

Undoubtedly, some of these VC-backed companies will be large, standalone companies. They just need to be evaluated with a cautious lens given the incentives at play. I worry that the VC deployment velocity will outpace the venture-backed wins.

AI needs to be slowed down

There are many people whose incentive is to slow LLM development down: coping non-winning foundation model developers, the hopeful CCP, workers who risk job displacement, AI doomers.

Nth place LLM developers: The “Pause Giant AI Experiments” open letter reads as thinly veiled competitive FUD. The motives were too clearly self-interested: many of the signatories were AI developers behind the frontier, and wanted time to catch up.

The CCP: A dead giveaway that China is behind the US is that their state reporting focuses on the number of LLMs being released, rather than the quality or ubiquity. Like the signatories on the open letter, China would love for the US to slow down on LLMs, to narrow their competitive disadvantage.

Workers who could be displaced: There are many people who have a professional reason to slow down AI due to substitutive fear: copywriters, call center employees, clerical workers.

When GPT-4 fails on particular tasks in math and logic, people are eager to call it out. This type of content always seems to go viral. It is comforting to know that the AI is not yet better than us at everything.

AI doomers: I take AI takeoff risk seriously – we simply do not know how substitutive or adversarial it will become in 5-10 years from now. At the same time, the AI doomer argument is unfalsifiable: an apocalyptic prophecy is impossible to prove wrong. The messengers’ value systems also create their own set of biases; I’ll leave this analysis as an exercise to the reader.

The slowdown message should be received skeptically, when there is no empirical evidence of tight enough feedback loops to lead to an AI takeoff. Aggressive alarmism too early will discredit AI safety; claiming an AI apocalypse before there is one will desensitize people to real risks. In the meantime, the productivity benefits to humanity are too great to wait.

Conclusion

During the early internet era of the mid-90s – and crypto in the early 2010s – very few people showed up in the first few years, but almost all of them made money.

AI seems like the inverse. After the ChatGPT launch, everyone has shown up to participate in the AI game: researchers, developers, VCs, startups. But unlike former technological shifts, it is possible that very few new players stand to win in the deployment era of LLMs. Unlike crypto, the value prop of LLMs is immediately obvious, driving rapid demand; unlike the internet, LLM products leverage cheap ubiquitous computing to deploy quickly. The few AI players in pole position are winning at lightning speed.

Because the AI revolution is so unevenly distributed, there are very few people defining the frontier – this unfortunately means that the rest of the technology industry is some combination of uninformed and biased. Most talkers are playing catch-up, hallucinating a self-serving worldview that’s a mix of despondence (mope), denial (cope), and salesmanship (hope).

I haven’t been able to cover all the AI narratives I regularly come across, but whenever you hear one shared with certainty, alarm bells should go off in your head. Nothing is certain this early in the cycle. The truth lies in what isn’t discussed publicly, and you have to seek it out yourself via conversation and experience. The only trustworthy source is first-hand.

Subscribe now

Thanks to Axel Ericsson, Scott Nolan, Melisa Tokmak, Philip Clark, Yasmin Razavi, Cat Wu, Divyahans Gupta, Brandon Camhi, Lachy Groom, and Roon for their thoughts and feedback on this article.

The nuclear winter techpocalypse arrived: software, SPACs, fintech, and crypto all entered a deep freeze. AI may be the only sector wearing a parka.

Peak AI indicators abound. VC thought pieces and tweetstorms have reached all-time highs. Fair-weather fans from the crypto boom have migrated. MBAs may soon outnumber nerds.

Just a year ago, crypto had comparable froth. Can AI avoid crypto’s bubble-bursting fate? Crypto provides a cautionary tale for excessive hype, with four critical flaws:

1. Capital: Crypto-specific capital sources cleaved from traditional venture capital, distorting valuations and liquidity timelines.

2. Mission: Founders lost sight of crypto’s decentralizing founding mission.

3. People: Token extractionists overpowered ideological builders.

4. Value creation: Founders generated more value for token holders than users.

Both sectors overheated, but that doesn’t mean they’ll collapse in the same way. If navigated carefully, AI can escape the downfall of inflated expectations.

Capital

In the early 2010s, no institutional investors spent time on crypto. It wasn’t even possible: venture capital LP agreements restricted investments into assets like cryptocurrencies. A few VCs solved the LPA crypto constraint by 2015.

In 2017, institutional crypto investing changed. Every VC started looking seriously. Most startups had a crypto strategy. Dedicated crypto funds emerged. And in 2020, a free-money environment gave funds excess capital to deploy, adding rocket fuel to a speculation-prone category.

Two capital problems took hold: a capital vs. progress mismatch, and VC short-termism.

Constant capital, non-constant progress: A myriad of crypto-specific funds provided a steady flow of capital into the ecosystem. This seemed like a good division of labor: generalist VCs could tune out crypto noise; crypto founders could raise from specialists.

On the way up, fast liquidity for crypto funds meant the specialists quickly scaled up to billions in AUM, rivaling generalist venture fund sizes.

But when the music stopped, the specialist VCs couldn’t evaluate crypto deals against the opportunity cost of startups in other industries. This opportunity cost may seem trivial, but acts as a rate limiter: if industry B’s progress exceeds industry A, more incremental venture dollars flow into industry B.

Constant capital inflows juxtaposed non-constant (declining?) progress in user adoption and commercial success, decoupling valuations from traditional startups. Investors justified valuations self-referentially: “X token is better than Y token worth $10B, so it’s a no-brainer investment at $1B.”

VC short-termism: As a financialized technology, crypto’s allure of getting rich quickly altered the founders and VCs alike.

VCs typically wait 10+ years for a return. But short-termism plagued investors after they saw that they could get token distributions in less than 3 years, in defiance of historical norms. VCs chose short lockups, and encouraged pre-launch token sales to get fast liquidity.

The economic incentives simply haven’t been to build and fund real applications over a long time horizon. Why continue building if you already cashed out?

AI avoids these capital challenges.

Generalist funds keep AI in check: There will continue to be AI-specific funds, but they won’t dominate the industry like crypto funds did. Technology-specific funds tend to focus on infrastructure companies, which is highly oligopolistic – there won’t be many new VC-backed AI infra winners in the 2023 vintage. The real opportunity for VCs today is specific applications.

When a technology becomes ubiquitous, funds dedicated to that technology become meaningless: it wouldn’t make sense to create a fund to invest in companies using databases. The distinction of "AI company" will cease as a common phraseology, save for infrastructure. It is assumed that you leverage machine learning.

What VC firms do well in an AI bull market? Probably the generalists, who see the big picture: buyer psychology, distribution strategy, compounding advantages.

With generalist investors in control, AI shouldn’t have a runaway capital problem. ML-driven products interact deeply with the real economy – customers, competitors, and investors provide regular reality checks.

AI is structurally long-term: Some early employees will profit from secondary tenders, but AI has no obvious pump-and-dump schemes. Quite the opposite:

• Unlike crypto, AI has no built-in liquidity mechanism.

• In spite of the hype, the commercialization of AI is nascent – real, durable businesses take time to mature.

• In a high interest rate environment, only real businesses can exit through the public markets or M&A.

If anything, profiting off of AI will take even longer than the last generation of software and internet companies.

The notion that AGI could be worth infinity dollars may distort people into overvaluing AI assets for the call option value. But the current macro and micro capital contexts should keep the bubble in check.

Mission

Entrepreneurs have tried to conjure digital currency since the 1990s, with years of failed attempts. In 2008, Bitcoin demonstrated that cryptographic proof of work + the blockchain was the first satisfactory solution.

The whitepaper made the use case clear: circumvent financial institutions. Crypto’s mission was anarcho-capitalist and revolutionary. Founders are the best missionaries for their respective technologies, and in Satoshi Nakamoto’s words:

“The central bank must be trusted not to debase the currency, but the history of fiat currencies is full of breaches of that trust. Banks must be trusted to hold our money and transfer it electronically, but they lend it out in waves of credit bubbles with barely a fraction in reserve. We have to trust them with our privacy…”

Crypto’s ideology fueled early applications. Decentralized money turned out to be useful, both legitimately and illegitimately: buying pizza pseudonymously, purchasing drugs on the Silk Road, escaping tyrannical governments’ fiat systems. Trade cash for Bitcoin, and you exited the supervision of centralized banking systems.

The mission later became murkier. Crypto’s decentralized purpose got lost: activity became centralized, transactions tracked, and on-ramps added KYC and AML provisions. Maybe crypto had to integrate with the financial system to popularize, but if the blockchain isn’t decentralizing, what is it really doing?

The mission scope crept as crypto grew into web3. Originally about decentralized money, crypto became about decentralizing everything. Builders focused on blockchain-based social networks, decentralized gaming, and NFT ticketing. Crypto is digital money, but people felt it needed to be more.

Even if now corrupted, crypto’s original reactionary mission hits you over the head. The mission is in the founding whitepaper title, the forum discussions, the architecture, the use cases.

If decentralization is the founding crypto mission – what is the AI equivalent? Hollywood isn’t painting a positive picture, featuring AI in every dystopian sci-fi movie: HAL 9000 in Space Odyssey, Skynet in The Terminator, Samantha in Her, Ava in Ex Machina.

Outside of film, people tell credible dystopian AI narratives around entrenching wokeism more deeply, or worse, replacing human work. AI can be a symbiotic partner to human creativity and thought, but the optimistic story needs to be shared.

AI’s positive mission is not as crystallized as crypto’s, or at least not yet.

Unclear founding mission: There’s nothing missionary built into the technology itself, unlike blockchains which are trustless and thus anti-institutional. This could be a good thing: AI is an ideological blank canvas, and can be shaped to match human will.

Many people are excited about the mission of AI. What exactly are they excited about? The ideology is some form of post-scarcity or abundance – a spectrum from abundance capitalism to communism. In AI’s early modern era (mid-2010s), the mission was effective altruism: build the AGI, then give away the profits.

The stated missions of many AI companies rhyme with failed socialist regimes from the 1900s, scaring off conservatives and libertarians. But it’s markedly different: generating abundance through technology has a better precedent than European socialism, which redistributes abundance through political force.

Centralized control: Control over AI’s foundational models seems likely to centralize. Open source models show promise, but large privatized models are far better. Even crypto ended up centralizing far more than its vision suggested.

My best mental model for AI centralization is semiconductors. There are many providers of lagging edge chips, but very few (and valuable) leading edge providers.

Centralization sounds dangerous if we think of political history, but decentralizing the technology isn’t necessarily good either – if AI becomes all-powerful, you would not want everyone to have nuclear launch codes.

Mission neutrality: Most industries have application-infrastructure duality. In software, for example, cloud providers handle infrastructure, while software companies build applications. Bitcoin was monistic: the infrastructure was the application. This meant the mission was defined at the infrastructure level.

Most enabling technologies are relatively missionless: cloud infrastructure, semiconductors, mobile, even the internet. Perhaps AI should be mission-neutral at the infrastructure level, while the mission gets defined at the application layer. DeepMind’s stated mission seems to point in the right neutral direction: “solve intelligence, and then use that to solve everything else.”

Foundational models should power a wide range of application-level missions – political, social, economic. Much like AWS shouldn’t censor developers except in the most extreme cases, large models should avoid censoring AI applications.

Unlike cloud infrastructure, foundational models have training bias. But they also have a fine-tuning lever, allowing application-level mission flexibility. Developers can generate Fox-weighted or NYT-weighted responses. The current political bias in large models has a long-term fix.

AI’s mission still needs refining, and who controls it is an open debate. But in a world where the infrastructure centralizes, the best mission may be neutrality.

Subscribe now

People

Crypto’s pioneers were libertarians and anarchists, but disorganized grifters flowed in as the ecosystem matured, attracting regulatory scrutiny.

• Cultural drift: In the 2008-2016 era, only technologists and ideologues bought and built crypto.
  
  Crypto’s original application required its original sin: tokenized architecture. The corrupting force of financialization ultimately overpowered the usefulness of the technology.
  
  A few years later, grifters piggybacked on the ICO mania. Easy money corrupted even the puritans: crypto founders vested millions of liquid tokens in 1-2 years without user adoption or managerial oversight.
  
  Crypto returns became a hyperstition – a positive feedback loop driven by culture. Hype-driven folklore drove retail adoption: WAGMI, the Bitcoin rainbow chart, the supercycle thesis, “few”. Anyone could profit off of the technological revolution!
  
  The hyperstitious retail wave created $3T in market cap during the easy-money ZIRP period of 2020 and 2021. Unfortunately, most people got burned.
  
  

• Regulatory creep: Under the rule of chaotic short-termists, crypto failed to garner regulatory buy-in. Some crypto companies worked to earn institutional trust, while others worked to undermine it. The company closest to regulators undermined regulatory trust the most.
  
  The crypto crowd was too anti-institutional to win over regulators, but too uncoordinated to pull off the anti-fiat mission. Buying Bitcoin in 2023 requires submitting your SSN and passport to centralized on-ramps. Each transaction is tracked by intelligence agencies. Regulators have crypto right where they want it.
  
  When the bubble popped, the SEC, FinCEN, and OFAC were unrelenting. Given entrenchment of centralized financial institutions, maybe regulatory support was never possible. Either way, centralization won.

AI has a purer talent evolution arc, but the same risk of regulatory scrutiny.

• Cultural scaling challenges: The formative days of AI development had an ethos of research and publishing, ensuring that technologists, not financialists, stayed in charge.
  
  In 2023, AI must absorb an influx of tourists – thin GPT wrappers, MBA tweet threads about AI trends, LinkedIn bios changing from #crypto to #AI. Genuine technologists will join the development effort, but filtering out negative human capital is challenging.
  
  Undoubtedly, more builders are needed to apply and productionize the latest technologies, but scaling headcount too fast will cause indigestion. Most great products are built by very few, highly talented people.
  
  The best cultural control mechanism is to keep technologists in charge. Crypto’s tokenization element amplified the voices of swindlers. In AI, there’s no trading – participants must build or leverage the core technology. AI promises productivity, not easy money.
  
  There is still hope for non-ML engineers. You don’t need a PhD in databases to use one – we’ve ascended the abstraction ladder. As foundational models improve exponentially, it’s becoming easier for generalist engineers to build AI apps. The limiting reagent will be founders and employees with end-market expertise and distribution advantages.
  

• Regulatory risks: As any technology scales, people can’t help but to burden it with politics. AI is no exception, particularly given it touches many end consumers. Depending on who you ask, current AI models are too woke or not woke enough.
  
  Tropes of AGI replacing human labor predispose regulators negatively. AI safety researchers are working hard to prevent dystopianism from escaping sci-fi movies, but regulators will create strict guardrails as models improve.
  
  The strategic importance of AI means that regulators will draw strong geopolitical boundaries. We won’t be using Chinese models to power our applications any time soon. Even AI-powered TikTok is under immense bipartisan scrutiny.

The research-heavy origins and longer-term business models should prevent tourists from taking over. But political controversy and regulation will slow things down.

Value creation

In crypto, people engage in serious debates about whether there’s a single web3 use case that creates user value. Let’s define net value created as:

It's unclear if this sum is positive, zero, or negative. The first half of the equation certainly has positive subcomponents – people use Bitcoin to escape tyrannical fiat systems, and pay journalists in oppressive countries.

But there are also negative subcomponents. Engineering time spent on crypto instead of other, higher-ROI technical endeavors. People losing their FTX holdings, BlockFi deposits, and altcoins down 90%.

Crypto leaders claim we’re still in the infra phase. Infrastructure is necessary but not sufficient to generating value – sustained usage trumps trading volume. The developers innovating at the application layer have faced launch delays or active regulatory resistance.

Successful crypto infrastructure startups provide infrastructure for other crypto infrastructure companies and traders. Leverage within the crypto ecosystem ballooned, in the form of trading leverage, crypto-denominated lending, and crypto-based reserve policies.

Web3 lingers in the value-neutral infrastructure phase, encumbered by self-referentiality. There are exciting research frontiers like zero-knowledge proofs, which could unlock new growth. But in most commercial dimensions, adoption has stalled.

On the other hand, people empirically use artificial intelligence to do valuable things.

Big tech has leveraged it for a decade: product recommendations, news feeds, spam filters, ad personalization. Within days of ChatGPT’s launch, there were myriad concrete use cases.

Maybe ChatGPT and generative AI won’t scale into robust, dependable systems. Journalists (who, ironically, may get obsoleted first) think it’s dumb and overhyped. But it’s hard to ignore the practical applications: Copilot accelerates coding, Jasper simplifies copywriting, Midjourney and DALL-E superpower artistry, ChatGPT provides analysis and answers questions.

The negative-value cases are harder to pinpoint. The TikTok algorithm may be the best example of wasted time, though in a world without AI-powered feeds, social media consumption is likely still very high.

Value capture cuts in two ways:

1. Infrastructure oligopolies: If you believe in continued returns on scaling compute, data, and parameters, the large model race should play out oligopolistically: scaling models in the 2020s is a game that requires billions to play. Value capture for large models could play out like semiconductors, where a handful of winners will emerge per geopolitical region.
   

2. Application proliferation: Unlike infrastructure, application value capture will be diffuse. It is late to start a new foundational model company, but VCs have yet to focus on the application layer. Many multi-billion dollar applications will emerge on top of large models.
   
   If the foundational models evolve too quickly, that spells trouble for thinner GPT wrappers. But applications with proprietary domain-specific training data, unique distribution, and complex integrations will endure.
   
   Some startups pitch themselves as “full stack”, straddling application and infrastructure, building custom models to power domain-specific applications. This will be tough: customizing excellent generalized models is easier than reinventing models from scratch. To use a present-day analogy, very few software companies should be building customized hardware.

A superintelligent AGI could generate negative value – a “Zeus [that] throws lightning bolts at people”, in Peter’s words. This tail risk makes AI safety work worthwhile. But alignment already works in other areas: financial regulations keep capital (an artificial intelligence, of sorts) aligned with humans.

The productionization of AI systems is nascent, but net net, it seems to crush crypto on value creation.

Conclusion

Can AI maintain the faith that crypto lost? It needs to get these things right:

• Capital: AI has had long capital feedback cycles, which keeps froth in check. But applications must productionize to justify continued capital influx.

• Mission: AI’s ideology is more of a blank canvas, and the technology is less intrinsically corrupting. But it needs a positive – or at least neutral and apolitical – mission to disprove the dystopian prophecies.

• People: The industry will have its fair share of grifters, but AI leaders can’t let them have real control over the ecosystem.

• Value creation: AI adoption seems promising, but transitioning from interesting toys to trusty tools is non-trivial.

In the long term, value creation should dominate, and AI seems to be winning handily. Businesses and consumers benefit from the technology even in its nascency.

People operate under the Gartner hype cycle framework: what goes up must come down.

Gartner gets the early shape of the curve right, but the plateau of productivity is misleading: it varies dramatically by industry. Some plateaus round to zero – say private cloud computing – while others surpass the peak of inflated expectations.

Crypto could plateau at a fraction of what Gartner suggests. If I had to guess, AI will plateau far above the historical peak.

Subscribe now

Thank you to Brandon Camhi, Axel Ericsson, Melisa Tokmak, Philip Clark, Sam Wolfe, Anna Mitchell, Russell Kaplan, Michael Solana, John Coogan, and Yasmin Razavi for their thoughts and feedback on this article.

Private financings could soften significantly…what would you do if fundraising on attractive terms proves difficult in 2020 and 2021?

With softening sales, you might find that your customer lifetime values have declined, in turn suggesting the need to rein in customer acquisition spending to maintain consistent returns on marketing spending.

But the prognosis was precisely wrong.

Most internet and software businesses would have been better off following the opposite strategy during the pandemic. Fundraising became easier than ever, and demand skyrocketed for tech substitutes to real-world products. Getting predictions wrong by two years means massive missed opportunities.

The pandemic brought unprecedented volatility to the tech industry – but with two years of hindsight, what has changed?

2020-2021: tech exponentialism

Months after the “imminent” tech crash in March 2020, the demand shock never came for most categories. In fact, indicators of tech acceleration surfaced by early summer: exponential e-commerce adoption, shortening software sales cycles, SaaS vitamins becoming painkillers.

Instead of a collapse, summer 2020 was the start of a historic tech bull run.

Two factors fueled tech growth:

1. Falling interest rates: Rates dropped to near-zero, making a dollar of profit in 2030 nearly as appealing as a dollar in 2021. Companies with profits far in the future – think high-growth and high-risk companies – got undue credit for those future dollars.

2. Rising tech demand: COVID raised the demand floor for online alternatives to physical businesses. E-commerce doubled. Food delivery tripled. Video conferencing quintupled. This created a credible argument for tech domination: DoorDash would crush restaurants, Zoom would render offices obsolete, Peloton would eliminate gyms.

The consensus view by summer of 2020 was that the pandemic would dramatically accelerate existing tech trends. Wishful thinking reigned king: if two months of COVID doubled tech trends, what could two years do?

There were three potential scenarios for the future of technological growth:

• Exponential: Pandemic era acceleration represented the new baseline growth rate. Instead of growing 10% annually, e-commerce and other categories would have a new sustained level of high growth.

• Step function: COVID established a “new normal” for tech demand, a step function above pre-pandemic levels. Old growth rates would continue post-COVID, just from a higher base established in 2020. This was a popular view among VCs and tech pundits.

• Mean-reversion: Acceleration through 2020 and 2021 was an aberration, and tech market growth will revert to the pre-COVID growth trendline.

In e-commerce, for example, these three stories you could tell about the future in summer 2020 would look something like this:

Public market prices started to reflect a combination of exponential and step function tech growth. This meant that all tech companies were undervalued by a wide margin pre-COVID. At-home fitness, security, e-commerce, video conferencing – if software was going to eat the world overnight, the respective stocks needed to reprice as such.

Tech growth engines have demonstrated a unique property in the last decade: durable growth rates. Many big tech and software companies – Facebook, Google, Amazon, Salesforce, Atlassian, PayPal – compounded at 20%+ for more than a decade.

When tech companies see growth acceleration, it is often permanent – look at Microsoft’s revenue from 2012-2017 (5.5% CAGR) versus 2017-2022 (15.6% CAGR). Once its growth engine proved it could compound faster, the growth rate had a new floor.

Did COVID establish a new baseline growth rate for tech?

2022: tech reverts to the mean

Mean-reversion seemed to be the least popular story to tell – who wants to rain on the tech parade? But the last six months clarified that much of the tech acceleration was transient.

Tech massively over-earned in 2021 – generating growth and profits at a temporarily high rate – driven by substitute demand from the pre-COVID economy. Companies with accelerated temporary growth got credit for a permanent acceleration.

But investors recently realized that underwriting growth permanence from COVID acceleration was wrong. The ensuing correction was harsh, but note that many high-growth companies are trading very close to their pre-COVID prices:

Under the surface, many pandemic era trends assumed to be permanent turned out to be temporary, and many seemingly temporary trends may be permanent. E-commerce seemed like the most inevitably permanent COVID trend. But in fact, e-commerce penetration has reverted to the trend line over the past few quarters:

Many other tech categories follow a similar path of mean-reversion. I wrote an article in April 2020, When Tailwinds Vanish, which was seemingly disproven just a few months later: tech had a banner year in 2020. But the last six months proved the opposite: tech’s over-earning during the pandemic means that tailwind decay will be even harsher now.

Bessemer’s Emerging Cloud Index (EMCLOUD), which is probably the best proxy index for public high-growth tech, is trading within 10% of March 2020 levels. So at a crude aggregate pricing level, COVID was net neutral for tech. But even compared to the major indices, high-growth technology companies fared neutrally relative to the old guard:

I have no doubt tech will continue to compound, but the laws of physics still apply. Tech and non-tech products are largely substitutes: fitness bikes versus gyms, gaming versus sports, takeout versus restaurant dining.

The nation’s COVID recovery was zero-sum with tech exuberance.

Subscribe now

The mid-2020s: stagflation is a lingering symptom

The tech exponentialism we saw during the pandemic was not only driven by a demand spike – it was exacerbated by low interest rates and money printing.

But it turns out you can’t print $10T without anyone noticing.

Macro factors like inflation didn't matter in tech for the 1990s through the 2010s. They were offset by total internet and software growth rates in the high 20s and 30s. Inflation was always below 4% – as a percentage of growth, it was an irrelevant headwind.

But many tech category tailwinds are flipping to headwinds as we revert to the mean – when you’re growing sub-30% and inflation is 10% a year, inflation becomes a far more consequential growth headwind.

Inflation sneaks in indirectly – it’s not simply the cost of gas and groceries. There are several hidden inflationary factors for startups today:

• Labor cost: Rising salary expectations are the most obvious and direct impact to startups. When inflation became common knowledge in the 1980s, unions pre-negotiated annual wage increases – this could happen in tech. When salaries rise, you have to burn more capital to build the same products. While labor cost inflation will slowly restabilize, expect regular wage hikes for the next several years.

• Productivity: In a more competitive tech talent market, employees expect not just higher compensation, but less work – a double whammy. The four-day work week is already gaining popularity. And many employers are acquiescing for fear of attrition.

• Equity burn: This is a painful inflationary backdrop to every startup, and under-appreciated even at the most sophisticated companies. The option pool is each startup’s internal Federal Reserve, setting an “inflation rate” for your share price. As competition for talent increases, so does your option pool inflation rate. Every year, your share price is offset by option pool expansion, which can range from 2-5% for growth-stage companies. Public companies are offenders too, with stock-based compensation constituting 10%+ of revenue for many tech companies.

• DCF impact: If your cash flows are far in the future (high-growth and sub-scale businesses), USD inflation and rising rates hurt you the most. As interest rates rise to match inflation, your valuation gets less credit for profit generated in ten years from now. If your average dollar of profit comes in thirty years from now, your stock would be down ⅔ from mid-2020 even if nothing changed about the odds of that dollar being earned.

Labor squeezes and compensation inflation will take years to stabilize. Tailwind reversion plus inflation will usher in an age of tech stagflation. Startups need to demonstrate exceptional growth and efficiency to overcome these inflationary headwinds.

How should we navigate tech stagflation?

Reversion to the mean takes time. But the easy wins are over, for both VCs and founders. In a way, the Sequoia memo from March 2020 is useful now, but because of the harsh return to normalcy, not because of a sudden demand shock.

Look at anything that grew massively in the past 2 years with a cautious eye: crypto gold rushes, venture capital expansion, diligence-free fundraises. I’m not saying these trends are not durable – a subset of pandemic-era shifts certainly will be. But mean-reversion will be a strong gravitational force across categories over the coming years.

Investors, founders, and employees must relentlessly seek signal within the noise generated over the past two years to assess what changes are durable vs. transitory.

Investors

The challenge for investors is to find non-obvious but permanent shifts. Unfortunately, ubiquitous 100x ARR multiples is unlikely to be one of them.

Investors and the media cheered on VC firm AUM expansion over the past 24 months. Deal sizes and deployment speed skyrocketed. Law firms were backlogged for months due to unprecedented deal pace. Hordes of venture tourists arrived.

2021 was peak VC imperialism.

The most bullish pandemic-era tech investors are now in a world of hurt. SoftBank registered a $27b loss on their Vision Fund in Q1. Tiger Global is down 44% on its public hedge fund through April, not counting the even steeper May correction. Their private book valuations are likely down even more (at least they have preferred shares).

I wouldn’t be surprised if some of these crossover firms get margin calls – their public books often run on leverage – and some could implode entirely. Going forward, LPs will grade new managers and markups on illiquid positions much more harshly. The private market correction is somewhat of a slow motion train wreck – the markdowns and failed fundraises don’t happen immediately.

The silver lining for investors is that many companies with stalled growth due to mean-reversion are now valued on a much more punitive profit multiple. Shopify, Netflix and Zoom, for example, saw multiple compression of 80% as their earnings reports showed slowing growth. Valuations for true compounders should renormalize as their growth engines prove durability.

Where should investors focus in an ecosystem reverting to the mean?

• Wish list: In 2021, every software company with a pulse warranted a 100x valuation multiple. This meant mediocre companies were massively overvalued. But the inverse is also true: in a market reset, multiple correction means that the best companies will go on sale, too. Come up with a target list of your top 10 picks so you’re ready if and when companies reach renormalized prices.

• Counter-cyclicality: Investors always tell their LPs that they invest in secular shifts – they’re currently figuring out how secular some of those shifts really are. But an even better strategy may be to find counter-cyclical categories: energy, healthcare, infrastructure, and defense are contenders for truly macro-proof industries.

• Consolidation: We’re likely to see strategic M&A and rollups, particularly in mid-stage software startups that don't achieve escape velocity. Companies that cannot grow fast enough to justify high burn will need to find acquirers as private funding markets soften.

• Remote work: Hybrid and fully remote work is far more permanent than people expected. Remember that in March 2020, people expected a return to the office within a couple of weeks. The return date slipped for weeks, then months, and in many cases, indefinitely. Today, hybrid work is commonplace for white collar workers at large companies, and many startups are building products to support the shift.

• Crypto caution: Crypto and web3 exploded during COVID. Non-BTC crypto is up over 10x since 2020. Billions flowed into projects promising fast profits. Enthusiasts argue crypto disruption of traditional financial institutions is imminent, but the 2021 run-up seemed more driven by a collective belief in the ROI of lottery tickets – a symptom of easy money. Is web3 overvalued?

• Institutional mistrust: Distrust of centralized institutions – government, big tech, schools – skyrocketed following pandemic-era authoritarianism. These second-order tailwinds could linger for years, opening up opportunities for startups building the future of privacy, transparency, decentralization, and education.

Founders

In a return to normalcy, founders should revert to 2019 benchmarks for fundraising and operations, and be careful in absorbing any pandemic tailwinds they benefited from.

• Agile spend: Some spend, like online ads, is easy to control: if you want to cut your ad spend in half, you can do so immediately. But some costs are hard to unwind: Peloton massively ramping up hardware spend assumed a sustained demand level that was in fact temporary. Hiring to meet demand spikes is tricky, too: see Hopin’s recent layoffs as a cautionary tale. Avoid fixed costs for variable demand, and be ready to unwind them quickly if demand softens.

• Fundraising expectations: VCs have been underwriting your business based on high-growth public comps, so look to public companies’ valuations and strategies as guidance for how to benchmark your business and build out your investor narrative. Getting fundraising tips from founders who raised in 2021 is irrelevant – it is far more useful to ask founders who raised pre-pandemic.

• Isolate pandemic and secular demand: An exercise in finance and data science, the best teams will isolate temporary and permanent shifts in business fundamentals – CAC, retention, LTVs. Assessing these variables requires ruthless intellectual honesty. Gauging retention means filtering through customers with a recent renewal option: are customers purchasing and engaging with your product at the same rates post-COVID?

• Geography: The pandemic shed light on San Francisco’s many problems. The positive network effects of Silicon Valley are increasingly offset by negative network effects – homelessness, taxes, political groupthink. But physical offices drive productivity and culture, while free-rein remote work lowers employee accountability and learning by osmosis. Is returning to the office undervalued?

• Profitability: The best time to learn how to build a finance model was when you raised outside capital. The second best time is now. Having visibility over your business levers is table stakes. A concerning number of companies raised on a simple deck or memo in 2021 – if you have never been through a challenging fundraise, you could be in a world of pain if you don’t control your financial destiny. Profit is the great equalizer.

Employees

Unsavvy employees can get crushed by the monstrous valuations of 2021 that take years to grow into.

• Scrutinize businesses: When evaluating prospective employers, candidates often substitute valuation for job security. If you’re assessing a company that last raised in 2021, don’t take solace in the billion dollar valuation they fetched. Understand their actual revenue scale and financial health (margins, runway, growth rate) to get a sense for the magnitude of the business. Great businesses will survive, but it could take years for the fundamentals to catch up to 2021 entry prices. If you’re uncertain on how to value various companies yourself, speaking with tech investors is a reasonable substitute.

• Assess product durability: When looking at job opportunities, be discerning on the durability of the product value proposition. If they experienced exponential growth during the pandemic, will it continue after trends revert to the mean? This should be fairly obvious at a category level: mission-critical infrastructure software is likely durable; a levered bet on virtual events may not be.

• Find recent data points: Put increased weight on current data points from startups. Recent fundraises and signs of product-market fit carry additional merit – bonus points for indicators from the past one or two months. Of course, joining seed-stage companies is a completely different risk profile, which is relatively macro-agnostic.

Conclusion

Two years with COVID have been hectic, but what actually changed? A return to normalcy is the boring answer that nobody wants to hear.

But reversion to the mean seems to be an increasingly powerful explanatory force, and not just in tech:

• Political unrest: From the BLM protests to the January 6 Capitol riots, protests spiked across the political spectrum during the pandemic. The reduction of face-to-face human interaction made humans far more volatile. With more human contact and less idle time, politics may stabilize going forward.

• Retail trading: Retail trading volumes reached a new baseline in the pandemic – volatility, boredom, and stimulus checks are a powerful cocktail. GameStop and Dogecoin represented peak retail trading. Volumes have already started to renormalize.

• Authoritarianism: From country-wide lockdowns to permanent Twitter bans, authoritarianism popularized during COVID. Thankfully, freedom antibodies have started to form: SF residents recalled the school board to focus on returning kids to school, the Twitter algorithm could become ban-resistant and open source, and the mask mandates have been struck down.

• Consumer habits: Travel halted, movie theaters shuttered, retail spending and GDP plummeted in 2020. The speed of return has been uneven across categories, but travel is nearing pre-pandemic levels, GDP has reverted to the trendline, and movie tickets are on the rebound. I wouldn’t invest in a movie theater chain in 2022, but the death of the old economy is further out than the 2020 forecast suggested.

Not everything will revert to the mean. Movie theaters may never reach 2019 levels, Zoom is here to stay, and even I have been surprised by the durability of remote work. And mean-reversion does not spread at the same rate for every phenomenon: inflation will persist far longer than home fitness sales.

But many pandemic-era trends will slowly revert to their pre-2020 trendline – inflated valuations, easy money, exponential tech demand, meme investing.

Mean-reversion and inflation will be lingering symptoms of the virus for years to come.

Subscribe now

Thank you to Harry Elliott, Brandon Camhi, Axel Ericsson, Michael Solana, Amin Mirzadegan, Philip Clark, Lisa Wehden, Pranav Singhvi, Melisa Tokmak, Russell Kaplan, and Everett Randle for their thoughts and feedback on this article.

Apple released its iconic Apple II computer in 1977 with just 25 employees. One engineer built the file system, one engineer built the input/output, and one engineer built the graphics. Microsoft, too, shipped its first operating system with just 100 employees. Even in the late 1990s, Google was launched with just 4 employees.

Each company had to chart its own path for product strategy, go-to-market, business model, and hiring.

As Silicon Valley matured, market sizes for tech companies grew 10x bigger than anticipated. Thousands of startups arose to chase these opportunities, fueled by internet adoption and cheap capital.

Then came an insatiable demand for talent. Consultants, analysts, and bankers flocked to high tech salaries. Career ambition replaced technological ambition as the driving force of startups. Employee count scaled faster than company productivity. Accelerating growth in teams and revenue masked declining productivity per employee.

Eventually, startups became so similar that playbooks emerged – and generally, they worked. Today, any startup can access an instruction manual for every aspect of marketing, sales, engineering, and even product strategy. For each playbook you follow, you’re outsourcing your strategy in return for jumpstarting the respective function.

But cheap cash, big teams, and playbook dependence come at a cost: incrementalism.

Tech’s growth in spite of declining productivity means that incrementalism can still yield success. Founders and execs can tackle increasingly narrow problems and still make millions.

Incrementalism is squandering Silicon Valley’s potential. Many of the nation’s most talented people are iterating on a tiny product surface area. First principles thinking is being replaced by shortcuts like NPS optimization and bookings-to-burn ratios.

Rippling is one of the few beacons of ambition in an increasingly incremental world. To understand the magnitude of its vision, we first need to understand the magnitude of the problem it is trying to solve: the administrative crisis.

The administrative crisis

The world sees the United States as a model for corporate prowess and agility.

But I worry the opposite is true: most companies sink far too many resources into menial tasks instead of company building. 

The post-covid world is even more burdensome. Managing employees across geographical borders brings a new set of HR, tax, compliance, and IT security challenges. 

For fast-growing startups, this overhead requires sacrificing precious time and elevating burn rates. For non-tech companies, maintaining a web of HR, IT, and compliance obligations is a constant burden, prohibiting them from investing in growth, building a strong culture, and solving big problems.

Hundreds of point solutions across dozens of Gartner quadrants solve subsets of these problems, but these narrowly-focused products cannot solve administrative challenges across systems. It’s not surprising that total factor productivity has stagnated, despite an explosion of productivity software tools promising the opposite. Not only do point solutions fail to accelerate productivity, their sprawl is actually slowing things down:

Consider the lifecycle of a remote employee. First, you must register them with the correct state tax agencies. Then you need to ship them a laptop, ensure it has the right security settings, and provision their accounts to systems like Zoom or Slack. Once they’re onboard, you must manage their benefits, comply with a web of local regulations, and enforce IT security settings without being on the same network. And when they leave, you need to somehow retrieve their device and comply with employment law.

Disconnected employee data and workflows are a massive tax on productivity.

Given employee data lives across dozens of systems, point solutions cannot solve remote employment holistically – they solve symptoms, not root causes. Companies must either pay the administrative tax to support a remote team, or miss out on top global talent.

The administrative crisis is a fundamental bottleneck to company growth.

Rippling: ambition makes a comeback

Rippling is the ambitious solution to the nation’s sprawling administrative crisis.

Its strategy starts with the employee record, contextualizing each employee within the company. As the source of truth for employee data, Rippling can solve the administrative crisis holistically across software systems.

Questions like "what software systems can this person access?", "what tax regimes must we comply with?", and “how should we compensate this person?" all require the employee record that Rippling owns.

Employee data encompasses more than a username and password: start dates, compensation, role, function, group within the company, and even third-party app credentials all give context on the employee. In the same way that Salesforce deeply understands external customer relationship data, Rippling understands internal employee data.

At first glance, it is easy to confuse Rippling for a payroll system like Gusto. That is a category error – Rippling checks the HR system boxes, but the overall platform cannot be described by a single software category: it is a series of applications and workflows powered by employee data.

For example, employers can onboard employees remotely, provision their devices and apps, and manage their taxes in one place. With deep integrations into the business software stack, Rippling controls all administrative logic.

Many software companies are started opportunistically, so their product roadmaps are evolutionary. But the CEO and co-founder Parker Conrad built Rippling through intelligent design. Each product has a whitepaper, each funding round has a written memo. Releasing these memos publicly means Rippling is unafraid of competition: if you give your product roadmap to your competitors, you best be able to outpace them on building.

Rippling’s ambition permeates every aspect of the company, from engineering and distribution, to culture and terminal value. The engineering velocity is much faster than most startups. The product covers far more surface area. A broader platform means more potential buyers. 

Rejecting incrementalism has downsides. Ignoring SaaS playbooks makes the business harder for investors to evaluate using traditional methods. Running the culture at a fast pace could make employee retention challenging. Avoiding HRIS categorization could make it harder to position the product. Building a massively complex platform requires years of engineering work and a big balance sheet.

But if Rippling succeeds, it unblocks company scaling, freeing smart people to work on hard problems. Everyone assumes SaaS companies are fundamentally unambitious, while Rippling is hiding in plain sight.

Subscribe now

Product: breadth and depth

The relative ease of building point solution SaaS products drove hordes of new startups to saturate every niche of business system software, optimizing for depth over breadth.

But Rippling was not looking to build a software company. It was looking to solve the employee management problem, requiring coordination across employee data silos. Parker started with the employee data asset: by building on top of the employee record, Rippling could develop a compound product quickly. The product’s ambitious surface area enabled Rippling to quickly scale to tens of millions in ARR.

Rippling also realized that choosing between breadth or depth is a false dichotomy.

There is a set of common infrastructure needed for many business applications. A typical Series C software roadmap is largely predictable middleware: permissioning, reporting, approvals, compliance. Every point solution company has to build these capabilities from scratch.

Rippling abstracted this common infrastructure into Unity: a set of middleware capabilities on top of its employee graph. Rippling only needs to build middleware once, and then can amortize the investment across all its modules, freeing engineers to work on new product functionality. As Unity gets stronger, the entire product suite becomes deeper. Breadth and depth don’t have to be zero-sum.

A common middleware layer on top of employee data empowers not only internal product teams, but also customers to build complex workflows on top of their employee data that go well beyond traditional HR use cases.

These workflows are otherwise nearly impossible to execute: for example, Rippling customers can configure custom workflows that allow salespeople to give customer pricing discounts based on their level within the org. Or ingest Zendesk tickets to see payroll costs per support ticket resolved. Or run reports across silos of business data.

The breadth of use cases requires a large engineering team. SaaS playbooks focus on efficiency metrics like maintaining a 1:1 ratio of NNARR to burn – this inadvertently discourages ambition, trading off long-term upside for a narrow definition of efficiency. Rippling, on the other hand, has R&D spend that is off the charts compared to other companies of its size.

Building a unified middleware layer reflects the scope of Rippling’s ambition. But it is a capital intensive endeavor, requiring an unusually large engineering team. This is a challenge for the company: it must simultaneously fund a massive engineering endeavor and translate the investment into operational leverage. And because Rippling touches so many systems, it has higher risk impact: if its infrastructure breaks or underperforms, customers suffer more deeply.

Distribution: the rebundling era

Accumulating advantages are rare in software. You can increase brand and product depth, but things get harder as startups scale: as your customer acquisition engine saturates the core market, incremental customers become more expensive to acquire and likelier to churn. 

CAC increases, retention declines, and product development slows.

Bundling is Rippling’s accumulating advantage that offsets classical diseconomies of scale. This means that, unlike most growth-stage companies, Rippling’s core metrics improve over time. CAC paybacks decrease. Growth accelerates. Retention improves. Cross-sell rates increase.

It is common knowledge that companies build product bundles over time. But this is mostly out of convenience – it is easy to add new SKUs to sell into a captive audience. Rippling’s bundle is core to its strategy from day one, so the effects ripple across the entire business.

Of course, an expansive product increases contract sizes, and thus LTV. But a bundled product also drives sales and marketing efficiencies. Entirely new cross-functional use cases emerge to market the overall suite – for example, onboarding new hires touches SSO, payroll, and benefits products. And more product functionality increases the likelihood of resonating with different buyer personas. These efficiencies lower CAC.

Bundling is a tailwind to sales conversations. Why pay $30 per employee per month for five different products ($150) if you can get a comparable offering in a single bundle for $50? And companies with smaller budgets can buy additional products that were out of reach as point solutions. This is one of the few true accumulating advantages in software – amortizing S&M across several products means you can charge less while making more.

Having many SKUs also fuels net dollar retention well above the industry average. A multi-product suite enables piecemeal expansion: companies can start with one or two products, then buy more over time.

Once customers know you’re a multi-trick pony, they’ll start looking to you for new product offerings. Today, Rippling’s product launches are nearly guaranteed to succeed given its existing distribution and customer trust. New offerings have revenue and engagement metrics within months that most startups would be excited to reach in years. The extreme version of this is Microsoft Azure: cross-sell was how it built the fastest growing b2b business of all time.

The proliferation of software products over the past five years suggests that there will be thousands of multi-$B software companies in a decade, but the gravity of bundling will pull software towards consolidation. An expansive product portfolio should lead to higher terminal margins, as Microsoft has demonstrated, suggesting the end-game of software companies will have a stark power law distribution.

Bundling comes with risks. Because Rippling is bigger than a single software market, it doesn’t neatly fit into each of its product’s Gartner categories, making it harder to find customers. The product takes time for users to understand, so those looking for a quick evaluation may get lost. And vertical software players may be able to cater to customers in their respective vertical more acutely than a horizontal player like Rippling.

Subscribe now

Culture: pushing the limits of the possible

In the war for tech talent, many companies are competing aggressively on employee benefits, and even the number of work days in the week. Rippling’s ambition demands excellence over balance, which is not for everyone. The company’s core value is to “push the limits of the possible”, a refreshing contrast to the muted values of modern-day startups.

Rippling’s ambitious culture starts with its founder Parker Conrad, whose idiosyncrasies reflect a relentless focus on product.

His job title on LinkedIn is "Customer Support" – he still personally responds to support tickets as a way to get closer to the customer and improve the product. His desk is squarely in the middle of product and engineering.

Parker also runs Rippling’s Rippling instance. He manages payroll, approves every hire, and makes administrative changes in the system. He has personally hired more than a thousand employees through the system. The product roadmap writes itself when it is solving your own problems.

Unconventional for a company of Rippling’s size, Parker has no executive assistant. This means you can’t simply slide onto his calendar. The result is that he spends far less of his time in meetings, and far more on product, than the average founder. A compound product requires undivided attention.

Parker’s distaste for menial tasks led to expensive lessons at Zenefits, his first HRIS startup. But the positive externality of Zenefits is that Parker learned from his mistakes and earned a chip on his shoulder. He hired a general counsel at Rippling within a year of launch to ensure they’re running a tight ship, reducing cultural risk.

It would have been hard to devise the employee data strategy without a deep understanding of employee management. With the learnings from Zenefits, Parker crafted Rippling with higher intentionality and caution.

Rippling’s expansive product strategy requires a unique modular org structure. Each product has a relatively autonomous GM and engineering team. Given each sub-product has massive scope, they’re like startups within a bigger startup. In fact, Rippling has over 50 former founders within the company, many of whom are running various product divisions.

Each micro-startup gets to leverage the technical infrastructure, distribution engine, and existing customer base of a large company. But the GM model requires complex coordination of resources across the company – new product leaders need to push the marketing team to highlight their capabilities, train the sales team on selling their product, and work with the infra team to ensure the backend supports their needs. This can get complicated with several GMs.

Of course, it should be no surprise that Rippling is ambitious in its marketing, too. It turned a cease and desist letter from Gusto into a viral campaign by writing its legal response in poem form. Rippling even flipped a customer support latency problem on its head via radical transparency, turning it into a marketing benefit:

The scope of Rippling’s ambition translates into a workplace that runs at a faster pace than most startups today. When employees ask Parker for meetings, he is known to show up at their desk to confront the issues in real-time.

Many tech employees would be surprised by that intensity – it could drive higher attrition. With high demand for tech employees, Rippling will need to fight to retain its talent.

TAM: more than HRIS or payroll

If you let others define your product category and follow accepted playbooks, you’re playing the game on others’ terms, including their market size.

Building a platform around employee data enables Rippling to transcend the traditional point solution market analysis. Because Rippling is not a singular product, its market size cannot be easily retrieved from an industry report.

Most narrowly, in HRIS, Rippling could steal market share from incumbents including ADP, Paychex, and Workday, representing over $40b in combined revenue. There is a chance this overestimates Rippling’s addressable market: industry precedent suggests it is challenging to serve customers across company sizes: upmarket HRIS is a complex product lift. Though it has already started to prove it can: several 1000+ employee companies use Rippling today (including Rippling itself). Other startups like Gusto churned off their own products years ago.

But I’d argue that traditional TAM math underestimates the Rippling opportunity. Singular TAM analysis does not do justice to multi-product companies. Salesforce’s CRM, for example, is now less than half its revenue. Its customer service software and marketing cloud far outweigh the original customer management business – the CRM TAM is too constraining for them. Similarly, Rippling’s adjacencies like SSO, international payroll, software provisioning, and workflow automation massively expand their market size beyond HRIS.

Rippling also controls software distribution as the provisioner of end-user licenses, giving it leverage over the business software ecosystem. It is hard to quantify the value capture from being the software distribution layer, but it certainly has been valuable for players in other arenas like Apple and Google.

If Rippling becomes the system of record for employee data, it will rewire the entire business software landscape. With a breadth of software integrations, Rippling can become the connective tissue across traditionally siloed applications.

Conclusion

SaaS wisdom teaches us narrow lessons: start by conquering a narrow market and expand from there, add at least as much in ARR as you burn each year, and build your product either horizontally or vertically.

Rippling rejects these norms entirely. It ignores the mimetic warfare of the Gartner quadrant. Its wedge was a data asset, not a narrow product. It doesn’t let SaaS metrics dictate its strategy.

Reminiscent of the foundational computing companies, Rippling shows what ambition in modern business software looks like. In the age of software incrementalism, Rippling is an anachronism.

With ambition comes volatility, simultaneously enabling outsized success and idiosyncratic risk.

An ambitious product strategy means an expansionary surface area, like the Roman Empire. It can become vast, but is expensive to build, there is far more territory to defend against competition, and you can burn employees out along the way.

But if Rippling works, it shrinks the administrative drag on American business.

Administration drives process and conformity. Less time on overhead, more time on creation.

Subscribe now

Disclaimer: Founders Fund invested in Rippling in 2020 and has a board seat. I’m writing this piece not as an investment note, but as a case study for founders.

Thank you to Napoleon Ta, Everett Randle, Melisa Tokmak, Lisa Wehden, Axel Ericsson, Philip Clark, Sam Wolfe, and Pranav Singhvi for their thoughts and feedback on this article.

1. It is not a beloved consumer brand like Apple, Facebook, Amazon, or Google.

2. It was not a venture capital success story: Microsoft was too profitable to raise real VC money, so the founders owned 70% at IPO.

3. It is the oldest of FAMGA, hidden away in a different state.

But there is a lot more to Microsoft than meets the eye. If it plays its cards right, Microsoft can become the first $10T company. And startup founders would be wise to learn from the behemoth in Redmond.

This piece undertakes a daunting set of tasks: 1) understand what Microsoft is, 2) chart a path for its global domination, and 3) apply learnings from the company to the startup ecosystem.

Part 1: What is Microsoft?

Even to avid Silicon Valley historians, Microsoft is hard to define succinctly. No singular power law product defines Microsoft like Google’s Search, Apple’s iPhone, Amazon’s e-commerce, or Facebook’s social network.

Understanding Microsoft’s hundreds of products is daunting. With historical context, we can learn what Microsoft was, in order to discover what it is today.

1970s: the founding

In the early 1970s, most people thought personal computers were toys, relegated to hobbyist magazines like Popular Electronics. But Bill Gates and Paul Allen disagreed. As hardware rapidly improved, Gates and Allen realized that the bigger opportunity was not in building the next computer kit, but in building software to make these hobbyist devices accessible to a mass audience.

In 1975, Gates and Allen saw that the Altair 8800 was the first minicomputer with the hardware specs to support a BASIC software complement. They founded Microsoft that year with the founding mission of putting a computer in every home. As Gates described Microsoft: “What’s a microprocessor without it?”

2000s: diversification is for losers

Microsoft rode the PC growth wave for the next 25 years. In 1999, Microsoft was worth $620B, making it the most valuable company in the world. But by 2002, its market cap had dropped 60%. It was stuck in a massive hole dug by the inflated expectations of the dot com boom.

In the 2000s, Bill Gates’ successors took his mission to put a computer in every home too literally, expanding the Windows empire indiscriminately in every direction. Business lines proliferated, and Microsoft’s identity became murky. Was it the Windows company? Office? Xbox? Developer tools? By 2009, the company was worth 30% of its 1999 all-time high.

But diversification was not damning in itself, it was a symptom of a bigger problem: Microsoft’s growth was indexed to the wrong trend.

PC sales were plateauing. Apple, Amazon, Google, and Facebook were each indexed to mobile or consumer internet – both trends inflected in the late 2000s. Microsoft was indexed to neither.

Without a clear growth driver, Microsoft went on the defensive against the rest of FAMGA in the late 2000s: Bing, Skype, Surface, and Windows Phone were reactive moves.

2010s: finding the next wave

In hindsight, Microsoft’s defensive moves in consumer internet did not align with its DNA. But Microsoft has deep enterprise distribution and trust, something the other FAMGA members do not. Those advantages positioned Microsoft to capture the next wave: cloud infrastructure. Azure, Microsoft’s new cloud computing operating system, quickly became an offensive move.

Azure started in 2006 as an experimental project under Chief Software Architect Ray Ozzie. Guided by the Windows-everywhere dream, Ozzie pitched Azure as “Windows in the Cloud”, or a cloud computing operating system. But what customers actually wanted was an easy-to-use virtual machine in the cloud that Amazon’s EC2 delivered. By the time Azure became generally available in 2010, it had cobbled together product-market fit.

AWS’ four-year head start gave Azure a second mover advantage: it instantly had the accumulated learnings from AWS, which launched 4 years earlier. But unlike Amazon, Microsoft already had enterprise distribution. Once Microsoft started including Azure credits and consumption in its existing enterprise deals, the business took off.

By the late 2010s, Azure’s edge in enterprise distribution created a powerhouse, moving the needle for the entire company.

Azure reached a $10b run rate by early 2019, just nine years after launch. This is faster than AWS, which became generally available in 2006 and hit a $10b run rate by 2016, and GCP which launched in 2009 and hit $10b revenue in 2020.

Today, Azure is a monster with over $30b in revenue, and several $100m+ TCV deals in the Fortune 100. And while Microsoft doesn’t break out the Azure margins separately, we know it’s an attractive business at scale: AWS has 30% operating margins, driving the majority of Amazon’s net income.

In the 2020s, Azure will power most of the company’s growth. Azure gave Microsoft what it needed: a new wave to ride.

Is Azure the fastest B2B product of all time to reach $10 billion of revenue? Azure puts the fastest growing software companies of the past 25 years to shame:

The extraordinary growth of cloud infra businesses wasn’t tracked closely because 1) the key players weren’t venture backed or even standalone companies, and 2) GCP, AWS, Azure were all highly secretive about their early growth, lumping it into “other” revenue buckets to sandbag the breathtaking growth of the segment. The breathtaking cloud computing growth story has only become widely understood in the past couple of years.

Despite AWS brand dominance in cloud computing, Azure is a real threat. While AWS remains the market leader, it has maintained a flat ⅓ of cloud market share, whereas Microsoft doubled since 2017 and is growing faster than both AWS and GCP.

The fact that Microsoft was a PC company and Amazon was a web retailer shows that it almost didn't matter where you started. It matters that you rode the growth wave of the cloud. Microsoft may have too many internal competing interests to recognize that it has become the Azure company, but it should be clear by the late 2020s.

Satya vs. Gates

The MBA case study narrative gives Satya Nadella credit for reinventing Microsoft via Azure, particularly after the lackluster years under Steve Ballmer.

A better framework to understand Microsoft’s leadership is the Roman Empire. Naturally, Gates is Romulus. Nadella is closer to Augustus: undoubtedly a great emperor, but not a founder. Calling him visionary may be too generous, but at the scale of the Roman Empire circa 27 BC, marginal differences in emperorship could have huge impact on civilizational outcomes.

Undoubtedly, Nadella has been an exceptional CEO. He is highly attuned to culture: his superpower was transforming Microsoft from a culture of “no” into a culture of “yes”. By unblocking execs across divisions, Microsoft reaccelerated by the late 2010s, from single digit revenue growth rates to high teens today.

While philanthropist Gates is in favor, business Gates has been out of favor since his notorious antitrust loss in 2000, United States vs. Microsoft Corp, when he got whacked for bundling Internet Explorer with Windows. But let’s not forget the empire he built: by the time he left the company, Microsoft had over $20b in revenue and millions of users.

Nadella inherited an iron man suit which now defines Microsoft: enterprise distribution, user trust, and engineering talent. Azure’s success is the result of a great leader taking advantage of a latent opportunity. Now that we know Microsoft is capable of riding new waves under Nadella, we can explore other latent opportunities to expand the Microsoft empire.

Subscribe now

Part 2: The path to $10T

With the right strategy and execution, Microsoft can become the first $10T company.

What would I do if I were running the company? First, I would take stock of its relative strengths: $130B in cash. $2.3T of market cap. Unrivaled distribution in every Fortune 5000 company. 96,000 talented engineers.

Microsoft has powerful intangible assets. Relative to the rest of FAMGA, it may be the only player with some antitrust immunity. This is somewhat structural: software markets are less monopolistic than consumer internet. Even its fastest-growing business line, Azure, only has 19% market share (AWS is at ~30%) – far below any legal definition of monopoly. With nearly unlimited cash and equity, plus some antitrust immunity, Microsoft should be acquisitive.

There is also a financial engineering “growth” strategy that will probably appease shareholders. Tim Cook, for example, started share buybacks in late 2012 after taking over as CEO in 2011, and now consistently buys back ~$20B in shares per quarter. Apple’s share price has grown by more than 10x since Cook took over. It is no surprise that Microsoft started aggressively ramping share buybacks in 2014, the year Nadella took over as CEO.

I would like to think that Microsoft has more compelling growth opportunities than share buybacks, though. We can’t do justice to the entirety of Microsoft’s product suite in 4,000 words, but on the path to $10 trillion, we can explore four new waves Microsoft can ride: demographics, data, developers, and depth.

Demographics expand Microsoft’s TAM

Microsoft’s products are ubiquitous in the Fortune 5000, but conspicuously absent in two growing segments: 1) young users, and 2) growing tech companies. Only 18% of Teams users are below age 35. These are the enterprise customers of the future. Microsoft failed to capture the self-serve motion that unlocked younger demographics and propelled many productivity decacorns in the past five years.

By acquiring companies with complementary demographics, Microsoft can recapture these segments. There is an entire swath of Office-style products with distribution among young + tech audiences, filling Microsoft’s gaps over the past decade:

• Airtable: Airtable is a lightweight CRM + Excel, with some automation on top. Microsoft would not need to acquire Airtable if it built out more Excel integrations and improves the UX, but it is trapped in an innovators’ dilemma and can’t alienate its existing Excel user base. A standalone complementary product has more degrees of freedom to capture the next generation of data entry users.

• Notion: To oversimplify Notion to its demographics, it is Office 365 for people below age 35. At a minimum, Office should aspire to the ease of use that Google Docs has (Office 365 for people below age 45), but shooting for the stars may help them to land in the clouds.

• Miro: This could represent an entirely new Office suite product line – digital whiteboarding. Every enterprise started digital whiteboarding during COVID, which is likely a permanent fixture in a remote and distributed world.

Acquiring these companies would capture the young productivity tool users that Microsoft needs to make Office successful in the 2020s. Given the distinct user experiences, it may be best to keep the products independent, but offer them as part of the core Office suite. This makes Office compelling to a wider audience, reduces its customer acquisition cost, and expands Microsoft’s addressable market.

Microsoft also needs to invest in demographics from the inside: young ex-startup employees could bolster a bottom-up approach to building products for young audiences. According to LinkedIn, Microsoft makes fewer recent grad hires in EPD compared to Facebook and Google. It also has far fewer young VPs and directors (<10 years of experience) than Facebook and Apple. In other words, there are relatively few young people in charge of things at Microsoft. Empowering the new guard can help its products flourish with Millennials and Gen Z.

Developers should love Azure

Azure is undoubtedly a monster business, but it is not cool among the next generation of developers. Without them, Azure will be stuck with top-down sales. In fact, Microsoft’s unpopularity with developers started in the 1990s: the Halloween documents from 1998 show Microsoft’s antagonism towards Linux and open source, which would haunt them for decades.

Microsoft realized this faux pas, and is making up for lost time: Nadella claimed that “Microsoft ❤️  Linux” in 2015, acquired GitHub in 2018 in a splashy $7.5B acquisition, and Azure has deep Linux support (half of Azure workloads run on Linux).

But owning GitHub doesn’t immediately translate into young developers that love Azure. Microsoft’s task is to gracefully harness its newfound developer love. But you can’t force love, like Google tries to force Meet on its unsuspecting users. Microsoft needs deeper OSS support and developer tooling before pushing users towards Azure.

In other words, Azure needs organic adoption to fully penetrate the enterprise. GitHub has developer credibility, but its primary usage is via its command-line tool, where pushing Azure is unnatural. Owning the developer experience at the IDE or Terminal is the path to cross-sell cloud infrastructure downstream.

Microsoft now has a next-gen IDE play through Codespaces, but IDEs are highly fragmented. Rolling up the long tail of IDEs is attractive – they tend to be weak standalone businesses, but increase developer mindshare and offer an on-ramp to Azure. Acquiring Replit is also worth exploring given its strong positioning among young developers who are still malleable.

Data strengthens Microsoft’s core products

Spreadsheets were the killer app for business computing, and Microsoft Excel ultimately dominated the market. It created a generation of data analysts and financial experts. For decades, Excel was the best-in-class business data store.

But in supporting the median, Microsoft lost the frontier. The modern productivity stack fills the gap of Microsoft Excel’s stalled innovation. Weak UX made room for Airtable and Asana. Weak data primitives made room for modern CRUD-based applications. Weak customization and integrations made room for vertical software companies. Weak collaboration made room for Google Sheets. Weak backend architecture and interoperability made room for the modern data stack. Instead of Excel realizing its platform potential, it remained simply a great application.

Microsoft has many pieces that can theoretically support the modern data analyst – data warehousing through Synapse, data pipelining through Data Factory, visualization through PowerBI. But modern teams don’t choose Microsoft, opting for the modern data stack of Fivetran + dbt + Snowflake. Is there a chance Microsoft can win them over?

Microsoft should pay up for best-in-class companies to make up for lost time in winning the business data race. I’d personally consider an acquisition in each step of ELT: Fivetran as the best-in-class E and L, and dbt as the corresponding T. This kills three birds with one stone: position Microsoft to win data pipelining with a leading product suite, inherit distribution to tens of thousands of the best data teams, and push them towards Azure.

Even within Microsoft’s existing product suite, there are massive opportunities for owning customer data. For one, Microsoft is one of the best-positioned companies to compete with Salesforce. Microsoft’s CRM product Dynamics is a sliver of what it could be. Microsoft has the most strategic asset in all of sales and marketing, LinkedIn, and the technological infrastructure to disrupt Salesforce.

Microsoft has a path to becoming the source of truth for customer data through Azure, which would make it mission-critical to all business software. If Microsoft convinces customers to store their customer data in an Azure warehouse (enriched by LinkedIn’s data) instead of a CRM, then the CRM becomes a simple window through which companies access customer data. Other business applications would then build on top of Azure, not Salesforce. If Microsoft divorces the CRM from the customer data system of record, it commoditizes its complement to beat Salesforce.

Depth is Microsoft’s path to winning new markets

Much of Microsoft’s history suggests that its defensive product lines are less successful: Bing lives in the shadow of Google, Surface lives in the shadow of iPad, Skype lives in the shadow of Zoom.

But these product lines were trying to compensate for Microsoft’s weaknesses in consumer internet, not its strengths in enterprise. When Microsoft leans into its enterprise distribution, even defensively, it wins.

If I were in charge of Microsoft’s product strategy, I’d seriously consider competing with any enterprise software company that reaches $100m in ARR. Given its lofty mission and AI-fueled vision, Microsoft may be too proud to clone simple SaaS products. But Facebook has shown that ruthless practicality in copying new startups’ products is a relatively successful strategy.

Enabling and encouraging entrepreneurship internally to build new product lines is challenging, but if Azure could catch up to AWS when it was already at hundreds of millions in revenue, then Microsoft can do it for other categories too. Each incremental product benefits from Microsoft’s distribution, so double-digit market share in these categories is nearly guaranteed.

At the risk of being unfashionably imperialistic: as the largest software company in the world, Microsoft should have a leading product in every meaningful business software category. It had this in the 1990s (Windows, Office, Access, et al.). In the 2010s, the surface area of business software expanded too quickly for Microsoft to keep up in every category.

Microsoft is playing catch up in most markets, so internal product development alone is not enough – it must be complemented by M&A. A few ideas come to mind:

• DocuSign: E-signatures are one of the most obvious secular trends in software, and DocuSign is the clear leader. An acquisition could be particularly accretive given Microsoft’s ability to embed the product across its productivity product line.

• Figma: Surprisingly, Microsoft has no answer to Adobe, despite being the largest SaaS company in the world. Figma captures both the demographics and depth opportunities.

• Zoom: Microsoft has Teams video calling, but Zoom has network effects and became the standard first among young users. This is an opportunity to win the web conferencing market and acquire a distinct audience from Microsoft’s current customer base.

These are not meant to be comprehensive examples. Microsoft should run a triaging exercise to build, buy, or ignore every company in the Bessemer Cloud Index. Plugging new software companies into the Microsoft distribution engine immediately turbocharges their growth, meaning that it can afford to pay above-market prices for great products. As Microsoft competes in new software markets, it expands its TAM, lowers its cost of sales, and gets a new wave to ride.

Gaming

Gaming doesn’t fit neatly into the strategy, so the path to $10T may not seem fully cohesive. But that is the real lesson: it doesn’t need to be. Microsoft is essentially a holding company. New acquisitions and products don’t need to be accretive holistically, but rather to a subsidiary.

In the gaming context, Blizzard helps build the network effects for Game Pass’ 25 million subscribers, and scale de-risks gaming development as a hits-driven business. Now Microsoft is the largest US gaming company – better for the US to own our gaming future instead of China. Gaming is seemingly distinct from its enterprise advantage, but Microsoft is a conglomerate.

$10T: Microsoft’s to lose

Given its market share for its core product lines is much lower than the rest of big tech, the path to $10T is much more in Microsoft’s hands than for other players. How much can Google really do to grow market share? It already won search. Microsoft has many growth levers to pull.

Becoming the first $10T company will require aggressive M&A and product development across categories, and some pride-swallowing to compete directly in emerging software markets. Microsoft has the capital, talent, and distribution to pull it off.

Growth at this scale is unprecedented, but a wealth of opportunities across its product suite suggests the path to $10T is Microsoft’s to lose.

Subscribe now

Part III: Lessons for startups

As the largest tech company in the world, Microsoft may not appear to have lessons applicable to startups. But the re-acceleration of the company in the 2010s and the path to $10T tell us about the power of S-curves, compound products, M&A, and second-mover advantages.

Market over execution

Execution is critical, but riding an S curve is the path to win in tech.

Steve Ballmer gets a lot of flack for Microsoft’s underperformance in the 2010s. Undoubtedly, he wasn’t as exceptional as Gates, and company culture stagnated. But the real challenge was category: PC sales were decelerating, while Microsoft was betting on their growth. Reorienting the company towards the growth of cloud infrastructure and business software (which Ballmer should get some credit for) gave Microsoft life again.

Traditional wisdom tells us that founders are the only determinant of startup success. Great founders are necessary but not sufficient. Great product theses and fast-growing categories are increasingly the true bottleneck.

Be acquisitive

Be acquisitive when you have a lot of cash and equity. Nadella seems to appreciate this given his relatively aggressive M&A track record. The path to $10T will require an acceleration of acquisition pace: even the $68.7B Blizzard acquisition represents a mere 3% of Microsoft’s market cap.

As the surface area of new software markets plateaus, tech will transition towards a consolidation era. Startup M&A will become a core part of growth – think of the rollup eras for oil, telecom, and cable. Late-stage startups that are valuation-rich should be considering M&A much more aggressively.

Compound products win

In the 2010s, unbundling made sense because software adoption outpaced the ability for unified product companies like Microsoft to build software. If that pace is flipped, do we see rebundling?

The classic consideration for VCs is whether incumbents can copy the startup’s technology before the startup copies the incumbents' distribution. For the past 20 years, the answer was almost always no – startups achieved escape velocity across categories, seemingly immune from incumbents’ distribution power.

But don’t let that fool you: there is a real bundling effect in software. See Parker Conrad’s Compound Startup thesis, playing out in real-time via Rippling. Office 365 has its weaknesses but is a truly compound product – Microsoft Teams, for example, outpaced Silicon Valley darling Slack just 3 years after launch.

The power of bundling may not have been obvious in the past 20 years because there was so much surface area – every point solution had a high ceiling before incumbents copied it. But the strength of bundling will become a lot more obvious going forward. Tool fatigue and bias towards simplicity are increasingly powerful forces.

The power of cross-sell

Microsoft already has distribution into every Fortune 1000 IT department. This makes it much easier to sell new products than from a cold start. A myriad product line suggests a lack of focus, yet Microsoft is more profitable than all of FAANG – that is the power of cross-sell.

Marc Benioff already understands this deeply: Salesforce has more revenue today in customer service than in its core sales CRM product, and is now tackling marketing and analytics.

Silicon Valley wisdom says to focus on a single product and market. Ballmer did the opposite and built products across categories. This was the right strategy, but in the wrong decade: cross-sell positions Microsoft uniquely well for the 2020s, when antitrust is more threatening than ever. I would rather command 30% market share across many categories than be regulated to death as a monopoly.

Definitionally, cross-sell benefits incumbents more than startups. But it also informs startup strategy: if you have lower CAC on selling incremental products to your existing customer base, it is advantageous to build a deep product suite before selling horizontally.

First mover advantage is overrated

Azure is a real competitive threat to AWS, despite being four years behind. We’ve seen this in the startup context too: Facebook surpassed Myspace and Friendster, Ramp is now a real threat to Brex, Modern Health is a real threat to Lyra. Second movers short-circuit the learning curve of a new market.

For first movers, the lesson is clear: don’t rest on your laurels. But this should also be encouraging for second movers: there is probably more room for new entrants than you think. Think of the degree of competition in other industries like retail or finance – tech has lots of room for new companies before we reach a saturation point.

Consumption-based pricing is a bet on yourself

Consumption-based pricing is the business model that is most aligned with customers. The opposite is simply a poor customer experience: why pay for something before you receive it, or if you may not use it at all? Unsurprisingly, many of the fastest growing companies have some version of consumption-based pricing: AWS, Azure, GCP, Snowflake, Twilio, Scale.

It may feel less “safe” than pure SaaS because the customers are not guaranteed to renew, but it lowers the barrier to adoption. Consumption-based pricing is the best way to bet on the quality of your own product: it perfectly aligns product, customer success, and sales. If customers are engaging with your product, it is a win-win.

Capital: a true moat

The tech industry doesn’t talk about capital as a moat, probably because it benefits those that have already made it. But Azure proves that it works: it spent billions to achieve economies of scale, but the prize is a comparably massive profit center.

In recent startup history, massive capital scale often has failed to change the trajectory of companies: think of the SoftBank mega-rounds that went sideways. Even in the case of Uber, a relatively good company with economies of scale, capital hardly helped: the equity has been roughly flat for the past seven years.

Few companies have the competence to ingest truly scaled capital. When it works, it can really work. The $535m DoorDash Series C, one of the first Silicon Valley mega-rounds, was an extraordinary case study: it quickly won over 50% market share, and was poised to capture the food delivery market expansion during COVID.

Megaproject success can be hard to see when it happens within large companies. Azure and AWS are two of the most successful megaprojects of this century, but were hidden from the public for years, nested inside much larger corporations. Starlink, the global satellite internet network, is only possible given the scale of SpaceX’s core launch business, but could be one of the most successful megaprojects of our time.

Conclusion

As a student of Microsoft, 4,000 words is woefully insufficient to fully digest its complexity. Everyone in tech should spend their own time trying to learn from the greats, and Microsoft deserves special attention.

So what is Microsoft? It was the Windows company in the 2000s, became the Office company in the 2010s, and is becoming the cloud company in the 2020s. It is the sum of its core advantages: enterprise distribution, user trust, and an engineering talent vortex.

With these advantages on its side, my money is on Microsoft as the first $10T company. Azure alone has a path to hundreds of billions in revenue, and would be one of the largest standalone companies in the world today. Getting to $1T in revenue will require ruthless expansion across product and M&A in every software market.

And in the startup context, Microsoft is a treasure trove of lessons: it teaches us the power of distribution, product bundling, M&A, and compound growth.

Subscribe now

Thank you to Brandon Camhi, Pranav Singhvi, Sam Wolfe, Harry Elliott, Everett Randle, Melisa Tokmak, James McGillicuddy, and Tomcat Sanderson for their thoughts and feedback on this article.

Yet as active managers underperformed the index 80% of the time in the coming decades, the argument favoring passive funds became a no-brainer: why pay higher fees for worse performance?

Even Warren Buffett, famous for his concentrated portfolio, now tells us to simply invest in index funds. Correspondingly, public index funds have doubled their market share in the last decade – a multi-trillion dollar transition from active to passive assets.

This pro-index tendency pervades the private tech markets, startups, and even our culture through what I call the index mindset: a focus on preservation over creation, optionality over decisiveness, general over specific. Public companies are an obvious thing to index, but the index mindset manifests in many domains:

Internet and software companies are far less risky than they used to be, even at the early stages: there hasn’t been a venture vintage since 2002 with negative median returns. Big tech is now huge tech. Risk-averse people and money have flooded in. When people have something to lose, they protect their downside – think of wealth managers, encouraging a “safe” mix of stocks and bonds. The tech industry has too much to lose.

Indexing may be the correct default for public investors, but can be dangerous when replicated in other domains. The public markets show us the second-order effects of indexing, so we can learn how it affects the private markets, startups, and culture.

Indexing in the public markets

Given the index mindset has snowballed the longest in the public markets, we can see that when the plurality of investors are passive, the implications are not passive at all.

Record-low interest rates since the 1980s enticed retirement money to flow into the US stock market. With low fees and returns that were hard to beat, index funds became the dominant sub-strategy, now constituting 14% of the US stock market.

Everyone knows that “past performance is not indicative of future results”, but nobody believes it. In the case of index funds, it has been indicative. Index funds are a reflexive asset: because US funds have attained consistent returns, retirement money pours in, driving asset inflation and more buyers.

In the public markets, investors can no longer solely rely on the DCF model of investing. As index fund capital flows approach 50% of the stock market, the indices themselves distort asset prices. It is important to ask ourselves, why should a low-conviction index bet generate any return at all? There must be some external force: that’s retirement savings pouring into the assets behind you. As index funds became a generational asset class, their returns were driven as much by demographic trends as by the underlying companies’ fundamentals.

The common intuition is that when everyone zigs, you zag... but index funds make another zig the optimal strategy. There is an uncomfortable pyramid scheme dynamic to capital flows: your investment performance is aided by the volume of capital behind you. The retirement savings capital inflows could revert, but are alive and well for now.

The importance of capital flows back-propagates into privates, too. Blackrock, Vanguard, and State Street are the largest buyers of public companies. Given every private investor underwrites valuations based on what the next round’s investors will pay, all private investors are beholden to the IPO buyers as the Series N+1 investors.

When capital inflows increase for private markets, it reinforces the indexing approach, in the same way that public index funds become reflexively good.

Indexing in venture capital

The index mindset is more obvious in venture capital than in any other asset class. Tiger Global, the private tech index fund, dominates VC partner meeting discussions. Those VC firms themselves are scaling and diversifying. And the increasingly popular venture capital job itself is the ultimate indexed career, a bet on a basket of companies.

Growth investors often use a framework of how quickly an investment will be “in the money”; in other words, how quickly another growth investor will pay more for it. This is implicitly a bet on capital flows more than intrinsic company value. And when everyone thinks this way, momentum becomes self-fulfilling. Capital flows govern the public markets, where index funds prevail. This is increasingly true in venture capital.

Venture fund indices prioritize momentum over value, such that fundamental business characteristics matter less: efficiency metrics and company-level profitability give way to growth. “Checkbox” investment strategies from the last decade won, which was in essence frontrunning future index fund algorithms. 

Public macro investor Michael Green speculates that index funds weaken mean reversion characteristics: overvalued things should be purchased less, but market cap-weighted index funds purchase them more. We’re already seeing this in growth stage investing: companies that check enough boxes are purchased with increasing price agnosticism.

To make outsized returns in venture capital, you need to believe in companies when others don’t. But for the last decade, many investors who claimed to be VCs were actually index investors chasing a checklist of 3x+ growth, 100%+ NDR, 70%+ margins, and recurring revenue. These investors are commoditizing overnight, just like stock brokers on Wall Street in the 1980s.

Most seed-stage investors have a high-volume investment approach: the conventional wisdom is that “you can’t know which company will return your fund, so take a lot of shots.” This could accelerate significantly into a true seed index that aims for high market share and deployment pace: a scaled seed index fund still doesn’t exist. But things start to break down at this stage: seed investors often act as advisors to fledgling companies, which doesn’t scale.

VC is commoditizing more broadly – finance types have flooded in, growth funds are raised and deployed far more quickly, investment teams have scaled, and deal-level competition has intensified. It is no longer the boutique asset class that small funds were betting on. You need scale to make an indexing strategy work – this could drive a consolidation era of venture capital, with far fewer funds.

Indexing in startups

The index mindset has pervaded startups, too. Employee tenure is short and shortening, with many employees opting to collect a portfolio of equity across several companies, hedging their downside along the way.

The proxy war of the index mindset in startups is quadratic vesting: employers are trying to combat micro-tenures with back-weighted vesting for employee equity. This is partially a function of scale: tech is now a place where normal people come to build careers, not just nerds tinkering around.

In day-to-day strategic trade offs, the index mindset also reigns. In marketing, conventional wisdom says to eventually build a portfolio of distribution channels to prevent concentration risk, instead of focusing on your best channel. In sales, it’s generally preferable to have lots of small customers versus a few whales. In positioning, startups try to check as many boxes for as many customers as possible, instead of resonating deeply with a narrow audience.

Even founders, who are on the highly concentrated end of the indexing spectrum, have started to diversify. Secondary sales happen earlier and are more common than ever before. Many founders have side hustles as angel investors, or even as part-time fund managers.

The limits of the index model are clear when it comes to company formation. In a world where people speak in probabilities, it is easy to forget that great startup ideas and teams stem from intelligent design, not evolution.

Founders can’t afford to index their market: a seed round only allows for a handful of product bets, particularly in a mature ecosystem with immense company-level competition. As Keith Rabois put it, “product market fit is forged, not discovered”.

Indexing in culture

We can also see the rise of the index mindset in culture, in the form of diversification and risk avoidance.

In philosophy, we’ve seen a shift from a singular truth in moral absolutism to a multitude of truths within moral relativism. In physics, a shift from a deterministic Newtonian physics to probabilistic quantum mechanics.

In politics, globalism swiftly replaced nationalism in the political zeitgeist post-World War II. This is the easiest to see vis a vis the establishment of the European Union and the Euro, a way for the member states to bet on an index of other countries instead of themselves.

Private life has also become indexed. Just a few decades ago, marriage at a young age was common, with one partner assigned for life. Today, Tinder allows people to diversify their relationship capital across an index of dozens of partners.

Even career choices have become indexed and hedged. Nearly 40% of Harvard grads work in banking or consulting after graduation – choosing a basket of potential career opportunities without committing themselves to one. And the projects within consulting and finance jobs are indexed across many industries and disciplines.

Parents encourage the index mindset, too, disguised as well-roundedness: a collection of hobbies and extracurricular activities that make their children broadly appealing to academia and industry.

When America was founded, it was a frontier. Each citizen had minimal possessions. Over the centuries, Americans have accumulated financial and relationship capital, giving themselves something to lose. This wealth went parabolic in the 1960s, and since then, we’ve become afraid of taking risks: marriage comes six years later, schooling lasts 5 years longer.

The index mindset is comfortable – avoiding decisions requires the least amount of effort. But if you index across every domain, you lose your differentiating features, becoming an average of everyone else.

Subscribe now

Indexing in the 2020s

The index mindset emerges in any highly liquid market, and technology supercharged liquidity in many markets: financial markets, job markets, dating markets. Regardless of our participation, the index mindset is here to stay.

The biggest consequence in tech is already here: ecosystem-level inflation as money floods into tech indices. Investors must pay higher valuations. Startups must pay employees more. Sales and marketing teams must pay more to acquire customers. Private tech inflation will only accelerate. If opex inflation outpaces pricing (revenue) inflation, then tech could become far less profitable.

Other consequences like fragile valuations are harder to see. Employees should be wary of assessing startup equity in the context of index fund valuations. VCs should be careful about how price changes represent fundamental business changes; maybe their markups should be reported on a volume-weighted basis. Venture backing doesn’t qualify companies like it did in the early 2010s.

As the world shifts towards passivity, you have a personal choice: join the index, or avoid it entirely.

Adopting the index mindset is the easiest default. You don’t have to think too hard. And it often works, if you get the highest-order bit correct: indexing the tech sector in 2010 was a great decision. It can also be useful as a jumping off point; purchasing a basket of crypto or biotech stocks, for example, may encourage you to learn more about the technological frontier.

In the investing context, you can frontrun the index: compete on speed, think about where the index is going, and try to get there first. Maybe you invest in companies that are ESG friendly, suspecting they’ll earn a premium in the index moving forward. Or you invest in a lot of early-stage startups before they enter the index.

Trusting the index fully can be dangerous. Index funds are reminiscent of the CDOs of 2008: don’t worry about the underlying assets, we were told, and trust ETF algorithms and tax-loss harvesting to optimize returns. When nobody pays attention to the details, things can go wrong. The index may hold a lot of junk under the surface.

Many of the most successful individuals have thoroughly rejected the index mindset. Elon Musk, for example, isn’t spreading his bets – in fact, he’s buying more of his companies’ shares. And Warren Buffett, despite preaching index funds, has 75% portfolio concentration in five stocks.

Indexing provides a safety net against failure, so it can be hard to reject. But you can start small. Every index has underlying assets that are below average, so you can start by identifying which components are weighing it down. This could be empty relationships, weak investments, or unfulfilling hobbies. Over time, narrow the index into a handful of bets, each of which you can staunchly defend.

Abandoning the index mindset may be more valuable than ever. When everyone is indexing, their collective trance distorts reality. That’s your opportunity. There isn’t a singular answer about when to adopt versus abandon the index. But the index mindset is sedative, a substitute for conviction. And conviction pays off, as Howard Marks succinctly put it: “if you wait at a bus stop long enough, you're sure to catch your bus, while if you keep wandering all over the bus route, you may miss them all.”

Subscribe now

Thank you to Luke Constable, Brandon Camhi, Harry Elliott, Everett Randle, Sam Wolfe, Melisa Tokmak, Josephine Chen, and Axel Ericsson for their thoughts and feedback on this article.

In practice, what does financialization look like? We’ve all heard the institutional narrative. The US stock market has nearly tripled in the past 10 years. The FIRE sector (finance, insurance, real estate) now accounts for 20% of US GDP, versus 10% in 1947. S&P buybacks nearly doubled in the 2010s. IPOs grew 2.5x in 2020.

But financialization is no longer purely institutional; it has seeped into our culture. A combination of low interest rates, a historic tech bull run, and the resulting torrent of fomo has tethered us to our monitors to watch candlestick charts. The financialization of culture has manifested in two primary ways: lottery culture and equity culture.

Lottery culture

We’ve always had lottery culture, in which people trade assets hoping to make money without understanding or conviction of their fundamental value. In the summer of 1929, for example, Joe Kennedy’s shoe shine boy gave him stock tips, signaling the market peak.

But lottery culture has exploded in the past few years. Robinhood has achieved its vision of democratizing access to financial markets, boasting over ten million monthly active users and exponential growth. Retail investor trading as a percentage of stock market volume has more than doubled since 2010. When “stocks only go up”, people realize the stock market is a casino with much better payouts. Gamified trading has tethered us to our phones and bank accounts.

Through GameStop, even protest became financialized. At the peak of the GameStop saga, millions of Americans owned GME, arming themselves against hedge funds. This was part protest, part nihilistic hail mary to get rich. The number of subscribers to the wallstreetbets (WSB) subreddit certainly spiked in 2021, but the exponential compounding for years beforehand implies a degree of inevitability:

Memes took lottery culture to new heights. Stocks popular among a retail audience, like Apple, have historically traded at higher multiples than others in their category. Tesla accelerated the divergence between retail excitement and fundamentals: TSLA revenue grew 50% over the past two years, but meme culture helped its market cap grow by 12x. GameStop completed the meme-fundamentals duality: investors don’t even pretend that the company’s fundamentals will ever substantiate its market cap.

WSB revealed where Robinhood culture was headed all along: a nihilistic lottery. It was historically viewed as a disorganized and degenerate mob, but the mob has evolved into a laser-focused and motivated militia.

Even outside of WSB, another clear sign of lottery culture is the focus on technicals over fundamentals. Many “smart money” investors don’t rationalize investments by talking about the fundamentals of the business, or even conviction in its future. Instead, we’ve seen the emergence of chartism, which advocates for investing not based on where the business is today or where it’s headed, but its likeness to a “cup and handle” or “head and shoulders” pattern. This is the astrology of finance.

For those who want more risk than public equity markets provide, a long tail of speculative assets are exploding – digital collectibles, DeFi, Bitcoin. Bitcoin grew more than 5x YoY. Collectibles like NBA Top Shot did $60m in sales in a week. DeFi (decentralized finance), the infrastructure powering the casino of crypto, had a $40b influx in the past few months. Dogecoin is worth more than $6b.

Speculation now dominates culture.

Equity culture

Equity has been a great mechanism to align labor and capital. As tech equity has gone parabolic, there is now a prevailing sentiment that equity holders are getting rich, and you’re not. In the tech context, this means that startup jobs are viewed as an offer of equity, with four years of indentured servitude attached. To be a software engineer in 2021 is to be a financier.

It wasn’t always this way. In 1970, a young engineer named Mike Markkula was offered 1,000 options to join Intel, their standard grant size. He then asked for 20,000 options, and the company agreed. Very few employees – and companies – appreciated the value of equity at the time.

Today, it is hard not to notice the prevalence of equity culture. Every employee optimizes their stocks and options. People frequently move between companies to build a portfolio of equities. Employees of successful startups want to become angel investors. Second-year MBAs want to be founders or VCs. Nobody wants to miss out on the next equity success.

Equity culture has become labor emulating capital. People aggressively negotiate and hold equity more than ever, seen as the golden ticket out of the tyranny of 9-5.

How did we get here?

The combined effect of equity and lottery cultures is omnipresent. And there’s a reflexive loop between equity culture and lottery culture: lottery culture creates the perception of easy wealth creation, and equity culture responds with more aggressive buying and holding, lowering supply and pushing up asset prices.

2020 accelerated finance culture: stimulus checks, nothing to spend money on, and lockdown boredom drove people to the stock market. But the driving forces behind the financialization of culture have been accumulating for years. We need to look at interest rates, tech maturity, inequality, and social media to understand the full picture.

Low interest rates

Interest rates don’t tell the full story, but they tell a lot of it.

The value of a financial asset is classically calculated as the cash it is expected to generate, discounted into the future – this discount rate is closely tied to interest rates. The present value of a business is the multiplicative inverse of the discount rate: as the discount rate approaches zero, the present value of equity climbs exponentially.

Take a company generating $1m in annual cash flow, growing 30% per year with a 15% growth rate decay. We can then calculate the net present value of the next 50 years of cash flows. If we use a 2% discount rate instead of 10%, the present value is 5x higher:

Importantly, low discount rates most heavily impact the valuations for companies with cash flows far in the future. Nikola, for example, has practically no income but promises cash flows on a multi-decade timescale, earning a $7b market cap. A 1% change in interest rates makes a one-year bond change in value by ~1%, but could change Nikola’s value by 20% or more. Stated differently, if you don’t discount cash flows, you believe the future is all but certain.

But there’s a principal-agent problem: by giving Nikola the valuation it deserves in 50 years, a disconnect emerges between current equity holders (principal) and the employees that will eventually need to build the trucks (agent). If I’m an employee of Nikola and my stock is now worth what I was hoping it’d be worth in 10 years, what incentive do I have to stick around another decade to build the trucks?

In the tech context, it’s becoming more common to see founders selling shares before the company reaches escape velocity, and employees rotating between hot companies before they have a chance to own projects end-to-end. Too much capital makes it more compelling to chase easy financial gains than build new tunnels.

When interest rates decline, businesses take more risk, which is good for innovation. But as rates approach zero, everyone with a high payout expected far into the future suddenly drowns in money, perversely distracting from innovation. This leads to a counterintuitive relationship between interest rates and innovation:

Tech maturity

Financialization is a byproduct of the narrowing of the visions: there are more engineers working at Dropbox today than at Apple when they released the Apple II. The explosion of tech wealth and headcount translates into many smart ambitious people doing tech with the expectation of transitioning from labor to capital.

When reality is an S-curve but expectations are exponential, the gap between the two curves drives financialization as investors find new ways to crowd into the same trends. The wider the gap, the more crowded it feels.

As “naturally aspirated” internet growth slows, investors need to take more risk to generate the same returns. This is why we see a rotation from “safe” tech assets into far riskier alternative assets – Bitcoin, DeFi, early-stage startups.

I often hear comments that the tech industry is becoming financialized at the corporate and individual levels. But financialization is natural as an innovation cycle matures, and was amplified as the internet became far more lucrative than anyone expected. Simultaneously, tech is taking over pop culture – the real story behind Clubhouse is that SV is now a major cultural exporter. The product of the financialization of tech and the techification of culture is pervasive finance culture.

While it may seem like the party is over, the plateau of the internet is a good thing for technological innovation. As late stage and public internet companies become fully valued, investment dollars shift further down the risk curve into new technologies.

Inequality

Labor’s share of income is declining across countries; in the US, it fell from >63% in 2000 to <57% today. COVID accelerated this dislocation between labor and capital. The further the disconnect between labor and capital, the more resentment we see between them. GameStop was the greatest invasion of labor into capital since the Peasants’ Revolt – a financialized flavor of class warfare.

The Gini coefficient is a classical measure of inequality, and it did increase from 0.43 in 1990 to 0.48 in 2020. But the inequality is far clearer on industry lines: the NASDAQ has generated 2.5x better returns than the S&P 500 since 2008. When the growth of tech’s contribution to GDP far outpaces other industries, but its total employment stays roughly flat, it is hard not to see the rest of the country missing out.

As people drown in record student debt, the appeal of a lottery ticket win as a way out is stronger than ever. The easiest way to emulate the financial success of the tech sector is to get equity in tech companies. Robinhood has accelerated this process in the public markets. In the private markets, employees are far more clued into the equity equation, aggressively negotiating on options since they now viscerally understand this is the way tech money is made.

Social media cultural amplifier

Social media makes income inequality noticeable in a way it never was before. In the 70s, the mega-wealthy stayed to themselves in Greenwich and Monaco. The only way the masses saw wealthy lifestyles was through the Hollywood lens.

Today, social media is a magnifying glass for wealth. Posts of wealth generation abound: from DFV’s reddit posts, to Chamath’s IRR tweets, to Elon’s Dogecoin memes. Social media acts as an emotional coordination layer, increasing the amplitude and frequency of culture. Jealousy, resentment, and fomo are more viral and powerful than ever, particularly when everyone is on their computer all day post-lockdown.

What Instagram did to body image, wallstreetbets and Twitter are doing to bank account image.

Subscribe now

Where does this lead?

It’s not entirely clear where the financialization of culture goes from here. The accelerative version of the story seems impossible: if we're all sitting at our desks watching candlestick charts, who is going to build the future?

There’s a bubble version of the story, similar to the dot com era: at some point, we could see that the emperor is wearing no clothes and the market crashes 80%. Maybe we’ll see this effect localized, in nihilistic cases like GameStop or some crypto. And if rates rise, highly exotic companies drowning in money today may not get funded: a company not expecting to make money for two decades could see its value fall more than two-thirds if rates returned to 2019 levels, even if nothing about its expected cash flows changes.

Similarly, we could see a populist reversion of the forces that have favored capital over labor: the economy may outperform assets if we see a pro-labor shift in tax policy, increased union presence, or accelerated fiscal transfers. In this scenario, maybe retail investors retreat from the investing landscape for a long time, and equity culture subsides.

The second version of the future is a squeeze, where low interest rates continue and the internet plateaus. Thus entry prices get higher, we see lower market CAGRs for key industries, there’s more competition at the company and fund levels. This may be a good thing: low yields for internet companies means we’ll shift to investing in new platforms like biotech, machine learning, space, and robotics. Speculative companies will grow into their valuations, and we’ll get a steady realignment between equity and reality.

There’s a third version where finance as culture continues. The aggregate S&P P/E ratio is high, but not wildly out of whack on a historical basis. We’ll continue to see a rotation of meme stocks, each its own bubble. Intense fomo of equity holders will continue, accelerating a sense of urgency across lottery and equity culture.

I suspect finance as culture is here to stay.

Thank you to Harry Elliott, Eric Wang, Sam Wolfe, Pranav Singhvi, Delian Asparouhov, Philip Clark, Julian Shapiro, Melisa Tokmak, Brandon Camhi, Axel Ericsson, Everett Randle, Tomcat Sanderson, Napoleon Ta, Josephine Chen, and Daniel Singer for their thoughts and feedback on this article.

Subscribe now

Accidentally adhering to timeful advice is outsourcing a certain part of your life strategy to recent history. But when you outsource decisions that change your life trajectory, do you want to bet the house on the wisdom of yesteryear?

Silicon Valley was once uniquely good at identifying timeful advice. This power weakened as the tech industry scaled, but it is more important than ever to distinguish between timelessness and timefulness.

When I look at my own life, I’ve found that much of the advice I’ve received is dangerously timeful. Below are a few examples – my way of reconciling the cognitive dissonance between the advice I have received and my personal observations.

Should we read the news every day?

Timeful advice: Read the news every day. This is the best way to stay informed about the world.

Counterargument: Trusting news institutions is outsourcing your worldview. What important thing have you changed your mind about recently? If you can’t come up with an answer, that’s a bad sign. The media tends to reinforce your current opinions: you could think of media institutions as the original filter bubble.

The shift to ads-based revenue has weakened the business model of most internet media companies, encouraging clickbait to stay relevant. And clickbait exploits various cognitive weaknesses like fear, greed and envy.

The business model weakens more than just the titles: given the 51% layoffs in newsrooms over the past decade, news publishers have slimmer budgets for important journalistic functions like investigative journalism, fact-checking, and editorial teams.

If you take media content at face value, it deeply distorts your map of the world. For example, it may distort your understanding of common causes of death: a newsreader worries far more about homicide and terrorism (<1% of deaths) than cancer or heart disease (>50% of deaths). This leads to a sense of learned helplessness around problems; risk feels unbounded and out of your control.

Alternative ideas: I’ve limited my daily news intake, rebalancing it into concentrated information streams that are directly relevant to me. It’s become a cliche to say that the media has become a monoculture, but I’m amenable to a more binary approach of cutting out the news entirely.

Rather than outsource your trust in writers to institutions, establish direct relationships with information sources. Substack and Twitter enable this process. Vetting each information source carefully is unfortunately much harder than simply reading the news. It may take months or years to build up a good information source list. Many of the people I respect the most have an information source list curated over decades – writers, experts, friends.

One telling exercise is to read the global news headlines from a month ago – how many stories are still worth reading? The news is the popcorn of media consumption, with very low information density: 99% of it doesn’t apply to you or will be irrelevant in a week, so it ends up being a daily attention tax.

I prefer reading long-form writing, such as books and blogs. Seek high effort per word written – during the golden age of TV, the cost per episode skyrocketed, producing far superior content.

This is a uniquely bad time to be outsourcing your information sources to ML-driven algorithms and intermediaries – between filter bubbles and shadowbanning, it is best to form your own carefully curated information intake algorithm. How else might you improve your information flow systematically?

Should we buy houses?

Timeful advice: Buy real estate. It's a steady compounder investment and a safe place to put your money. In the late 20th century, there were immense tailwinds favoring homeownership:

• Demographic: Buying a home was a common goal among boomers, and was completely embedded into the American culture. It was the centerpiece of the "American Dream.” Everyone wanted to own property, and "everyone" was getting bigger.

  

• Legal: In a classic case of generational regulatory capture, the boomers changed the incentives around homeownership in their favor. The mortgage interest deduction sweetened tax bills for homeowners. And after we hit the western frontier, zoning laws ossified, fixing the housing supply while demand continued to rise.

• Low interest rates: Interest rates dropped from the low teens in the late 70s to ~2.5% today, halving the effective cost of buying a house.

Counterargument: Buying real estate was a great strategy when you could buy up the frontier at low costs: think John Arrillaga in Silicon Valley, or Stephen Ross in New York. But it was timeful advice. People don't get rich from real estate like they used to, and while there is admittedly an emotional component in favor of buying, the pro-buy tailwinds are weakening.

• Demographic: The demographic tailwinds are much weaker than before. The tripling of the likely buyer pool in the late 20th century was a one-time growth spurt that ended around 2010, as the last Boomers entered the prime age for homeownership. Today, the number of prospective buyers is flatlining.

• Regulatory: Homeownership tax incentives became fashionable when the legislators were primarily Boomers. The mortgage interest deduction and capital gains exclusion made homeownership particularly compelling. But as congress shifts from Boomer + Silent to Gen X + Millennial, these tax incentives are weakening. The mortgage interest deduction was reduced in 2017, and the capital gains exclusion may fall out of favor.
  
  The renting generation delayed home buying given increased home prices and student debt. As homeowners age out of legislative bodies, renters will not treat homeownership favorably.

Alternative ideas: Consider the IRR opportunity cost of the down payments: how could you invest that money towards your future? You could invest into equities in the literal sense, but you could also be betting on yourself or your family with that money – investing in education (e.g. a tutor or coach), better sleep and fitness, or starting your own business.

People typically rationalize purchasing real estate by arguing that the home value will appreciate by 2-5% per year. Unfortunately, most of these gains are eaten away by inflation, and sometimes in a big way when the money supply skyrockets.

The downside of buying real estate isn't losing money, but opportunity: what else could you be doing with the down payment? If you look at a typical “rent vs buy” model, the pro-buy argument depends heavily on 1) a 2%+ property value inflation (more than true during the “settlement era”), and 2) a modest ~5% personal investment IRR (your annual yield on investments).

But what if this is flipped – low property value inflation and high personal IRR? Consider a $1m condo with market-standard HOA fees, property taxes, and interest rates:

The opportunity cost of the down payment can often outweigh the property value appreciation. You can think of the rent vs. buy decision as a function of your personal IRR and belief in the housing market.

Which quadrant are you in?

Should we prepare for a rainy day?

Timeful advice: Always prepare for a rainy day. Put 20-30% of your income into savings or retirement accounts. Invest in a safe balance of index funds and bonds.

Counterargument: This is good advice for most people – if losing your job would put you or your family at risk, high savings is a must. However, in the tech sector specifically, there is high employability, so you don’t want to over-optimize for downside protection. Insulate from existential risk, but beyond that, there are highly diminishing returns to savings.

Insurance on smartphones and consumer items, for example, has always baffled me. Many people rationalize the purchase with the logic “you never know what might happen.” Why would Apple sell you AppleCare? Because it’s profitable – AppleCare is a ~70% margin business, meaning you could save ⅔ of that money on average by skipping it. Don’t spend money to protect against limited downside – why give an insurer margin if you can afford the downside yourself?

Alternative ideas: Invest everything you can into compounding assets. But keep your expenses low, so you can focus on upside maximization without worrying about high fixed costs.

Each dollar of net income should be an investment in yourself with asymmetric upside. And investments don’t always look like investments: books and educational content are low cost but have asymmetric upside. Similarly, investing in your sleep setup or home office has compounding returns. At a larger financial scale, you could invest in your own business or those started by friends you trust.

Public equities and crypto can also be lucrative, particularly if concentrated into assets you are uniquely bullish on. In the investing context, I’ve found that conviction is the limiting reagent: it is becoming increasingly popular to trust algorithms and indices, as opposed to investing with conviction. It is convenient to trust the indexes, but resist the force of passivity. Index funds have had a good run, but 1) have bubble-like properties, 2) dampen asymmetric upside by definition, and 3) have no learning value like direct investments do.

Should we get a stable job?

Timeful advice: Start off in a stable professional track after college (consulting, banking), or get an MBA – this career move ostensibly helps establish credibility and unlock higher level jobs in any field.

Despite people vaguely knowing this is timeful advice, I still regularly see seniors at top universities follow it, as the easiest post-college default.

Counterargument: Starting in a professional role narrows the band of outcomes, so you’re less likely to fail, but no more likely to to have outlier success. A lot of careers – consultant, lawyer, banker – are still good, but they were timeful for the mid-20th century, and there are much higher leverage opportunities today if you look for them.

Alternative ideas: Many people seem to understand intellectually that following a safe career path is likely bad advice, but it is hard to identify the tactical things one could be doing differently.

What’s the exact opposite strategy to a safe career? Follow the volatility. History seems to agree: the best move historically has been to work in the most volatile industries, which tends to imply there is a frontier being forged.

Tech startups, for example, were considered highly volatile for the last 20 years – but the vast majority of people who have worked in startups for a 10-year period since 1995 have done well. But following the volatility holds true even further back in history: Wall Street in the 70s, railroads or steel in the late 1800s, international trade in the 1600s. Today, industries like biotech, robotics, machine learning, and crypto are all highly volatile.

Even within your job, you can seek volatility. Within investing, this could be learning about a frontier category that others don’t yet appreciate. Or within a startup, it could be working on the new project within the company. Or it could be taking new bets within your company, like starting a new team or initiative. It’s important to escape the 9-5 employee archetype, instead entering the payoff space, where volatility generates luck.

Admittedly, seeking career volatility is much easier when you’re young and have limited obligations to family. But importantly, many types of volatility have no career risk. Talking to strangers, writing publicly, or building side projects – things without an immediate reward alchemize luck.

Should we travel the world?

Timeful advice: Travel the world to get exposure to as many cultures as possible. This helps you to get out of your comfort zone and develop empathy for other cultures.

Counterargument: I worry that questioning the value of travel will be particularly offensive to people in my generation – nobody questions that travel is the highest form of pleasure.

But most international travel is legitimized hedonism – not harmful, but a highly privileged hobby that doesn’t serve its alleged purpose. Most tourism is highly curated, such that it doesn’t enable learning about new cultures at all.

Alternative ideas: The fact that travel destinations pitch themselves as an “escape” or “getaway” should be telling. What are we escaping from?

In the context of intellectual and psychological development, traveling to most places is no different from staying home – because you are yourself the same. Sure, the “escape” may be a drug for your current problems, but drugs don’t solve problems.

If you’re looking to explore new ideas, you can find differentiated viewpoints and cultures right around the corner, if you look in the right places. Engage with your community at a civic and social level – even San Francisco has lots of intellectual and socioeconomic diversity if you search for it. Once you escape the filter bubble, the Internet is also a good place for finding differentiated viewpoints.

Most dangerously, the Instagram era has magnified the social signaling of travel. Tulum and Mykonos more than doubled their tourist traffic since 2010. Think about whether your choice of destination is fundamentally vanity-driven. Then exclude vanity and consider the motivations behind your trip, maybe learning new things or relaxing: would you choose a different destination, different activities, or perhaps find an even more fulfilling activity locally?

Should we be in no rush to find love?

Timeful advice: You have plenty of time to find love. Enjoy your 20s and don’t worry about finding love too soon. Love happens when you aren’t looking, and you can’t force it.

Counterargument: We’re in a particularly non-Lindy time for dating: the mate-finding algorithm has shifted dramatically in just a few decades, altering the trajectory of our global progeny. A few tailwinds propel this shift: online dating driving the perception of infinite choice in mates, delay of marriage as a cultural norm, and Instagram as a jealousy farm.

Some of my computer scientist friends represent the limit case of this brave new world. They use dating app APIs to “optimize” the top of the online dating funnel, tracking their conversion rates quantitatively from match to chat to date to ____. Does something seem amiss to you?

Too many comparables makes people unhappy in new relationships – people become even less happy as they evaluate more choices, an instantiation of the paradox of choice. We are children who have eaten too much candy, but won’t realize it for ten years.

Alternative ideas: Dating is a counterintuitive area where we may not want to deviate from the timeless advice that has worked for centuries. Finding a life partner should be a serious priority in your early 20s. At some level, it is a simple but somewhat inefficient drafting problem. If you wait to commit, the pool narrows.

There’s an odd contradiction here: people optimize dating more than ever through swipes and Facetune, but commit less than ever. The blissful ignorance of pre-internet dating may be sorely missed in a few decades. Choosing someone and committing to grow with them is much more battle-tested than ten years of partner shopping.

Should we avoid talking to strangers?

Timeful advice: Don't talk to strangers.

Counterargument: Still certainly true for young children. But hearing this advice repeatedly trains you to become too insular and avoid talking to people outside of your immediate circle. And after college, it becomes a lot easier to avoid talking to strangers.

Alternative ideas: You're not talking to enough strangers. The internet massively unlocked our ability to reach anyone in the world at any time. Use this to your advantage.

A personal example: my girlfriend wrote her story about coming from a small Turkish village into Silicon Valley. The key inflection point was when she cold emailed dozens of successful businesspeople to ask them to pay for her plane tickets to Stanford. One responded and changed her life trajectory. Anyone can send an email.

Now is a better time than ever to be establishing yourself publicly early. Things that are hard to measure tend to be undervalued – talking to people online increases your luck surface area.

The world has never been more malleable, as the internet provides high leverage on new ideas. But it requires active effort to create these opportunities for yourself. Maybe there’s a company you want to learn about, an expert in a field you’re interested in, or people with similar interest areas you could befriend. What strangers could you be talking to right now?

Conclusion

There are many types of advice you should follow more, so come to your own conclusions on which advice is timeless versus timeful. Approach non-Lindy advice with caution, as it hasn’t withstood the test of time.

Know when to outsource vs. insource your strategy. Things you want to look for when evaluating strategic decisions:

• Compounding: If the outcome of your decision will compound, insource. Your personal financial strategy, for example: the dollars in a given financial decision may seem small, but the compounding effect of finance makes the outcomes highly consequential, such that the seemingly small benefit of insourcing is larger than it appears.

• Irreversibility: Can the decision, once made, be reversed? If not, insource. Career paths and partners are hard to reverse.

• Stereotype: If you’re making a decision because it’ll help you fit in, you should think twice about it. People often travel to fit in, and they subconsciously understand this: if you didn’t post on social media about your trip to Cabo, did you really go? Not all cultural consensus is bad, but stereotypes are a form of ill-supported consensus, so they’re worth questioning.

• Iteration: If you do something repeatedly, like reading the news, you should think more deeply about how and why you’re doing it. Sleep, diet, and entertainment also fall into this category.

• Magnitude: This one is most intuitively obvious. If the decision you’re making is obviously big, like your choice of career, it’s worth insourcing.

As a rule of thumb, when there may be asymmetry in the outcome, it is probably best to insource the decision.

When an important strategic decision arises, ask yourself if you’ve inherited defaults from people around you. Owning your future requires active questioning of “timeless” advice and coming to your own conclusions.

There is probably lots of advice that I haven’t even thought to question yet. What am I missing?

Thank you to Brandon Camhi, Julian Shapiro, Melisa Tokmak, Sam Altman, Russell Kaplan, Axel Ericsson, Michael Solana, Anna Mitchell, Sam Wolfe, Gaby Goldberg, Jason Lopata, Bala Chandrasekaran, Anna-Sofia Lesiv, Quinn Barry, and Tomcat Sanderson for their thoughts and feedback on this article.

Subscribe now

Hot takes range from SPACs “disrupting IPOs” to being “a lousy deal for companies going public”. There’s a philosophical underpinning to the SPAC debate: SPACs let us dream of a world in which companies aren’t dependent on stodgy bankers to go public. In theory, this world is fundamentally more decentralized and equitable.

But putting philosophy aside, as a founder going public, choosing the wrong path could cost you and your team millions in cash and equity. I found it surprisingly difficult to find a balanced and accessible introduction to SPACs and their tradeoffs, so I compiled everything you need to know.

What is a SPAC?

A SPAC is a Special Purpose Acquisition Company. In layperson’s terms, this means a shell company is formed, raises money from investors, IPOs immediately after, finds a late-stage private company, then merges with them so that the late-stage private company is now public.

From the SPAC sponsor’s viewpoint, it's like running a growth or PE fund that can only make one investment. From the late-stage company’s viewpoint, it's a quicker and easier alternative to an IPO, corporate M&A process, or private fundraise.

Why now?

SPACs have been around since the 1990s – why are they now back in vogue?

Let's not get carried away, SPACs are still relatively niche. There is $25B of SPACs looking for targets right now, which is the equivalent of one large PE firm – pretty small. And only $34.4B across 139 SPAC-driven transactions took place between 2017-2019, compared to $128B across 512 IPOs in the same period.

There are a number of reasons why SPACs are gaining public attention today – and record market volatility and retail investor trading volume have a lot to do with it.

Recent successes: SPACs have historically lost money for the SPAC shareholders, underperforming the market by 3% per year between 2010 and 2017. But there have been some recent high-profile successes in the past couple of years, all generating strong returns within a matter of months:

Robinhood era: IPO bankers typically target companies with revenue scale and a clear path to profitability, whereas SPACs historically have targeted companies that'd otherwise struggle to IPO in a stable manner via a traditional banking process. In contrast to IPOs, SPACs are empowered to acquire more "exciting" businesses – in industries like space exploration, fantasy sports, or electric vehicles – that may appeal to retail investors (read: Robinhood traders). The recent Fisker SPAC announcement, for example, led to a 67% increase in the SPAC’s stock in a matter of days – what Robinhood trader wouldn’t want to see Tesla-style gains from a copycat player? Another electric vehicle example: a SPAC called VectoIQ had their stock jump 3.4x pre-merger on solely the rumor of buying pre-launch electric trucking firm Nikola! Retail investors should be aware that these aren't always premium companies.

Volatility: Volatility can make IPOs expensive for companies going public. If your investment bank is too conservative and encourages you to price your equity at half of its public market value (euphemized as an “IPO pop” or “high coverage”), you’re diluting yourself and your team twice as much as you need to. This effect is particularly acute in a year like 2020, when volatility is the highest it has been since 2008. High volatility periods make it preferable to lock in a price if you can – which is exactly what SPACs accomplish.

Institutional credibility: SPAC sponsorship used to be a common fee-generating strategy among tier 2 or 3 PE funds. While the SPAC sponsors made money through guaranteed fees, the investments themselves often lost money, which led to adverse selection on the deals. In the past few years, tier one names like Apollo, Michael Klein, Bill Ackman, and Chamath Palihapitiya have begun sponsoring SPACs, popularizing the process for higher quality companies. We’ve also seen top tier banks step in to facilitate SPAC IPOs. Increasingly, there are tech-forward people thinking about SPACs, both as sponsors and investors.

Decreasing structure: SPACs were historically used as an LBO equivalent with heavy deal structure: management changes, earn-outs, secondary sales, debt financing, and other financial engineering tools. Today, SPACs are becoming truer to an IPO-equivalent – fewer management changes, and a greater focus on technology companies with high growth potential.

Shifting access from private to public: In the dot com era, the conventional wisdom was to IPO after six quarters of revenue growth. This left massive gains for public investors – Amazon, Apple, Microsoft all saw 800x gains after their IPOs.

But then VCs got greedy. By the Softbank Vision Fund era, the IPO window had shifted massively: conventional wisdom has been to stay private until you simply can’t raise any more private capital – which could get you to a $50B+ valuation. This minimizes gains for public investors.

Will the 2020s see a goldilocks IPO window? That’s one promise of SPACs – at scale, SPACs should help to shift the IPO window earlier for companies with a more nuanced story. This gives public investors access to great companies early, and unlocks liquidity for early investors, employees, and founders.

SPACs vs. the alternatives

As a founder, does it make sense to take your company public via a SPAC? This is mostly a question of alternatives – corporate M&A, PE buyout, direct listing, or a traditional IPO.

From the founder’s perspective, the closest alternative to a SPAC is a bank-led IPO. Existing shareholders get to keep the same upside potential, and the company ends up public either way. Unlike an IPO though, the process for the company is shorter – on the order of 4-6 months instead of 18 for an IPO. This makes SPACs generally less disruptive and closer to a financing round or M&A process in terms of team overhead.

There are a few key differences from a traditional IPO:

• The “IPO pop” myth: SPACs reduce the market volatility inherent in an IPO process, since you’re negotiating a fixed price with a single buyer. When you IPO, you don’t have a concrete sense for pricing until you run a full roadshow process with bankers – and if pricing changes during your roadshow, it’s too late to change banks and too expensive to cancel the process. Recent bank-led IPOs like nCino show how bad of a deal IPOs can be for late-stage companies. As Bill Gurley pointed out, if your IPO is 50x oversubscribed, that probably means the existing shareholders are getting a bad deal. I spoke to Jay Ritter, professor at the UF Warrington College of Business, who shared how egregious IPO pops can be: of the 58 operating company IPOs in the first six months of 2020, the equally weighted average first-day return was 31%, higher than any year since the dot com bubble. This means that $130m in existing shareholder value was left on the table per IPO! Admittedly, investors can’t pull this money out until their lockup is over and the float is substantive, but the cultural encouragement of the IPO pop is a clear market inefficiency. Bankers are playing an iterated game with IPO investors, but a one-time game with your company. Bankers need to make sure they don’t lose money on their IPO underwritings, but don’t forget that a post-IPO pop is a direct wealth transfer from existing shareholders to IPO investors.

• Deal structuring flexibility: In the SPAC universe, it is more common to have milestone-based compensation for execs, contingent SPAC fees, secondary sales, and other deal structures that can align company and investor incentives. Every facet of SPACs is bespoke, so the sponsor and the company can negotiate management incentives, secondary sales, dilution, debt, lockup periods, or any other bells and whistles they want. While this is theoretically possible via IPOs, they tend to have more rigid deal structures than SPACs.

• Financial forecasts: S-1s prohibit companies from including financial forecasts. SPACs, on the other hand, have a private due diligence process which allows companies to present financial forecasts. This opens up SPACs to merge with more vision-driven or high-growth earlier stage companies, where the trailing financial profile is not representative of the company’s trajectory.

Will SPACs eat IPOs? Probably not. IPOs still have a lot of benefits for founders:

• Potentially higher price: Founders may not want a guaranteed SPAC price if the IPO price could (in theory) be even higher.

• Fees: IPO fees can be lower than SPAC fees, depending on the company and deal structure. I expect that over time, the fee structure of SPACs will become more compelling as PIPE:SPAC ratios increase (fees taken primarily as % of SPAC), warrants come down, and more founder-friendly SPAC sponsors come onto the market over the next 2-3 years.

• Control: Founders have more control over their press cycle and investor base.

• Culture: There is still a big cultural premium on "ringing the bell".

Some businesses are structurally better suited for IPOs today – particularly those with 1) a straightforward growth and margin story, 2) a clear set of comparable public companies, and 3) strong pre-existing investor demand. But as SPACs mature and become increasingly founder friendly in terms of fees and pricing, an increasing percentage of companies may be better off going public via SPACs.

SPAC sponsor timeline

From the SPAC sponsor’s perspective, the timeline from SPAC formation to closed transaction can take anywhere from 5 to 30 months.

💰 Raise SPAC, IPO (2+ months): This can take as little as 45 days (but could be much longer), where the key bottleneck is your ability to raise a few hundred million dollars for a blank check. SPAC investors are buying into your ability to find a great late-stage acquisition target, with no certainty as to what that company may be. This period involves auditor engagement, SPAC incorporation, S-1 preparation, SEC filing, a roadshow, and closing / funding the IPO. The average SPAC size was $230m in 2019 across 59 deals. The convention is to price SPAC equity at $10 per share, and the shares trade within a narrow band around $10 until an acquisition target is found. Pre-merger SPAC shares may trade at a slight premium if investors are confident in the SPAC sponsor – e.g. Hedosophia II, run by Chamath Palihapitiya, was trading around $13/share at the time of this article despite still searching for an acquisition target. There are currently 99 active SPACs seeking acquisition targets.

🔍 Find acquisition target (6-18 months): SPAC agreements typically allow for a 24 month window to find and acquire a company. During this period, the SPAC is a public company with nearly blank financial statements. Given the merger itself takes 4-6 months, the SPAC sponsor must find their target in the first 18 months in the SPAC’s lifespan.

The SPAC sponsor typically comes up with a list of target late-stage companies, and engages each one to better understand their business and financing needs.

🤝 Execute transaction (4-6 months): The execution of the acquisition takes 4-6 months, requiring due diligence, preliminary tender offer documents, and 8-K filings. SPAC investors must then approve the transaction, which is the biggest risk in the de-SPAC process – if they reject the transaction, it means deal renegotiation or back to the search process. During the tail end of this process, the SPAC sponsor will coordinate a second group of investors who invest into the company alongside the SPAC via a PIPE (private investment in public equity).

PIPE investments typically accompany the de-SPAC transaction, which essentially adds more equity leverage to the SPAC. SPAC sponsors coordinate the PIPE capital raise from hedge funds and PE firms, who are tagging along for the ride. The ratio of PIPE to SPAC money is typically between 2:1 and 3:1. This means that a $400M SPAC could effectively generate a total transaction size of $1.2-1.6B.

There will be fees of around 20% of the size of the original SPAC (excluding the PIPE amount), which goes to the SPAC sponsor in the form of equity – this effectively means a blended fee of 5-6% for the company as a percentage of total capital raised, fairly similar to the 5-7% for a traditional IPO. SPAC fees are mostly equity-based to align the SPAC sponsor and the company, in contrast to the primarily cash-driven fees for IPO bankers. SPAC fees can also be performance-triggered to incentivize fair pricing, such that a portion of the fees will be withheld unless the stock price crosses a certain threshold. Some SPACs use outside bankers to execute the de-SPAC process, which can add some cash fee overhead.

🕴 Manage investment: After the acquisition, the SPAC sponsor looks like a large investor, where they can take a board seat or two. For more PE-style SPACs, this could involve active management and control by the SPAC sponsor. I expect the SPACs of the 2020s, run increasingly by techies, to adopt the founder-friendliness ethos that defines Silicon Valley.

SPACs often incorporate warrants and earn-outs to further align sponsor and company incentives. This gives the SPAC investors the right to purchase additional stock post-merger, unlocking asymmetric upside for them. These equity and milestone-driven compensation structures tend to align investors and companies more than the historical SPAC structures that were focused on squeezing out fees.

Future of SPACs

Between direct listings and SPACs, banks will no longer be the sole gatekeepers of what goes public. And the SPAC sponsors may shift from structure-oriented PE firms to tech-forward VCs – there are natural synergies given late-stage VCs’ synergies with their existing portfolio companies and access to the best companies. I view this as an inevitability as VC firms build out an array of products to support companies throughout their entire lifecycle – a critical leap towards building Sand Hill Sachs.

As SPAC fees compress and founder-first SPAC sponsors enter the market over the coming months, they’ll become an increasingly compelling option for founders to take their companies public.

Bankers that lean into SPACs will get to capture more company upside and sell their services earlier in the company lifecycle. Those that don’t should be nervous: SPACs may be another step in the startup value creation process that slips out of their hands, and into the hands of private investors.

SPACs won’t disrupt IPOs entirely. But they give us hope of a future where companies can go public independent of the archaic and conservative banking institutions that have misaligned incentives on valuation and fees. I think this hope is well-founded. The SPAC model is more decentralized and ad hoc than the bank-led IPO model, which should be good for founders and investors alike. SPACs aren’t for every company, but every founder should consider them as an alternative path to going public.

Sign up here to get notified about my future posts:

Subscribe now

Thanks to Troy Steckenrider, Kevin Hartz, Harry Elliott, Everett Randle, Matias Van Thienen, Brandon Camhi, Melisa Tokmak, Anna Mitchell, Sam Wolfe, Philip Clark, and Anna-Sofia Lesiv for their feedback on early drafts of this article.

In the late 1990s, the tailwinds began when Internet usage surged from a nerdy hobby on the West Coast to a global household necessity. Since then, the Internet has consumed an ever-increasing percentage of consumers’ time and money. Smartphones and social networks became ubiquitous. 

These tailwinds swept businesses, too. Enterprises SaaS spend has skyrocketed by more than an order of magnitude. A Cambrian Explosion of cloud infrastructure and business tools shattered capex barriers to starting companies.

As market tailwinds grew at 20%+ CAGR, the market dynamics shifted so quickly that incumbents couldn’t react, leaving room for startups to emerge. When SaaS spend grew 50% per year, it was hard not to find green pastures as a new software startup. And consumers doubling their Internet spend and smartphone usage simultaneously every couple of years opened a massive window for new mobile applications.

The current exponential growth in VC fundraises, startup projections, and valuations for the past two decades all assume these tailwinds will continue in the 2020s. Many valuations are derived from expanding public multiples of next-twelve-month revenue, which in turn assume continued exponential market expansion. But if you look at these supposed tailwinds today, a gloomier picture appears: they are not exponential functions, but logistic functions with plateauing growth rates.

The vast majority of the country spends 6+ hours per day online, frequents social media, uses a smartphone, and shops online. This usage will undoubtedly continue to grow: e-commerce will continue taking over physical retail, and SaaS spend will continue replacing manual business processes. 

But people can’t spend more than 100% of their time or money on the Internet. As we approach full online penetration, new companies will need to steal revenue and users from Internet incumbents to grow.

Like any mature industry, Silicon Valley must battle to maintain growth in the face of immense economic gravity. For the first time in Internet history, startup growth will require a push from the company and not a pull from the market. Unlike the organic pull that drove many of the dotcom-era successes, today’s Internet startups need to fight for growth by investing more heavily into sales, marketing, and operations.

Tapering tailwinds mean that incremental spend will slow. This empowers Internet incumbents to capture the remaining incremental spend with their existing product and go-to-market organizations, leaving Internet startups with fewer shots on goal to capture spend:

What are the practical consequences of the Internet’s shift towards maturation? We can look through financial and cultural lenses to better understand the implications. We’ll see an acceleration of zero-sum games between Internet startups and incumbents. A shift from R&D to SG&A will operationalize Silicon Valley, leaving room for new financial infrastructure. VCs will need to take risks on vision, not numbers. And the founders and operators of tomorrow won’t look like those of the past 20 years.

Competition: an acceleration of zero-sum games

We are building an exponential number of Internet companies that compete over ever-shrinking slices of consumer attention and enterprise spend, increasingly locked down by incumbents.

We can look beyond tech to guess what will happen.

How does a company like Boeing continue to grow its commercial airplane business? Sure, there’s nominal annual growth in demand for airplanes, but the primary growth vectors are zero sum: stealing market share from Airbus, or acquiring smaller manufacturers. This will increasingly happen between Internet companies.

Software companies founded today are competing less with pen and paper than with other Internet-first incumbents. Put another way, as happens in every maturing industry before it, Internet company revenue will become zero-sum. As a corollary, the time between founding years of software startups and their competitive incumbents is shrinking:

When there are massive economies of scale – a marketplace or a social network, for example – the blitzscaling approach works well. The first to critical mass wins.

Does this strategy still work today?

Scale advantages still exist at the corporate level: brand accumulation, supplier purchasing power, and cheaper access to capital. But today, true network effects-driven business models are rarer than ever. Simultaneously, there are diseconomies of scale in the tech ecosystem: fiercer competition for talent, skyrocketing real estate prices, and sophisticated incumbents driving up customer acquisition costs.

The blitzscaling playbook is more fitting as a reflection on the past two decades than as a prescription for the 2020s. When the ecosystem-level diseconomies rival the company-level economies of scale – “first to scale” may become “first to fail”. Unit economics matter more than ever. Carefully measured growth will win.

Hiring: a shift from R&D to SG&A

Internet startups will face new tactical realities as a result of zero-sum competition. As we experience CAC inflation across industries due to more people bidding on static ad inventory, Internet startups need to invest more into SG&A than ever to maintain growth.

To run a thought exercise: what would happen if you needed to reduce your startup’s operating expenses by 50%? Many startups are running this unfortunate thought exercise right now.

If Google were to fire half of its employees, there would be large reorgs and engineers would be stretched thin, but the financial outlook would be fine. Its core business – search (supply) and ads (demand) – is an autonomous machine with limited human capital requirements. I would expect a diminutive decline in revenue, and even a large increase in net income.

As an enterprise software company, on the other hand, your growth would slow as you fire AEs, and churn would increase as you shrink your customer success team.

As an atoms-heavy marketplace, your local operations and customer support teams would fall apart, and supply acquisition would stall out.

Internet companies have spent the last 20 years capturing opportunities with the highest margins, lowest operational complexity, and strongest market pull: search, social networks, CRMs, ecommerce.

As the Internet growth tailwinds subside, what’s left? Harder problems. Today, startups tend to focus on problem spaces where there is higher operational and go-to-market complexity. This could mean companies with an atoms component, or with a tougher sales process, perhaps stemming from a weaker market need or more robust competition. This often means higher marginal costs to sell and provide services. More software sales reps chasing the same customers, and more ad dollars chasing the same clicks, means more expensive customer acquisition – it’s a simple matter of supply and demand. This translates into higher investments into SG&A, and eventually lower gross margins. 

To pose the inverse of the opex reduction question: if you had an extra million dollars for your startup, where would you spend it?

In the immature Internet era, a consumer Internet company would likely invest this money into R&D by hiring engineers, product managers, or designers. Consumer Internet companies famously grew without needing employees – Instagram grew to 30 million users with 13 employees in 15 months, and Whatsapp had 500 million users with 50 employees just 5 years after founding. They hired people to support their growing user base after achieving scale.

The dominant startups today tell a different story. As a marketplace with a physical component – say, food delivery or ridesharing – you might spend this money on local ops and supply acquisition. As a SaaS company, you’d spend an extra million to hire more sales reps or run a marketing campaign. These SG&A investments are a prerequisite to drive business growth. Relative to the R&D-driven growth of early Internet companies, SG&A will become the primary growth vector in the 2020s.

Growth: higher SG&A spend means more predictable growth

There is a silver lining: as Internet companies invest more into SG&A, they’ll have a clearer understanding of their growth levers.

As a founder, there’s an important question of which employees are critical to growing the core business.

During the consumer Internet era, an investment into R&D helped move the product forward, but not typically in a financially quantifiable way. The ROI on the marginal hire was positive, but relatively uncertain: you wouldn’t hire an engineer for $80k per year expecting them to generate six figures of enterprise value annually, but they would develop features to retain and grow your user base.

Investing in R&D was king for the early Internet era, but it has diminishing marginal returns – everyone has read cautionary tales around technical debt and engineering coordination problems. At scale, many consumer Internet companies prefer holding cash to reinvesting into R&D.

Paul Graham tweeted that “a visitor who walks around and is impressed by the magnitude of your operation is implicitly saying “Did it really take all these people to make that crappy product?”” But this observation is fixated on a world where R&D dominated Internet startup headcount, and small teams were preferable.

For SG&A-heavy companies that dominate today’s startup ecosystem, headcount is now more causal than resultative of growth. While consumer internet companies scaled up opex after scaling user growth, labor-driven marketplaces and enterprise software companies require heavy SG&A investments to drive growth:

• Consumer Internet: user growth → fundraise → invest into R&D

• Labor-driven marketplaces: fundraise → invest into SG&A → user growth

• Enterprise software: fundraise → invest into SG&A → user growth

This causal relationship may seem counterintuitive, given there’s more tooling than ever making it easier to start a company. Shouldn’t we soon have a one-employee unicorn? But this deflationary pressure on startup capex is offset by the inflation in opex: labor, operational, and distribution costs. As startup success becomes zero-sum, growth will be more expensive – both in terms of financial and human capital.

In the past five years, startup spend ROI has become more predictable as it shifts from R&D to SG&A: in contrast to an engineer, a salesperson or operations leader can drive a quantifiable amount of value to your top line (growing revenue) or bottom line (increasing LTV). But as with most investments with predictable yields, it will commoditize – companies will grow spend as fast as possible while generating acceptable yields. Like any low-yield business, companies will need to compensate for these yields in volume – investing more dollars to generate the same cash flows.

This creates an increasing need for software to operationalize growth. As growth becomes the key bottleneck of Silicon Valley Internet companies, software that measures and unlocks growth will become correspondingly desirable. This may take the form of cross-functional growth software – an orchestration layer between sales, marketing, finance, and operations – quantifying ROI tradeoffs across divisions to help founders make strategic decisions.

Financial operationalization: Sand Hill Sachs

As a prospective founder, don’t be scared off by the increasingly zero-sum nature of Silicon Valley. This dynamic presents an opportunity to build the financial layer of tech, similar to what Goldman Sachs did for the rest of corporate America.

What is Goldman Sachs? Goldman is a financial layer on top of corporate America, helping to fuel its growth. They offer services ranging from M&A advice, IPO underwriting, private equity, debt investing, prime brokerage, private wealth management, market making, and investment research.

As the ROI of SG&A spend becomes more predictable, a non-VC financial layer will emerge within Silicon Valley, similarly helping to fuel its growth. This capital layer can help partially compensate for the slowing market-based growth tailwinds. This suite of services will benefit from a tech-specific approach: real-time debt offerings based on operating KPIs, securitization of software ARR, and retail investor-facing SaaS bonds.

Today, many startups interface with banks only as they near IPO. In the 2020s, tech advisory services will emerge to help growth-stage startups achieve their financial goals – divorcing companies’ financial strategies from their operational strategies.

Why can’t a traditional bank do this themselves? 1) They don’t understand how to underwrite using tech industry metrics (ACVs, churn, LTV, engagement, et al.), and 2) they don’t have the tech DNA to underwrite programmatically, which Sand Hill Sachs will have. This may be a new company, an existing fintech player with distribution in tech, a VC firm, or a bank. My bet would be on a new startup or a late-stage fintech company.

As Alex Danco highlighted in his recent article Debt is Coming, it is clear that recurring revenue securitization – the notion of selling your future ARR bookings at a discount – is the future. The biggest barrier to adoption is cultural: the stigma that “venture debt is like a delicious sandwich that only costs ten cents, but occasionally explodes in your face” is deeply tied to the predatory reputation of old-school venture debt lenders. Companies like Pipe and Clearbanc are already starting to destigmatize securitization, and it will only become more culturally normalized in the coming years.

Once this securitization accelerates, ARR securities will become the next bond-like asset class for both institutions and individuals – irresponsible not to have some in your portfolio, as a fixed income product and a balance against equities. And who will make this market? Sand Hill Sachs.

VCs: take risk on vision, not on numbers

Once Sand Hill Sachs exists, it will become clear that VC dollars should be reserved for R&D, not S&M or G&A. 

A large portion of the collective effort of VCs is focused on late-stage software financings. This certainly will not go away, but companies with predictable metrics will have access to alternative financing options like debt and securitization products. VCs will increasingly need to distinguish between S&M, G&A, and R&D risk.

As a VC, the question becomes: how can you generate alpha through the financial commoditization of Silicon Valley? Thankfully for my industry, there are several risks VCs will continue to be uniquely good at taking:

• R&D risk – can this technology be built?

• Founder risk – can this team build it?

• Vision risk – can this idea become huge?

• Macro risk – will this startup win in 2030’s political, economic, and competitive climate?

VCs will be rewarded for investing in companies with limited empirical data or historical precedent – which, after all, is the foundational idea behind venture capital. Idiosyncratic companies with limited tech precedent will be relatively immune to this shift (defense or health insurance, e.g.), as will those with limited empirical data (e.g. biotech or pre-launch products).

When late-stage investing shifts away from primary capital-based fundraising, secondary investments will become important to generate VC returns for growth investors. Carta and Forge are uniquely well-positioned for this shift as market makers, as are funds registered as RIAs to capitalize on secondary opportunities.

Late-stage VCs often lament that other funds are willing to underwrite to a lower multiple – and thus give higher valuations to companies. If the growth-stage multiple expectations continue to decline, it will only accelerate the relative attractiveness of late-stage debt, both for founders (no dilution) and investors (guaranteed IRR).

Founders: a new playbook

In the late 1990s, the prototypical startup founding team was two engineers working out of a garage, seared into our brains by the successes of garage startups like HP, Apple, and Amazon. This formula worked because you needed computer nerds to reimagine industries for the online world, as far away from industry experts as possible.

Today, the Internet blueprint has been laid. If new Internet startups are increasingly competing with Internet-first incumbents, what would a promising Internet startup founder look like today? There will certainly always be value in fresh perspectives from industry outsiders. But on a relative basis, founders with 1) experience in growing online businesses, or 2) pre-existing distribution advantages will have a premium that didn’t exist before.

For startups taking R&D risk in new technological areas, the founding team may look like something we can’t pattern match to historical successes. Maybe it’s a scientist in his garage who escaped the tendrils of academia. Or your first hire for the founding team is no longer your college roommate, but an expert in your startup’s industry.

Career paths & education: will finance sneak into STEM?

In the 1990s and 2000s, computer science was the dominant major by almost any metric: financial success, career impact, social capital. Eric Schmidt, Meg Whitman, Marc Andreessen, Marissa Mayer, Mark Zuckerberg – all STEM majors.

In a maturing Silicon Valley, what career paths will be rewarded?

We may see a reacceleration in fields like biology or mechanical engineering as we build platforms beyond the Internet. 

In the Internet context, as growth is in short supply, the labor market seems to be providing. We are already seeing growth sneak into job titles: growth engineers, growth teams, growth marketing.

Perhaps ahead of the curve, many of the CS majors at Stanford no longer want to be engineers. Being a product manager or chief of staff is the new post-college status symbol. I think there is an organizational component to this: as there’s a shift towards SG&A spend in Silicon Valley, there’s an increased premium on cross-functional roles (or conversely, a lower premium on pure R&D roles).

A rise in the importance of S&M roles may lead to a revival of econ and statistics in undergrad. Since 2008, there has been a nearly 50% decline in Stanford economics graduates – which you can think of as a “pre-finance” degree – largely driven by a declining appeal of investment banking, and an increasing appeal of computer science. As Silicon Valley builds out its financial infrastructure to maintain growth, will we see a “CS + finance” hybrid major appear?

Conclusion

The Internet’s maturation will have ripple effects across every financial and cultural aspect of the technology industry. There certainly will be $10 billion dollar companies started within segments slow to adopt technology: legal tech, construction, agriculture, and mining are all prime candidates for massive new technology entrants. But new $100 billion dollar outcomes are less likely to come from pure Internet companies.

The Silicon Valley of tomorrow will not look like that of today – success stories rarely repeat themselves – but new Internet opportunities certainly aren’t going away. Quite the opposite: recognizing where we are in the Internet adoption curve clarifies the opportunities in front of us. Founders may seize this moment to build new tools to better understand operational investments, create the financial layer of the Internet, or look beyond the Internet to build new platforms in biotech or energy.

Sign up here to get notified about my future posts:

Sign up now

Thanks to Brandon Camhi, Sam Wolfe, Melisa Tokmak, Harry Elliott, Michael Solana, Everett Randle, Anna-Sofia Lesiv, Lisa Wehden, Gabriel Bianconi, and Lee Edwards for their feedback.

What I learned about: IAC & Match Group 💒💑
Ever since reading the OKCupid data science blog in 2009 (I was in 8th grade), I’ve been fascinated by the implications of online dating. It produced an unprecedented dataset around social preferences. But 10 years later, online dating has produced much more than an interesting dataset: nearly half of all couples in 2019 meet online:

In the online dating acceleration of the last two decades, Match Group – the owners of Match.com, Tinder, Plenty of Fish, and OKCupid – has captured the vast majority of that value and is now a $21B standalone company. They also own a variety of niche dating websites: OurTime (50+ y/o dating), Meetic (French dating), Pairs (Asian dating). Match Group’s success (~5x since selling 20% of IAC’s equity via IPO in 2015) is no surprise given the chart above and their ownership of 4 the top 5 dating sites by MAUs. I don’t see the red line reversing any time soon.

On mobile specifically, all dating apps were below 1M MAUs in the US until early 2013:

And today, the top 5 apps have more than 100m worldwide combined MAUs. This is a well-understood cultural shift, but I don’t think the implications have been widely discussed within SV. This is largely because the largest dating apps are outside of Silicon Valley’s purview: Tinder was acquired more than five years ago, Grindr was made in China, IAC has stayed largely under the radar, Bumble is owned by a Russian conglomerate.

We’re moving away from a world where dating “algorithms” are decentralized: there was historically no central authority deciding who should go on a date with who (except maybe the church). Instead, dating matches 100 years ago were informal and distributed: who happened to be in the same bar, who your friends were friends with, who your parents thought you would like. The matching process is now incredibly centralized, and controlled by relatively discreet companies. A handful of companies have a surprising amount of power in determining the future socioeconomic and demographic makeup of our country. Small UX and algorithmic tweaks can lead to massive changes in interracial and inter-socioeconomic marriages – and the opposite is also true. For example, encouraging users to widen socioeconomic filters for matches (or constraining the filter by default) can meaningfully impact preferences for users who are still determining their ideal candidate. Similarly, the apps can easily filter out certain demographic groups given your historical swipes, which could cut off your likelihood of matching with wide portions of the population.

These companies’ cultural and demographic importance seems to dwarf their market cap. From a regulatory perspective, this is unchartered territory. Is this something the government should have any say over? I’m not sure what’s worse: our dating algorithms being controlled by China and Russia, or the US government. What do you guys think?

See these links if you want to read more about:

• Race and reply rate matrix, race and attraction matrix (two interesting old OKCupid posts)

• OKCupid blog archive (this is a gold mine for pseudoscientific dating data)

• Hinge IRL blog post on how school affects dating app reception

• Match Group’s portfolio

• Match Group Q1 2019 presentation

• Match Group 2018 10-K

What I’m reading 📚

• How Rupert Murdoch’s Empire of Influence Remade the World – one of my favorite NYT feature pieces of all time (warning: long read)

• FBI takes down Deep Dot Web (index of dark web)

• Teenagers are becoming better behaved than they used to be

• Chinese computer vision company Megvii raises $750M ahead of IPO (still think the US is way behind on this)

• The Emerging Market Fallacy: has investing in EM actually been a good bet to take? How tightly correlated are GDP growth and equity growth in emerging markets?

• Can you make a whole society wealthier?

• Drug prices 101, helpful to understand high-level payment flows within pharma

• Justice Department opening antitrust review of big tech – I still expect Apple is the biggest culprit here (may do a newsletter on this, if you’re interested let me know!)

• Do generals matter?

• Is the gig economy a myth? Good WaPo article on how much the gig economy has actually shifted American jobs so far.

• What content dominates on YouTube?

• China Internet Report 2019

What I learned about: TSMC 🇹🇼🖥
Taiwan Semiconductor Manufacturing Company (TSMC) is a living counterexample to the common wisdom that successful tech platform companies are 1) software-only, and 2) run by young founders. TSMC has maintained a near monopoly on pure-play semiconductor manufacturing since pioneering the foundry model back in 1987. We now live in a world where computer hardware is abstracted away behind a nice LCD display, but let’s time travel back to when the tech industry was more electrical engineering than web design. TSMC was the first to solve what was a major cooperation challenge within the world of integrated circuit design: prior to TSMC’s founding, foundries were primarily used to manufacture their own circuit designs. In practice, this meant you couldn’t design circuits without owning a production line, or else you’d risk your manufacturer stealing your design IP. Consequently, only massive companies could design circuits because there was no such thing as a “small test run” of a new integrated circuit without building out your own clean room, production line, and specialized machinery. By solely focusing on fabrication and avoiding any chip design that would compete with their customers, TSMC became the Switzerland of chip production, and started what I’d argue is one of the least well-understood tech platform successes.

Despite being in what is perceived as a pure commodity industry, TSMC maintains 56% foundry market share, more than 5x the size of next-largest foundry. There are massive economies of scale given the extreme capital intensity, both opex and capex, which is increasing exponentially ($9B for their Fab 15 facility, $20B earmarked for their next 3nm facility, plus $10B of annual maintenance). TSMC’s neutrality enables companies like Apple to avoid chip fabrication at plants owned by arch rivals like Samsung. TMSC also operates at 35%+ net margins, higher than both Qualcomm and Apple – somewhat ironic given founder Morris Chang claims TSMC was created so that companies “don’t have to worry about the capital-intensive part of the business any more.” That said, given Apple’s scale, I wouldn’t be surprised if they start their own chip production lines to insource the highest-cost component of the iPhone.

What's impressive about TSMC is its founder: Morris Chang was 56 when he founded the company in the late 1980s, and served as CEO until he retired at 74 in 2005, leaving TSMC with a market cap of $46B. But even more impressive is in 2009 when, after revenue and margins were roughly flat for four years, Morris returned to the company where he served as CEO again for nearly a decade until he retired at age 86 (!) in June 2018. He left TSMC with a market cap of $200B.

See these links if you want to read more about:

• Morris Chang, the founder

• The foundry model of semiconductor production

• The Economist on TSMC’s “Fab Success”

• TSMC’s revenue and net income since 2005

• TSMC’s upcoming 3nm plant

What I’m reading 📚

• A brief intro to QSBS

• Alex Clayton on Zoom’s S-1, with interesting comps on growth, margins, and payback periods of SaaS companies.

• Chinese surveillance and racial profiling – alarmist, but informative.

• Rohan Pavuluri on conditional probability in your career.

• The New Revolution in Military Affairs, depicting what defense could look like if we used 21st century technology. Written by Christian Brose of Anduril.

• Uber S-1, specifically the “core platform contribution margin” waterfall on page 99.

• Uber valuation / TAM analysis

• Proposal to open-source the index of the web. Could be an interesting anchor point for discussion of anti-trust solutions for Google.

• Thoughtful reflection from the founder of Gumroad.

• Map wars (spoiler alert: Google is winning): Google Maps, Apple Maps

• Silicon Valley groupthink in the context of recent IPOs

• China’s Starbucks killer Luckin Coffee raised $150m

• Tesla’s robotaxi network, launching in 2020 according to Elon

• What tech hasn’t learn from science fiction

• Feature on Masayoshi Son, with good insight into his personality and process.

What subscribers said about the last edition (Servicetitan):
Wing highlighted ChefHero as another company aggregating supply within a specific vertical. “Similar to Servicetitan, they've also managed to fly relatively under the radar by being headquartered in Toronto, which raises an interesting point about the pros and cons of starting a company and investing centrally in the Bay Area. Obviously there is an incredible density of capital and talent here, but perhaps when most people believe that the next big thing will come out of this geographical area it's time to shift attention to undervalued and underserved geographies."

Brad surfaced Ivy, a similar vertically targeted SaaS offering for interior designers. They’ve now been acquired by Houzz, which is interesting as it becomes increasingly competitive with Ivy’s customer base.

David highlighted Lango, which he says has parallels to Servicetitan but oriented towards language translators.

What I learned about: Servicetitan 🛠
Nestled in Glendale, a non-techy LA suburb, Servicetitan is just far enough to fly under the Silicon Valley radar: Techcrunch’s first coverage of the company was their $800M valuation round last year (followed months later by a round at a $1.65B valuation), while most SF-based companies get funding announcements starting at their seed round. It's also not an overnight hyper-growth success – founded in 2012, they claim to have "at least doubled" in revenue for the past five years, eventually hitting ~$100M in revenue last year. Servicetitan offers a full software platform for HVAC and electrical workers: payments, payroll, marketing, sales, customer service, scheduling, contracts, reporting, and more.

Most of the key horizontal SMB SaaS applications have played out (CRM, POS, accounting, marketing, payroll, etc), but for an unsophisticated software buyer, it can be hard to navigate these purchasing decisions. This is why Servicetitan and other full-suite vertical-specific applications are intriguing: the messaging is extremely straightforward as they have a single buyer. The one-stop shop messaging removes all complexity from the purchasing experience, and once on-boarded, customers are much less likely to churn given the software effectively runs their entire business. An increasing ability to hyper-target on Facebook and Google means there is an opportunity for vertical-specific SMB software companies to finally reach the relevant SMBs efficiently, despite traditionally having been hard to target. Once you aggregate enough SMB customer density, you can create a consumer marketplace with proprietary supply (like Thumbtack or Opentable, but with deeper supplier integrations).

There are several Servicetitan analogs in other verticals: Toast (restaurants, founded 2011), Mindbody (gyms, founded 2000), Booster Fuels (fleets, founded 2014). What are other services industries that don’t yet have clear full-suite software winners?

See these links if you want to read more about:

• The benefits of Servicetitan’s Glendale HQ: "In the Bay Area, everyone is getting poached left and right. But we have a bunch of superstars, and nobody ever leaves."

• SMB software retention

• Booster Fuels

• Toast’s latest $250M financing

What I’m reading 📚

• John Cochrane describes “narrow banks” and explains why they haven’t taken off.

• Interesting article on retail in China shared by Jiani in response to my last newsletter; another industry in which China is leading the charge.

• Lengthy feature on Deepmind in 1843 Magazine.

• An argument that we're at Peak California.

• Blackpink’s “Kill This Love”, the fastest Youtube video ever to hit 100m views, and #1 trending video in the US – foretelling of Asian influence on US culture?

• A detailed account of a Mexican digital bank heist – probably will see more digital than physical bank robberies this century.

What subscribers said about the last edition (Chinese insurtech):
Eric pointed out that while Chinese incumbents can get away with collusion via JVs, it is much less common in the US. This is because if US tech incumbents form a JV to get a distribution advantage, it will expose their digital monopoly power, making them an easy antitrust target.

Athena devised a creative insurance arb scheme for someone who is into extreme sports but safe on paper – i.e. a bad risk for insurers to take. This person can take out a really expensive insurance policy and sell it for a profit. Do not try this at home!

Jeff noted the relatively common pattern of Chinese firms launching with giant war chests – Lufax, e.g., raised a $500M round very early on given their relationship with Ping An Insurance. Other startups regularly raise $100M seed rounds when affiliated with Chinese incumbents.

What I learned about: Chinese insurtech 🇨🇳🔏
A few editions ago, I wrote about the delta in average time from founding to unicorn in China (4 years) compared to the US (7 years). One such example: ZhongAn, a Chinese insurtech company founded in 2013 by industry veterans, raised at an $8B valuation two years later. They turned heads after selling 5.8B policies to 460M customers in 3 years, and rolling out unique products like “heat insurance” (paid out when temperatures exceed 37º), and “binge drinking insurance” (paid out for alcohol-related medical issues). Coupled with a quickly expanding Chinese insurance market, ZhongAn made for an investor darling. It has since deflated post-IPO – claims ratio higher than expected, expensive channel partnerships, and maybe a distracting blockchain strategy – but offers some good lessons on Chinese insurtech nonetheless, which is evidently far ahead of the US:

• JVs and the value of playing inside baseball: ZhongAn has a unique founding story, started as a JV between the “three Ma’s”: Ma Mingzhe of Ping An Insurance, Jack Ma of Alibaba, and Pony Ma of Tencent. This means very little insurtech competition given nobody wants to mess with the giants. You don’t see this type of startup collaboration from Google, Facebook, Amazon, and Apple, despite a massive distribution advantage. Chinese tech incumbents seem much more attuned to the coopetition mindset. In insurance specifically, JVs solve the distribution challenges that many US insurtech startups face.

• Creative distribution strategies mean creative products: Insurtech companies often struggle because online insurance distribution is expensive: insurance SEO is a textbook case study of a perfectly competitive market. This means insurtech customer acquisition is often even more expensive than incumbents using broker / agent models, particularly given startups’ limited brand awareness. ZhongAn’s integrations with large Chinese tech players solve this challenge, and lead to entirely different insurance products compared to what incumbents can offer. Their most popular products are upsold via other companies’ checkout flows: shipping return insurance sold via Alibaba, flight cancellation insurance sold via Ctrip, cracked screen insurance sold via Xiaomi. The unit economics for low-LTV plans only work if they’re seamlessly integrated with existing digital distribution channels. These partnerships aren’t fully ironed out yet, and may be too expensive at 36.5% of net written premiums, but still point to an under-explored strategy in the US.

• Normalcy around data sharing means better pricing: More data sharing in exchange for better insurance pricing isn’t a tradeoff the Western world is often willing to make, but Chinese insurance companies have been able to capitalize on personal data much more efficiently. For example, Taikang Insurance integrates with healthcare CRM systems to optimize claims management, and Ping An Insurance uses facial recognition to prevent fraudulent claims and increase insurance agent efficiency. We've seen the beginning of data-driven insurance with Metromile and Root Car Insurance in the US, but there is a lot of room for expansion.

See these links if you want to read more about:

• ZhongAn performance stats: claims ratios, cap table, channel partnership fees

• The struggles of ZhongAn since IPO

• Chinese insurtech: biggest players, penetration of insurance products by country

• ZhongAn’s quirky insurance products

• Distribution of startups in China by geography, sector, founders, investors

What I’m reading 📚

• OpenAI is becoming a for-profit.

• Hong Kong is opening up its banking market to increase competition.

• Thoughtful article by Adamant Capital on valuing BTC (haven’t seen a good new approach in a while).

• WSJ details the lack of financial coverage in rural America given the operating costs for local branches.

• NYT on 2019 IPOs: “Thousands of New Millionaires Are About to Eat San Francisco Alive”

• Long-form article by Eugene Wei on online social capital.

• The oil boom in America that nobody in SV talks about.

• Scott Belsky’s predictions for the future. Don’t agree with all of these but I like the idea of doing 5 to 10-year predictions.

What subscribers said about the last edition (checking & savings accounts):
Joe shared Sofi Money, a consumer banking option that offers “the convenience of checking and savings morphed into one hybrid account”. This seems to be the closest thing to what I suggested in the last email of blending checking and savings accounts, and gets around the standard checking / savings dichotomy by using a “cash management account”.

Adam shared a good resource for finding high-interest savings accounts, where some players get creative to drive up APY.

What I learned about: checking & savings accounts 💸🏦
I’m stating the obvious when I say the Robinhood checking account failure was the largest fintech faux pas of 2018. The tl;dr of the Robinhood scandal was that they launched a 3% APY (annual percentage yield) “checking & savings account” to consumers in December, which let them (or so they thought) circumvent the FDIC insurance required for both “checking accounts” and “savings accounts”, and instead insure the accounts with SIPC insurance which doesn’t actually cover a decline in the accounts’ value. This was a pretty blatant circumvention of the FDIC rules, and unsurprisingly, they rolled it back within days and issued a semi-apology. It sounded too good to be true when it launched, and it unfortunately was.

That made me think we may be quite far from a world with 3% APY checking accounts. But I then learned a bit more about a new wave of high-APY savings accounts (read: money market accounts) from Wealthfront, Varo, and Empower. There’s clear product-market fit for high-APY savings accounts, despite the friction of maintaining both a checking and savings account: Goldman’s 2.25% APY savings account via Marcus has earned them ~$25B in deposits with a sub-$80M annual marketing budget. I have become hopeful that there’s a wedge for fintech companies to win on acquiring new customers by getting creative:

1. APY optimization: Compete in the APY race, where the incumbents are surprisingly bad. Bank of America, for example, offers a comically low 0.07% APY for money market accounts and 0.03% APY for savings accounts. With the rise of financial product aggregators like Nerdwallet and Creditkarma, high APY accounts will be rewarded – think of what OTAs have done for flight price standardization. One creative example of APY optimization I’ve seen is Varo, which juices their effective APY up to 2.8% by making up margin in other departments: to get promoted to their highest APY tier, you must have under $50K in your savings, Varo must be your primary direct deposit destination, and you must use the Varo debit card 5 times per month. You can imagine companies getting creative here: “invite 10 friends to unlock 10bps of APY”.

2. Blending checking / savings accounts: Ideally a bit more subtly than Robinhood did. There are fairly stringent requirements that limit what banks can invest in from checking account deposits, making high-APY checking accounts tricky. But at the same time, savings accounts have federally-mandated limits like 6 transactions per month, so they’re impractical as a daily driver. If fintech challengers can control the money flow between checking and savings, they could effectively simulate a high-APY checking account by keeping the vast majority of the balance in the savings account, and transferring funds to the checking account on an as-needed basis (within the 6-transfer limit, of course).

The large banks are able to offer near-zero APY because they control an oligopoly on consumer trust, and the switching costs between banks eases competitive pressures. This is basically ~$11T of nearly interest-free money for them, until new players put pressure on banks’ standard interest rates. New fintech strategies to drive up APY will undoubtedly be a dance with the regulators. Robinhood went too far; Varo and Wealthfront perhaps haven’t gone far enough. Have you seen any creative solutions pop up, either in the US or abroad? How should we be regulating checking and savings accounts?

See these links if you want to read more about checking & savings accounts:

• The Robinhood checking account’s downfall

• FDIC vs SIPC insurance

• Europeans getting burned by uninsured money market accounts in 2008

• FDIC stats – total deposits, asset concentration, margins

What I’m reading 📚

• Easternization by Gideon Rachman – thanks to Shreyas for the recommendation!

• The State of European Tech by Atomico

• The State of UX 2019

What subscribers said about the last edition (remote driving, European advantages):
Axel on Europe’s greatest tech advantages: highly digitized governments, strong telecom infrastructure, cheap labor, free vocational education, regulatory arbitrage between countries (gambling, taxes, fintech), pro-esports culture.

Athena came up with interesting applications of virtual rigs beyond remote-controlled ridesharing: remotely controlled TSA screening, for example. This is likely a regulatory headache, so we may need to find other high-ROI but less regulated applications… remote-controlled farming or construction, perhaps? Intuitive Surgical has unlocked lots of possibilities in medicine with a remote-controlled robotic surgeon, though their core value proposition has been precision versus labor savings.

What I learned about: remote-controlled cars 🎮🚘
The standard narrative around self-driving cars is that the labor efficiency gains and freed-up driver time will have ripple effects across the economy. But as the list of edge cases proves to be longer and harder to solve than originally anticipated, some creative technologists have come up with short-term solutions to circumvent the remaining level 5 CV challenges. Specifically, I’ve found remote-control error handling to be one of the more creative solutions. Phantom Auto, for example, provides remote-control safety drivers as a service. And some full-stack players like Zoox are building this competency in-house.

But when you expand the scope of remote-control drivers to the entire driverless car problem, you can actually get nearly identical benefits to level 5 cars with much lower R&D overhead: full error handling coverage, human-level accuracy, time saving from being “driverless”, and labor savings by using much cheaper outsourced labor.

Full remote-control driving is best tested in private use cases like construction sites, where a network outage won’t cause a traffic jam or accident. For a broader consumer use case, you’d need a low-latency wireless network with perfect coverage, which requires massive country-level commitment to investment in infrastructure. In other words, you probably won’t see it first in the USA 😊 I’d keep an eye on countries like China, Singapore, and Germany.

What I’m wondering about: European unfair advantages 🇪🇺
I am headed to Europe next month to see how the major tech hubs are evolving. What are technology problems that Europe is uniquely suited to solve? Some potential candidates: gaming (Supercell, King Digital), fintech (Revolut, N26, Adyen), travel (Trivago, Skyscanner).

Top replies from last edition (virtual beings):
Gabriel pointed out that AI-based facial reenactment is just about here (source), and text-to-speech is already fairly robust (source). Still lots of technological developments remaining for virtual friends, but many core elements are there.

Sacha shared The Waldo Moment, a Black Mirror episode in which an AI runs for public office. It shows one possible version of a future with virtual beings.

In response to the Facebook Class article, Jungwon thinks it’s worth developing a more nuanced framework for determining when to use a “move fast and break things” mindset versus “extreme planning and testing” in product development. I suspect we’re moving towards an era where the latter dominates across most industries.

What I’m reading 📖

• WSJ on the challenges in the China 2025 plan. This is probably the exact wrong lesson to be taken from the China 2025 plan: “they’re over-leveraged, behind in XYZ dimensions, this isn’t a real threat, etc.” More on this later.

• How Tech Utopia Fostered Tyranny in The New Atlantis.

What I learned about: virtual beings 🤖
I saw the Fable Studio demo at Sundance last weekend, which is a Pixar-quality interactive short with a virtual being named Lucy. The experience was captivating, giving me hope that we may be closer than I originally thought to coexistence with virtual beings. I expect we’ve recently crossed the $100M mark of funding into virtual celebrities and avatars between Brud, Fable Studio, AI Foundation, Mica (Magic Leap), Shadows, Artie, and others.

After looking deeper into the industry, I learned about the power of virtual celebrities: specifically in terms of collapsing the agency-celebrity dualism that exists today. There exists a coordination layer between agencies and celebrities, meaning agencies need to enlist new celebrities when new trends or needs emerge. If you’re an influencer, you can’t easily change your persona to adapt to a new trend. But a virtual agency can quickly create new content – or new characters – to fill the gap.

That said, don’t hold your breath for free-roaming interactive virtual friends. Combining 1) NLP-driven understanding, 2) AI-based facial expression, 3) human-quality voice generation, and 4) realistic intelligence will take time to fully realize; we’re currently there on 0 of these 4 key dimensions. The most successful virtual beings of the next decade will likely lead a more confined existence – but entertaining nonetheless.

Top replies from last edition (food delivery, ISAs):
Ayush thinks ISAs are most exciting in the context of third world countries, given their ability to make high quality education more accessible, particularly in a world where remote work is becoming more normalized. He thinks US companies can save 40-60% on talent using a remote work + ISA model. Short-term credit is another key potential use case of ISAs, and could make the loan approval process much more straightforward.

Chris pointed to some players using ISAs in the context of coding schools that have been live for ~4 years. He suggested alternative ISA use cases outside of education such as Uber taking an ISA on future driving income, or truckers splitting their income in a formalized way to pay for ownership of an 18-wheeler.

Matias mentioned a key reason for food delivery market acceleration: food delivery startups successfully transitioned costs from consumers to merchants, which helps ease friction for consumer checkout flows.

Andrew noted the value of food delivery players capturing specific demographic subsegments – e.g. Doordash in dense urban areas like New York, which takes years to get full coverage in.

What I’m reading

• A 2011 NYT article on "The Facebook Class" at Stanford. Interesting to read in 2019. It makes user acquisition seem trivial: 75 students collectively got 16 million users and $1M in ad revenue in a 10-week period. How often do you see consumer apps getting this type of traction pre-funding in 2019?

• New dating research from analyzing online dating datasets. Interesting takeaways: 1) people pursue mates that are, on average, 25% more attractive than themselves, and 2) using advanced approach strategies doesn’t meaningfully improve your chances of success.

• Making 18th-century philosophy cool again: Michael Solana on capitalism.