Welcome to The AGI Files - a series of essays exploring superintelligence and its consequences.

Why write a series of essays on superintelligence?

Let’s start with an analogy: If you told me there was a chance of rain tomorrow where I live, I would carry an umbrella. If you told me there was a 1% chance tomorrow of deadly acid rain, I probably wouldn't step outside.

In short, humans often adjust their life in anticipation of impactful future events, even in the case of low probability events that are particularly impactful. However, strangely enough, very few people – even within the most tech-forward circles in Silicon Valley – spend enough time preparing for AGI, despite the fact that an AGI timeline is rapidly coming into focus and the impact of existing AI systems is increasingly measurable in society.

In my opinion, this is because the biggest changes in life are also the hardest to predict – and predicting AGI is a particularly hard and undefined challenge. But this doesn't mean we shouldn't try. I hope your presence indicates you are willing to try as well.

This essay in the AGI Files explores the emerging ideologies of AGI1 and how they will impact its creation. Historically, technology shifts are predated by strong emergent forces of human ideology. I will argue that the development of AGI in the US is being impacted by human ideology and will explore why this is consequential.

1 The Link Between Ideology and Technology

The creation of the atomic bomb is often cited as a potential analogy to the creation of AGI. The parallels, on a surface level, are compelling – they’re both impactful technological achievements, they were (or will be) built by accomplished research teams, and they’re both harbingers of doom or peace (depending who you ask). This analogy was actually the impetus for me thinking about the connection between human ideology and AGI – primarily because I think the comparison between the bomb and AGI succeeds and fails in misunderstood ways.

Let’s start by defining two ways to study technological change – P(Doom) and α(Ideology):

• P(Doom) is a well-known term that represents the probability of doomsday. In this context, it measures how dangerous a given technological change could be, is, or was to the human race.

• α(Ideology) is a term I’ve coined. In this context, it represents the strength or pervasiveness of an ideology predating a technological change.

These are obviously not the only two lenses under which you could study technological change, but I believe they are among the most important for studying AGI. It’s also worth noting that P(Doom) and α(Ideology) are both highly qualitative metrics.

If we focus these two lenses onto important technological creations throughout history, the connection between ideology and technology begins to emerge…

It's worth calling out - aside from the atomic bomb, bioweapons of mass destruction are the only known human technology in existence with a high P(Doom).

So, what conclusions can we draw from this road trip through history?

• Conclusion #1: Human ideology can strongly influence the creation and acceleration of technology, and there is strong historical basis for this fact.

• Conclusion #2: Humanity has yet to create a technology with a high value of P(Doom) and a high value of α(Ideology).

• Conclusion #3: Historically, all high P(Doom) technologies have been created and owned by governments. Equivalently, the corporate sector/free market has never produced high P(Doom) technology.

Hypothesis: AGI appears to have the characteristics of the first ever technology with a high value of P(Doom)2 and a high value of α(Ideology), likely because it is being built through the corporate sector and not within the government.

Let’s test and explore this hypothesis.

2 The Emerging (and Powerful) Ideology of AGI

2.1 The Current Ideological Paradigm

The emerging ideologies of AGI, particularly if you measure their size by the number of existing acronyms, resemble something like a fast-multiplying, many-headed hydra. Stay on X long enough and it feels as if every time you learn an acronym, three new ones have sprouted up, rearing their ugly heads and demanding your attention.

It's no longer so simple to represent AI ideology according to the effective altruism (EA) vs. effective accelerationism (E/ACC) two-party system. I prefer a slightly different heuristic where all emerging ideological approaches to AGI have a direction (forward or backward), speed (fast or slow), and acceleration (increasing or decreasing). Today, there are three primary categories of ideology – a desire to move forward faster, a desire to move forward slower, or a desire stay still or go backward. A few examples:

It's worth noting that while I strongly believe the three-way framework presented above is the best way to think about existing AGI ideologies, some of the more specific views straddle these categories, depending on the individual. For example, some individuals claim to be effective altruist through the lens of halting AI. Similarly, some transhumanists claim to be effective altruists. In short, it’s important to listen to what someone actually believes vs. simply taking a label or acronym at face value.

2.2 Processing the Current Ideological Paradigm

Perhaps the most immediately striking aspect of the current ideological paradigm for AGI is how prescriptive it is. In general, human ideology can be prescriptive, meaning it aims to guide behavior or thinking (e.g., the field of normative ethics investigates how one morally should act), or it can be descriptive, meaning it doesn’t espouse certain behaviors or actions, but simply tries to understand things without judgment (e.g. many branches of existentialism describe the human experience without prescribing how to live). You could argue that a sound descriptive basis is often necessary to be prescriptive (for example, how can you tell someone what to do if you don't know anything?).

It is interesting that AGI ideology is almost entirely prescriptive, and particularly strongly so. AGI ideology strongly seeks to influence its creation and is not satisfied with merely describing it. This underscores the particular influence of ideology on this technology, while also illustrating that we are seeking to change something that we don’t yet fully understand.

This is not too surprising – considering AGI’s potential of high P(Doom). Stakeholders want to get things right because the cost of getting things wrong is simply too high. Equivalently, we could say that fear is a significant motivating factor between the three groups. The “stay still or go backwards” crowd fears the dangers of AGI and sees no reasonable path to avoiding these dangers while creating AGI. The “move forward slower” crowd fears the dangers of certain kinds of AGI but believes we can carefully navigate a path towards a safe AGI. And the “move forward faster” crowd fears that the disadvantages (economic, geopolitical, regulatory, etc.) of moving slowly outweigh the risks of accelerating.

Unfortunately, we don’t have strong evidence to support any of these ideological claims today. This is because none of what any group fears has been meaningfully realized. Until then, we exist in an era of ideological speculation, which allows ideologies to run rampant and to possess fairly coherent supporting arguments. Of course, there are numerous smaller data points that ideologists will argue support their claims. For example, we could point to recent research demonstrating that smart agents tend to scheme/seek power or cite recent comments about AI-related layoffs (e.g., Salesforce and Meta) as evidence that we should slow down or halt AGI progress. Alternatively, we could cite China’s rapid frontier model progress and the historical importance of technology for economic/geopolitical hegemony as arguments that we need to accelerate.

2.3 Can AGI Ideology Influence AGI?

By this point, I’ve hopefully convinced you that AGI ideology exists. But now we need to ask –are there mechanisms for AGI ideology to impact its own creation? I would argue there are:

• AI researchers can choose where they work based on ideological alignment, effectively voting with their time and research contributions.

• AI customers – both consumer and enterprise – can vote with their wallet, giving capital to research labs that align with their individual ideology. (VCs also technically vote with their wallets but I don't give us much credit - I don't think we're a particularly ideological bunch).

• While nascent, we might also point to regulators, who may produce AI legislation based on their (or their party's/administration's) ideological beliefs.

Thus, unlike secret government projects like the Manhattan Project or bioweaponry, AGI is more readily influenceable by a broader ideological public (n.b., I use "broader" on a relative basis – the mechanisms for influencing AGI clearly do not resemble a true democracy). This is not necessarily a good thing by the way (to be explored later in this essay).

2.4 Will AGI Ideology Influence AGI?

We now have two conclusions – that AGI ideology exists AND there exist mechanisms for these ideologies to influence the creation of AGI. But still, this does not guarantee that AGI ideology WILL influence AGI.

Other factors may have a much stronger effect – such as the laws of capitalism. For example, AI researchers may only choose where to work where they get paid the most, AI customers may only choose to use the cheapest or most performant models (n.b., any company experimenting with Chinese models are decidedly NOT ideologically motivated), AI research labs may shut down if their unit economics suck (looking at you, Anthropic), and regulators may be influenced by lobbying efforts from commercial interests.

I do believe that other effects, such as capitalism, are real – even extremely strong. But I wouldn't discount the effect of ideology either. Safety matters to many consumers and enterprises, and labs are pursuing radically different approaches to appeal to different types of users (consider Grok vs. Gemini). Furthermore, I also believe that AI researchers, perhaps more than any other group, profoundly care about AGI ideology, while also having one of the most important and impactful “votes.” I directly and indirectly know many researchers who have made career decisions or have switched labs (or were even fired) due to their ideological alignment, sometimes even at the expense of salary. This is not just an employee-led movement, it's also led by executives. Anthropic is well-known for vetting ideological alignment during the hiring process. Anecdotally, I have a friend who was recently denied a senior research position at xAI due to perceived ideological misalignment.

3 Implications and Consequences of a Radicalized AGI

So here we are…on the cusp of creating one of humanity’s most important and potentially dangerous technologies and doing so through a powerful ideological mechanism. Aside from historical novelty, what are the implications?

• Potential division of AI talent: We do not exist in a world where we can take our smartest AI researchers, aggregate their talents into a single lab, and sprint at building AGI. This creates a division of best-in-class talent, likely slowing down the AGI timeline. Of course, there may be efficiency gains from this division of talent (because ideologically opposed people tend not to work well together, or because competition breeds excellence), but these efficiency gains are only because talent is already ideologically opposed in the first place.

• Major research breakthroughs will be scattered, reducing competitive advantages: The potential division of AI talent will result in scattered research breakthroughs. Achieving AGI requires research advances in a multitude of domains, and since domain experts will be at multiple labs, we can expect independent breakthroughs at multiple labs (this has already played out many times in the last few years). Scattered breakthroughs make it more challenging for labs to create robust and consistent research advantages, limiting competitive edge. This also helps close the gap between open source and closed source models, which may have significant implications on the commercial viability of closed source.

• Inefficient allocation of resources: With many labs shooting for AGI, increasingly scarce resources are being allocated to redundant tasks. Each independent lab needs to raise funding, collect data, hire researchers, invest into infrastructure and compute, use energy…the list goes on. Many experts believe the infrastructure and energy requirements required to create AGI alone will require a significant amount of investment relative to what exists today. Each research lab, assuming they stay in the game, effectively represents a significant stepwise increase in redundant resource requirements. Over time, I expect the resource allocation problem to become more pronounced and unwieldy.

• Ideological division may may stymie geopolitical decision-making and stunt American geopolitical first mover advantage: While ideological division may create a healthy dialogue around what AGI should look like, it also can erode America’s geopolitical first move advantage by slowing down its creation. Authoritarian regimes like China are theoretically far better equipped in this regard to centralize talent, ignore ideological concerns, and produce AGI as rapidly as their research breakthroughs and infrastructure allow. AGI has the potential to be an enormous source of geopolitical + economic hegemony and America may be slowly relinquishing its lead.

• Research lab ideological alignment may take on a political dimension: Just as other technological debates (big tech, regulation, etc.) have taken on a political dimension and found homes on opposing party lines, I anticipate AGI and its ideology to eventually take on a political dimension and align itself with our two-party system. I predict this will be a key development during the current Trump administration. In future administrations under different party leadership, we may see the U.S. flip-flop its stance and approach to AGI development to adhere to party ideology. I will even go out on a limb and predict that the Democrats will ultimately adopt a “stay still or go backwards” AGI platform. In addition to this resonating with their recent antagonism to big tech, it will appeal to anyone experiencing layoffs and may help them recapture their historically “pro-worker” platform.

• It may be challenging to create AGI for the silent majority: Ideological polarization is a common phenomenon in systems with competing points of view, where the representatives of a viewpoint gravitate towards extreme positions rather than moderate stances, even when moderate stances are more inclusive of a populace. The obvious example is the U.S. two-party system. If AGI ideological polarizes further, it’s possible we will create an overly-radicalized version of a technology that caters to niche radicals vs. a more moderate majority.

• Technology that impacts the national interest is exposed to capitalism and other effects: As mentioned, human ideology is not the only factor impacting the creation of AGI. To the extent that ideology is, or becomes, an incredibly desirable factor – we have to consider the possibility that ideologically desirable labs may be undermined by things like basic market forces. For example, ideologically desirable labs may not produce the best products, nor have commercially capable executives, nor attract enough venture funding.

4 Should Ideology Impact AGI?

This is the last important question to consider. Unfortunately, it doesn’t have a clear answer. The implications and consequences of a radicalized AGI are a mixed bag. What we gain from certain types of efficiency, ideological variability, and economic incentives can also be offset by certain types of inefficiency, division, or geopolitical risk. Many of the consequences listed in Section 3 are themselves double-edged swords – capable of producing both good and harm.

Another way to answer this question would be to consider the alternative – that AGI is created a la atomic bomb or bioweapons – i.e., by the government without the influence of ideology. While this is largely speculative, I believe this would reduce the influence of economics but vastly increase the influence of geopolitics on the technology. I believe that governments are more likely to consider safety risks and implement guardrails than commercially-motivated research labs, which of course would make many accelerationists unhappy and many cautious ideologists proud. I also believe that creating AGI through independent research labs is significantly more open and democratic than creating AGI in a government lab. It’s also worth mentioning that the US government and its agencies have not demonstrated the ability to keep up with technology, especially in recent years, and so it’s unclear they would even be capable of producing AGI under a similar timeline.

Of course, unless the US government intervenes and attempts to nationalize or pseudo-nationalize elements of AGI, this is currently not a possible path for American AGI. In other words, the question of should ideology impact AGI is largely rhetorical, as it seems, at this point, inevitable.

5 Parting Thought

I first pitched this essay to a friend of mine who leads a safety team at a well-known research lab. His response struck a chord. He said something to the effect of:

"The AGI ideologies we see today are young and probably naive. I have a feeling the most impactful ideologies around AGI have yet emerge."

He is 100% correct. This essay explores how human ideology can influence the creation of technology. However, it is also true that the creation of technology can produce powerful reactionary ideologies. Think of the ideological (and geopolitical) debate around the atomic bomb. Or consider the Luddites, the group of 19th-century textile workers who struggled against the automation created by the Industrial Revolution.

The concept of post-technology reactionary ideology and specifically the story of the Luddites are topics that deserve their own separate essay, which I will be sharing soon. Stay tuned!

General-purpose, humanoid robots – robots that look like humans and promise to do all the things that humans can do – are not a new concept. We can trace the humanoid at least as far back as 700 BCE, when the ancient Greek poet Hesiod described the story of Talos, a giant bronze man built by Hephaestus who protected the island of Crete from invaders. Nor were the Greeks the only people fascinated by the humanoid - an ancient Taoist text written by Chinese philosopher Lie Yukou in 300 BCE spoke of a humanoid automaton, and the earliest known references of the Jewish golem date back to 300 CE, among many other examples scattered throughout history.

In 2024, it seems like we are on the cusp of realizing a millennia-old dream of a functional robot made in our own form. Entrepreneurs and investors have flocked to the category, investing countless engineering hours and billions of dollars this year alone. Many believe that humanoids will soon flourish in the workplace, factory floor, assembly line, warehouse, hospital, and even at home. Many companies are seeking to build the humanoid themselves (e.g., Tesla Optimus, Figure, 1X Technologies, Agility Robotics, Boston Dynamics, Sanctuary AI, Unitree) while others are focused on the underlying foundation models that will power these machines (e.g., Physical Intelligence, Skild AI).

What was once perhaps an idle fantasy is now, at least on the surface, an inevitability. Believers in the general-purpose humanoid all share some version of the following thesis:

If our world has been designed by and for humans, then the human form will be the most effective means of interfacing with it.

I believe this thesis is deeply flawed and misunderstood. I do not believe the humanoid is an inevitability. In principle, a human form factor may be highly versatile, but in practice, non-humanoid robot form factors[1] are better. I would go a step further and argue that general-purpose humanoids will hold the robotics industry back and could delay the adoption of beneficial robotic automation in our society. To understand why this is the case, it's critical to understand the challenges that robot companies face in getting to production – and why humanoids uniquely struggle to overcome these challenges.

Why Robots Fail

"Most people think robotics is a hardware problem – but it's a software problem." So starts a recent post from Sequoia Capital, announcing their investment into Skild AI, a company building foundational models for general-purpose robotic intelligence. Reframing robotics as software has been a recent trend amongst entrepreneurs and investors. It makes sense why this is desirable. It's nice to imagine that a traditionally capital-intensive category can be "solved" with a positive FCF-margin solution (a pitch, no doubt, that is particularly enticing to a VC investment committee). And to be fair, there's at least some technological evidence supporting this optimism - with hardware costs declining and large foundational AI models (potentially) presenting a new paradigm for robotic intelligence.

Unfortunately, the software vs. hardware narrative is not just wrong – it completely misses the point. Hardware and software are challenges that robotics companies encounter but they are not the primary reasons why these companies fail. There is a third understated and underrated challenge, and it's the hardest to overcome. I call it the operationalization problem.

Operationalization is the process of taking a robot, which has hardware and software components, and getting it to work in a real customer environment. It means integrating the robot both upstream and downstream of some existing process and connecting the robot to all the systems that exist in the customer environment. It means incorporating the robot into human workflows and teaching the human how to work with the robot (and sometimes vice versa). It means grappling with a seemingly infinite number of edge cases that were not identified in the lab or initial POC. It can mean starting this process from scratch for every new customer. It means knowing how to communicate with the customer in a language they can understand.

Over the past year, I have had hundreds of conversations with roboticists. The conversations that have been the most compelling were with individuals who have already been in the trenches. To give a sense of the scope of the "operationalization problem," I've shared a small handful of quotes below (anonymized to protect identities and privileged information).

In short, when you ask roboticists to talk about their battle scars, they talk about operationalization – not hardware nor AI. This is a significant narrative violation, and it unfortunately catches too many first-time founders unaware.

So why is this relevant to this post? Well, once we accept that operationalization is the key challenge, it becomes clearer why humanoids are not the ideal form factor for the future of robotics.

Why Humanoids Fail Harder

Simply put, humanoids are particularly bad at operationalization, in addition to also struggling at the hardware and AI layers. Here's why:

• Humanoid hardware is overly complex and integrating complex hardware is challenging: If building a performant robot is a complex engineering problem, building a performant humanoid robot is an extremely complex engineering problem. Humanoids have more degrees of freedom, more balance and stability issues, higher mechanical intricacy from trying to mimic the human anatomy (joints, limbs, hands, etc.), and more complex actuator problems and limitations, among many other challenges. This introduces complexity to the hardware layer, an area where most robotics companies would prefer to avoid complexity. Furthermore, it exacerbates many separate operational issues that I’m about to detail further – cost, reliability, safety, time to market, error rates, etc.

• Humanoid hardware is unnecessarily costly: The cost of humanoids currently dwarfs that of their non-humanoid counterparts. Many believe that the cost of a humanoid can eventually fall to low- to mid-5-figures, which is cheaper than the cost of a human laborer (from a corporation's perspective) or cheaper than the cost of some consumer purchases (from the consumer's perspective). However, no company has so far demonstrated they can produce such a robot. Furthermore, training, integration, maintenance, and upgrade costs for humanoids remain unknown until more of these humanoids are deployed. Of course, these costs will come down in the long run - but so will the costs of non-humanoid robots. Humanoid companies will attempt to pass the higher costs down to their customers, who will struggle to justify the high price tag when they can just buy a more specialized and cheaper non-humanoid.

• Complex maintenance supply chain: The supply chain for new humanoid parts can be quite complex. The hardware components are delicate, and due to the nature of the tasks, they need to be replaced quite frequently (n.b., this is truer of humanoids because they are mobile vs. stationary). Furthermore, customers are typically bad at knowing how to replace parts, which can create significant ROI-destructive downtime. One famous general-purpose robot company told me that they had to fly out employees to customer locations weekly to fix and replace parts. For another famous humanoid robot company, their solution was simply to replace the robot instead of fixing certain parts (i.e., they literally just ate the cost of the components that did not need to be replaced).

• Humanoid companies are overhyping their capabilities: Humanoid robot companies are unnecessarily exaggerating their capabilities to a public audience. Insiders at several high-profile humanoid robot companies have confided that their demos range from misleading (e.g., cherry-picking 1 out of 100 attempts to make the robot appear more capable) to borderline fraudulent (e.g., saying the humanoid in the demo is standing on two legs, when in reality it is pillared to the ground and obscured by a table). While unfounded claims may serve to attract investment dollars, I believe they will also serve to disappoint customers who expect something quite different than what they will get – a dangerous double-edged sword. In some places, the damage is already done – I've spoken with several robotics R&D/innovation leaders across multiple end markets who roll their eyes when I mention the word humanoid.

• Humanoids are relatively unsafe: Humanoids are more prone to collapse than their non-humanoid counterparts. This is driven by physics – humanoids often have two points of contact with the ground (vs. three or four, which is passively stable), a high center of mass, and often carry payloads outside the cone of stability (n.b., the area in which a humanoid can sway without losing balance). They can also be quite heavy, further posing a safety risk to nearby humans. These environments can range from manufacturing settings populated by healthy adults to home environments filled with children, babies, and pets.

• Foundation models that can power humanoids are far from realization: While advances in ML/AI are occurring rapidly, the industry is far from having a useful foundation model. Companies like Physical Intelligence and Skild AI are doing interesting work in trying to recreate a GPT-like model for robotics. But this will take longer than most people think. First, there's a massive data shortage and no ideal or easy method of scaling data collection, unlike how the internet facilitated training LLMs. Second, the nature of the real world is far more open-ended and complex. Already hard problems, such as walking up different types of stairs, can become infinitely more challenging when you want robots to deal with other factors simultaneously, such as carrying objects of different weights and shapes. Third, if we pattern match the length of time it has taken to achieve useful computer vision models or LLMs, it's probably safe to assume at least a similar trajectory for robot models. And fourth, even if we were to have a robot model on par with the best LLMs today, it would be useless in my opinion, as the number of errors and hallucinations that are acceptable for language-based tasks is higher than what is acceptable for robotics tasks.

• Too many humanoids are solutions in search of problems: One piece of advice I consistently heard from robotics veterans was to never build a robot before you've identified a problem you want that robot to solve. I am not saying that robot companies can't have a more ambitious, long-term vision of becoming general-purpose. I also am not saying that all humanoid robot companies are making this mistake. What I am saying is that some - if not many - humanoid robot companies do make this mistake, and it’s because they believe the humanoid form factor will be so useful that commercialization will be easy. Spoiler: it won't be.

• Humanoids will take too long to reach market: Given the factors already outlined, we can conclude that humanoids will take longer to reach market than non-humanoids. Non-humanoids have fewer open-ended research problems and fewer hurdles to achieving production readiness. By the time these problems have been solved for humanoids, many non-humanoids will already be embedded in the most relevant end markets. Competition will be fiercer and true whitespace will be less abundant. Furthermore, robots are usually sticky products – so good luck ripping them out. Strategically speaking, it just makes more sense to build a non-humanoid that can achieve PMF quickly and then slowly embed more general-purpose capabilities into the robot over time.

• Humanoids will be left with immensely challenging tasks: By the time humanoids reach market, only the most challenging tasks will be available as whitespace. For less challenging tasks, humanoids will need to be able to displace existing automation solutions, which will either be handled by general-purpose non-humanoids or traditional robots. It’s worth noting that traditional automation providers have already automated as much as possible in most end environments today. Furthermore, due to rising labor costs, traditional providers are in fact expanding their capabilities because more expensive projects are becoming economically viable. This is all to say that humanoids will face exacerbated deployment challenges relative to other approaches.

• Humanoid mobility creates additional operationalization challenges: Humanoids trade efficiency for versatility. From a customer’s perspective, if the humanoid is not able to accomplish tasks faster than non-humanoid alternatives, then the humanoid must be able to make up for that loss in efficiency by being able to go handle other tasks that the non-humanoid cannot do during idle time. However, this introduces a new operationalization problem that has not yet been solved – how do you coordinate and herd a group of humanoid robots to accomplish a set of (probably) amorphous tasks (because the amorphous, weird tasks will be the only ones left untouched)? This isn’t impossible, but it is another layer of challenging operational complexity.

• Humanoids will struggle with reliability and ROI:  Given the factors already outlined, humanoids will struggle to be reliable and provide as much ROI as their non-humanoid counterparts. Failing on reliability and ROI has been a death knell for many robotics companies. This is the heart of the operationalization problem.

Where Does My Thesis Go Wrong

A lot of smart people will disagree with this post. I think they're wrong, but for fun I'll play devil's advocate.

On a surface level, you could take the opposite stance of anything I've said so far. For example, you could try to argue that humanoid hardware costs and complexity will be on par with that of non-humanoids very soon, or that we're only a couple years away from a GPT-like foundation model for robotics, or that the operationalization problem is trivial. But that's a bit boring and will sound fairly repetitive. Instead, I'd like to call out a few other arguments:

• Humanoids are psychologically appealing: Maybe it doesn't matter that the humanoid is a suboptimal form factor. The concept of a humanoid is undeniably captivating on a deep level with the human psyche – humanity has been talking about humanoids for thousands of years! It's possible that our desire to realize the humanoid will eclipse the barriers to their realization. Who knows...if all the smart people decide to go spin their wheels on humanoids, that might even give humanoids an advantage.

• Certain environments and end markets may be so chaotic that only a humanoid form factor is useful: I previously argued that in most end markets general-purpose non-humanoids will be able to commercialize more quickly than humanoids, even if the humanoid form factor is more versatile. But what if there are certain environments that are so chaotic that only a humanoid will ever be useful? I'm willing to entertain this. One potential example that comes to mind is the home environment, which is an enormous end market. It's possible that the non-humanoid wins in enterprise applications in the next ten years but the humanoid wins the consumer segment over the next twenty (though I would still bet against it). 

• Humanoid Convergence: It’s possible the non-humanoids that are successful in the near-term will eventually adopt more humanoid features, whereas the unsuccessful humanoids will need to become less humanoid in order to survive. In other words, over a longer time horizon, some of these form factors, which look quite different today, may converge.

Parting Thoughts: Hype -> Grind -> Graveyard

Normally, when you build a company, your trajectory should look like this:

In other words, you start your company, you grind really hard, and then either your company dies or it succeeds. If it succeeds, you earn the hype. This is “Grind -> Hype or Graveyard.” It’s the natural order of things.

Robot companies don't look like “Grind -> Hype or Graveyard.” Instead, they look like “Hype -> Grind -> Graveyard”:

For robot companies, hype is concentrated at the early-stage – typically at the Seed and Series A. During this phase, the robotics companies have a long-term vision, deliver an early prototype/MVP, and engage with a few design partners. Afterward, robotics companies enter an excruciatingly long period of "Grind." This is where the operationalization problem becomes apparent - customers fail to convert, integration costs remain high, the long-tail of errors appears unsolvable, and it's hard to manage burn. Inevitably, almost all these companies end up in the graveyard. The few historical success stories, frankly, aren’t that remarkable.

This is basically the opposite of how startups should work. It simply shouldn’t be the case that robotics companies are most attractive when they are all vision and no product, and least attractive after they build the product.

It's incredibly important that this changes. There are so many exciting problems that robots can eventually solve, particularly after embedding next-gen AI into their stack. However, “Hype -> Grind -> Graveyard” cannot change until the industry better understands why robot companies fail. And the humanoid is unfortunately not pushing the industry forward, but holding it back.

I'm investing significantly in technologies that I believe can transform the real world. If you're a founder and this post resonates with you, my DMs are open. If you hated this post, hit me up for a healthy debate.

[1] Not to be pedantic, but it's worth noting that the term humanoid is a spectrum - you could call a robot that looks virtually indistinguishable from a human a humanoid (e.g., Ava from the film Ex Machina), but you could also call a stationary arm a humanoid because, well, it has an arm. In this post, when I refer to humanoid, I'm generally referring to robots that are almost fully human in form (i.e., two arms, two legs, hands, feet, a torso, a head).

Subscribe now

When I joined Vine Ventures as the first hire a little under three years ago, I asked myself what it would take to build the world’s best venture fund. A lot was at stake – not just me and my colleagues’ careers, but the capital from our limited partners and the relationships with our future founders.

The last three years have been the best kind of whirlwind. We are now the second-largest seed-stage venture fund based in New York with additional offices in San Francisco and Tel Aviv, and we are proud long-term investors in over 25 exceptional companies led by high-integrity, problem-obsessed founders.

I wanted to share this info at the start of this blog post for two reasons. First, I’m excited to announce my recent promotion to partner at Vine. It’s an honor to have built such wonderful trust among my colleagues, founders, and broader community, and I couldn’t be more fired up for the journey ahead. And second, I wanted to revisit the same question I asked myself three years ago – what does it take to build the world’s best venture fund – and share a few thoughts.

The funny thing about this question is that very few investors ever seem to ask it. At any moment in time, there are a handful of key challenges that impact the venture ecosystem. Some of these challenges are exogenous – for example, the fact that valuations for exciting AI technologies are high or the fact that there are so many venture funds competing for too few compelling deals. In my experience, these exogenous variables are the only ones that people talk about, and they’re also the least addressable.

What I find more important – and yet rarely discussed – are the endogenous challenges in VC. These are the internal variables at a VC firm, the same variables that we, at Vine, have been tweaking and adjusting as we scale.

I strongly believe the following to be true: the vast majority of VC firms are severely self-unoptimized, meaning they construct and run themselves in a way that is detrimental to operations and returns. Unlike exogenous variables, endogenous variables are fully within VC firms’ control, which is why it is surprising that they have addressed them inadequately.

I realize this post will be more “VC-ey” than normal and that most of my subscribers are founders. This is by design. I believe it’s incredibly important for founders to be aware of these dynamics, as working with a “self-optimized” VC will produce a far better partnership than working with one that is “self-unoptimized.” I will circle back to this point at the end.

Endogenous Variables in VC

The following represents my list of key endogenous variables that matter to a VC, and that are often significantly self-unoptimized.

• Fund size: In theory, it’s common knowledge that fund size equals fund strategy. In practice, many VCs treat these concepts as separate. There are many factors that drive VCs to increase fund size beyond reason, but a combination of overconfidence in ability and desire for more management fees are the usual culprits. But what constitutes the “correct” fund size for a given strategy? The key input is a reasonable assumption of the total enterprise value of a venture portfolio at exit. Let’s create an illustrative example for Seed and Series A funds. Let’s say $7.5bn is your assumption for average total EV at exit because you expect to hit ~1-2 unicorns per fund and a decacorn every few fund cycles. With this as a starting point and a few other simplifying assumptions we can estimate the following ideal fund size for Seed and Series A firms (separated by those who lead and don’t lead deals). For fun, I then show how a few well-known funds stack up to these figures. As you’ll see, some fund sizes make sense relative to the returns they promise, and others do not.

• Size of investment team: VC firms should strive to hire the absolute minimum number of investors needed to identify and invest in the best technology companies for a given strategy. Bloated teams slow down decision-making processes, reduce overall talent quality, can frustrate high-performers and incentivize them to leave, and force people to spend time on nonessential workflows and opportunities. Bloated teams will also exacerbate many of the other factors that I’m about to discuss in more detail (e.g., delineation of responsibilities, decision-making, promotion path, compensation, etc.).

• Delineation of responsibilities: How are responsibilities split within a VC firm? Are the employees who source deals different than those who invest in a company? Is the person who knows the founder’s business on the board, or was that responsibility given to a more senior partner? Do the investors have direct relationships with relevant operators, or is that handled by someone on the platform team? At Vine, we aspire to be full-stack investors, a term we use to describe someone who can source an investment, run point on diligence, sit on the board post-investment, and introduce founders to potential customers and hires, among other responsibilities. Too many firms fragment responsibilities among too many people. While it may be easier to build this kind of VC firm (e.g., it’s easier to hire someone who is good at one thing vs. good at many things), it inevitably results in founders receiving fragmented and inconsistent value from their investors.

• Decision-making: Who makes investment decisions at a VC firm? From the founder’s perspective, it’s important that the person who best understands your business has the most influence on the final investment decision (otherwise you’re wasting your time). From the VC’s perspective, it’s important that processes are in place to ensure that decisions are more right than wrong. Too many VC firms have junior investors who are closest to an opportunity but furthest from the final investment decision. Furthermore, too many VC firms have broken decision-making processes, either by giving individual partners too much leeway to call their shots or overconcentrating decision-making power to the whim of a single GP. Instead, perfecting the decision-making process requires a system of checks and balances, where a deal partner has the most informed and respected voice in the room, but where the voices of other partners on a diverse investment committee can approve (or veto) the deal partner’s recommendation.

• Promotion path: Many VC firms do not offer a clear path to promotion for junior investors. This is a huge mistake. If potential for promotion explicitly or implicitly does not exist, then junior investors may make decisions that are fundamentally antagonistic to fund goals. For example, it is common for junior investors to recommend highly competitive, overpriced investments and then use that track record to pivot to another fund before they can be evaluated on those deals.

• Performance assessment: To run a successful company, it’s important to retain top-performers and let go of poor-performers. However, most VC firms struggle to assess performance. If you’re trading liquid assets and have a short investment duration (e.g., a long-short public equity hedge fund), it’s more intuitive to track an investor’s performance. But venture capital deals with illiquid assets over incredibly long durations – often success (or failure) is over a decade away. Consequently, most VCs assess performance on 1) deal access, 2) velocity of deals, and 3) markups. But these measures can be misleading. Good access often means you’re seeing the same deals as everyone else (i.e., no alpha), doing a lot of deals means nothing if those deals do not generate returns, and paper markups can diverge significantly from underlying business health. Generally speaking, it is the crossover funds and the funds with the most competitive internal cultures who struggle to correctly assess performance because they default to applying the same methodology of performance assessment across their public and private deal teams.

• Compensation: Building a world-class venture firm requires world-class talent, but world-class talent is expensive and hard to retain. VC firms tend to make three common mistakes when it comes to compensation. First, GP’s often will allocate too much compensation to themselves relative to high-performing employees. Inevitably, the high-performing employees realize that they are better off joining another firm that is more meritocratic or spinning off to start their own fund. Second, GP’s will pay well on salary but not allocate enough carry to their high-performing employees. Not only is the upside potential associated with carry higher than cash compensation, but it locks in investors and makes them feel like true owners of their fund. Third, GP’s fail to recognize junior talent and over allocate compensation to underperforming senior talent. Funds that have been around for a long time most often run into this problem, and often struggle to share upside with junior employees who are responsible for the go-forward success of the firm.

• Carry structure: Carry structure is related to compensation but deserves a separate section. Funds approach carry distributions to partners in different ways. In some cases, carry is allocated (or at least weighted) on a deal-by-deal basis – i.e., a partner is rewarded more for their deals and less for other partners’ deals. In other cases, carry is solely allocated to the GP, who elects to pay out large bonuses to partners at his/her discretion. Lastly, in rare cases, the fund is an equal partnership, meaning all carry is evenly split between all partners at a fund. This last approach, in my opinion, is optimal. It fully aligns partners’ incentives with the incentives of the fund and creates a collaborative environment between the decision-makers at a fund.

• Culture: In an ideal world, VC firms have collaborative internal cultures, where partners are supportive of each other and each other’s portfolio companies. Too often, however, VC firms inadvertently create highly competitive cultures, and partners operate out of self-interest.

Why Founders Should Care

What happens when a VC fund is self-unoptimized?  More importantly, who should care about it?

For certain stakeholders, the consequences are more obvious. For example, VC funds that are self-unoptimized will struggle to retain key talent and will produce poor returns - two outcomes that are clearly not ideal for the fund itself and its limited partners.

But what about founders? Should they care whether a VC fund is self-optimized or not? When I ask what kind of VC a founder is looking for – I rarely hear about any of the variables I’ve discussed in this post. Instead, I hear things like “I want a VC that can find me customers,” or “a VC that can help me hire,” or “a VC that understands my industry.”

To be clear, these asks are all super important. My point is not that founders should stop asking about customers, new hires, and industry-expertise. Rather, I believe founders should be asking more than that – specifically whether a VC firm knows how to run itself.

What are the consequences of taking money from a VC that is self-unoptimized? At a conceptual level, if a VC demonstrates it can’t run its own business well, I don’t believe they should be a in position to advise founders on how to run their own businesses. But I’ll also be more direct. Below are a few specific consequences founders face when partnering with self-unoptimized VCs.

Parting Thoughts

In this post, I’ve shown how VCs are self-unoptimized and how these dynamics impact founders. But I haven’t answered why funds self-unoptimize in the first place. Why should VCs choose to build their firms in ways that hurt returns? Moreover, why would LPs give self-unoptimized VCs capital?

In brief, there are multiple factors at play:

1. It takes a long time to validate or invalidate a VC’s strategy, largely due to the long fund cycles.

2. Many long-term suboptimal decisions are short-term enticing (e.g., receiving more management fees from a larger fund).

3. Maintaining an optimal state is difficult (e.g., maintaining a meritocracy, in terms of both compensation and decision-making, becomes increasingly more difficult as a firm matures, particularly given the natural generational turnover in technology).

4. Outside of fund size, LPs traditionally do not exert meaningful influence on the other variables. Maybe they should.

5. It’s just not on every GPs radar. I genuinely believe a lot of managers simply don’t think about these dynamics.

I titled this blog post “Building the World’s Best Venture Firm.” You’ll notice that I never mentioned that Vine is the world’s best venture firm. This is because I truly believe no one is a finished product. I do know, however, that we, at Vine, are obsessed with providing our founders the best user experience possible. And one of the subtle ways we do that is by constantly trying to self-optimize.

Subscribe now

An exit is the key measure of a startup’s success. For founders, regardless of whether you’re mission-driven or financially-driven, your company’s IPO or acquisition demonstrates that you built something special. Similarly, for investors, success means generating exceptional cash on cash returns through liquidity events (M&A, IPO, or secondary sale).

In 2023, I began tracking startup exits in earnest after the announcement that MosaicML had been acquired by Databricks for $1.3bn. With Adobe - Figma called off, MosaicML was the most successful strategic acquisition for a venture backed company in 2023…at least so it seemed. After all, the acquisition price was over $1bn (a significant premium to the prior two funding rounds) and the time to exit was incredibly short (~3 years from company inception and <2 years from first fundraise). By all accounts, this should be a career-defining investment for anyone who backed the company. So what went wrong?

As a colleague pointed out, the fund math for Lux Capital, who led MosaicML’s seed round, didn’t make much sense. In short, Lux turned ~$10M into ~$175M, but only returned ~0.25x of their enormous early-stage fund. In other words, Lux will need to find 3 more MosaicMLs just to breakeven.

This case fascinates me to this day. If one of the highest profile strategic acquisitions of the last 2 years couldn’t produce a fund-returning venture outcome for investors, then what was the state of the rest of the market? What do exits look like for founders and investors in a down market and what lessons can be learned?

M&A

I first aggregated all the notable startup acquisitions of 2023, where “notable” is defined as a venture-backed company that 1) was acquired by a strategic and 2) had raised at least ~$10M in venture funding to-date. The following chart illustrates 50+ acquisitions from 2023, including information on the target, the acquirer, and the acquisition price. On the right-hand side, I included details on the most recent venture round prior to the acquisition as well as an estimated return for the lead investor in that round. It's worth noting that the estimated returns are directional in nature, and make several simplifying assumptions when necessary – i) dilution on latest round equals 20%, ii) announced fundraises were completed at a single valuation, and iii) acquisitions were completed in all cash.

Key M&A Takeaways

• Mega M&A outcomes increasingly blocked by regulators: The Adobe - Figma acquisition would have been the most successful strategic venture outcome in 2023 had it not come under intense regulatory scrutiny in the UK and EU. The implications are clear – if you want to build the next important tech company, you should assume your only exit option will be to take the company public. Not only does this take away a significant amount of optionality (I would argue there exist many incredibly valuable businesses that can/could never go public), but there’s no guarantee that IPO markets will be favorable (or even open) should you build a successful company.

• Estimated median IRR / MOIC for the most recent investors in businesses that completed strategic acquisitions in 2023 is upper bounded by 17% / 1.3x: This is a likely an overestimate, as most acquisitions did not report a price and were immaterial. In short, investors on average are barely (if at all) seeing returns on their investments that exited via acquisition in 2023. From another perspective, the average acquisition price needs to be >3.0x higher than what acquirers are willing to pay today in order for VC’s to achieve a typical return target of 4.0x.

• Fund size matters: As noted earlier, the MosaicML deal was a spectacular deal for Lux, but only returned 1/4^th of their latest early-stage fund, meaning they would need to back three more similar deals just to give their LP’s their money back, ignoring management fees. I ran through a similar exercise for other “successful” venture acquisitions. Coincidentally, these Seed/Series A deals had been completed by VCs with oversized funds and the math looked even worse.

• Valuation matters: It feels like too many VCs have yet to grasp this concept - or at least they don’t apply it in practice. Looking at the above table, multi-billion-dollar exits still would not have returned some of these funds. What should be simple mental math is apparently absent from many IC and LP discussions.

• Founders did slightly better than VCs, but its highly context dependent: In acquisitions with a large pref stack where the price is at a discount to the prior round, founders (and employees) likely struggled to capture value from the acquisition. This can be highly demoralizing. Imagine dedicating years building a company that’s valuable, but being unable to receive a financial reward for your hard work. However, in acquisitions with smaller pref stacks and higher valuations, founders will fare better than their investors. Furthermore, it’s worth noting that founders will often renegotiate meaningful economic incentives in excess of existing guarantees as part of the acquisition. The relationship between founder and investor is crucial during these processes. A good VC will work with the founder to reach an outcome that optimizes for everyone. Unfortunately, there are numerous horror stories where VCs work against founders’ interests (e.g., attempting to block founder-friendly deals).

• Bad exits today have downstream consequences for future founders: Limited partners are likely to pull back their VC allocation, meaning less capital will be available for founders. Furthermore, VC’s may start avoiding industries (or stages) that struggle to produce fund returning outcomes. Industries like cybersecurity for example, where successful outcomes are significantly dominated by M&A (rather than IPO), may find fewer capital sources willing to invest.

IPO

The IPO market was largely closed in 2023. Instacart, Klaviyo, and Arm were the primary exceptions to this rule. For a brief period of time, their public offerings were seen as a sign that the IPO market for tech was finally reopening. Looking back, these proclamations were premature. Furthermore, for Klaviyo and Instacart, the two traditionally venture-backed companies, performance since IPO has been suboptimal with declines of 17% and 22%, respectively, as of 12/31/23.

Key IPO Takeaways

• There are a number of IPO-ready companies unable to go public: Stripe, Databricks, Reddit, Chime, Klarna, Discord…the list of companies that have indicated or are rumored to go public “soon” goes on. It remains to be seen what is needed for a successful IPO market to return. If the public markets stabilize, interest rates are cut as promised, and a few early movers have strong public debuts, perhaps 2024 will shift the narrative in the way that people were expecting in 2023.

• IPOs promise to generate liquidity, but not returns: As The Information noted, the Instacart IPO price was far below the implied share price of the numerous investors who backed the company after 2018. IPOs will likely give investors the opportunity to exit their investments, but for many this will still represent a loss.

• Antitrust regulation may create a private market “dead zone”: A good rule thumb is that eventually the bad companies will fail, the ok companies will get acquired, and the great companies will IPO. But what happens if you’re a decent company? Too big to get acquired without a regulatory snafu, but too small (or perhaps growing too slowly, or associated with the wrong narrative, etc.) to produce a successful IPO? People talk about the number of “zombie unicorns” created in the ZIRP era, but no one talks about the companies that are better than the zombies but worse than the all-stars. In a world where the IPO market returns, but only tepidly, it’s unclear to me where, and when, these companies will find a home.

A Brief Note on the Secondary Market

The secondary market has been open and active. If you aren’t transacting shares of a high-flying AI company, the secondary market has been a buyer’s market, and there are too many unhappy sellers, several of which are looking to unload entire portfolios at significant discounts. As time goes on and the value potential of each private company becomes more clear, the Bid-Ask spread will narrow and I expect more (fire) sales to continue. However, a proper assessment of the secondary market likely requires a separate post and a lengthy deep dive into opaque data, so I’ll skip it for now.

Parting Thoughts

Throughout this post, I talk about startup exits through the lenses of different stakeholders – founders, VCs, and sometimes VC limited partners. If you fall into one of these camps, it may be easy to ignore the perspectives of the others. For example, founders often don’t spend their time thinking about VC fund math or how limited partners might react to the current exit environment.

But this would be a mistake. The VC – limited partner – founder relationship is symbiotic. All three stakeholders need to be happy for this ecosystem to thrive. As long as tech businesses require venture funding to scale (TBD whether AI changes this dynamic, but that’s a conversation for another day), any negative outcome for a stakeholder, whether investor or founder, is a loss for everyone.

The AI ecosystem is often described in “layers” – the Application Layer (e.g., Codium), the Tooling Layer (e.g., LangChain), the Hardware Layer (e.g., Nvidia), and the Model Layer (e.g., OpenAI).

The primary reason the layer framework became popular – aside from giving VC’s an excuse to create market maps – is that most AI companies fit neatly in a single layer. After all, modern AI companies are nascent and usually good at only one thing. But talk to any founder and you will realize that most AI companies eventually want to do more than just one thing. In other words, the future of AI is “multi-layer.”

Companies that occupy the model-layer are the most likely to go multi-layer – and the least likely to admit it. Furthermore, this transition will not be driven by choice, but will be born out of necessity. Unfortunately, too many founders still see the model-layer as enabler rather than competitor, a mistake that has led (and will continue to lead) to suboptimal outcomes for AI startups.

Foundational Model or Business Model?

It’s first important to understand why model-layer companies are highly incentivized to go multi-layer:

• Open-source alternatives are numerous and robust: Open-source alternatives to state-of-the-art closed source models are increasingly common and comparable in terms of accuracy, context window length, use cases, etc. Not to mention the myriad open-source small-scale models that exist for more specific tasks (and can be superior in those niches to their more generalized, large-scale cousins).

• Foundational models are expensive with unclear ROI: The cost to create cutting edge LLM’s is significant (e.g., the cost to train GPT-4 was ~$100M), so much so that model layer companies are searching for ways to improve the next generation of models beyond scaling known techniques. For an investment of this scale to make sense, the revenue generated from these models needs to outweigh the costs to create them. Today, that cost is high, but the longevity of those models remaining cutting edge is surprisingly short. Assuming longevity correlates to revenue, model-layer companies will need to generate significant revenues in short periods of time or meaningfully drive down costs to justify the economics.

• Perpetual demand for higher performance is not guaranteed: I’ve had hundreds of conversations with users of AI models across multiple industries. Regardless of use case or modality, many buyers have a threshold of performance that they care about (I like to call this the “Threshold of Commercial Viability”). For example, a call center with thousands of hours of audio data wanted to turn that audio into text but was indifferent between solutions that provided >90% accurate speech-to-text accuracy (90% accuracy is shockingly low!). Once such a threshold is reached, customers only care about price, which is the definition of a commodity market. Commoditization will further accelerate when open-source alternatives also reach the Threshold of Commercial Viability.

• Sky high valuations will force expansion: Many model-layer companies have unicorn valuations with minimal (if any) revenue to show for it. Due to their ongoing capital requirements and margin profiles, they are unlikely to receive long-term valuation multiples in excess of public SaaS companies. Consequently, the model-layer faces pressure to grow their revenue and margins and will seek new opportunities in adjacent layers of the stack.

• Going multi-layer makes simple operational sense: Better AI applications can be built with better foundational models (especially if those models are inaccessible to other applications). Those applications can then become defensible through other means, such as via network effects, customer service, or bundling. Alternatively, AI tools can make models sticky and easier to use. In other words, even if the model-layer primarily focuses on models, it still makes sense to expand to other layers to support that effort.

The Multi-Layer Today

To what extent has the multi-layer transition already started? Below, I’ve mapped out the product suites of the major model-layer companies, occasionally providing color on their recent strategic directions and decisions.

OpenAI is the most interesting example of a model-layer company going multi-layer, primarily because its public positioning (to this day) put it at odds with this transition. OpenAI was founded as a non-profit organization whose “mission is to ensure that AGI benefits all of humanity.” Even as the company has shifted from non-profit to for-profit, it continues to position itself as an enabling layer of the AI ecosystem, providing powerful research models that companies and people can use and build on top of. OpenAI even has an investment fund dedicated to backing early-stage founders.

The release of ChatGPT precipitated the first major shift within OpenAI towards the multi-layer. No one expected ChatGPT to succeed in the way that it did – perhaps not even OpenAI. Since its release, OpenAI has poured significant resources into the product, releasing Plugins, ChatGPT for Enterprise, and most recently multi-modality.

Microsoft’s investment into OpenAI was the company’s second major shift towards the multi-layer. Today, OpenAI’s finances and resources are intrinsically intertwined with Microsoft, and it comes as no surprise that the two companies have collaborated in releasing a slew of applications.

Ironically, OpenAI’s applications have brought it into direct competition with the ecosystem it brands itself as enabling. The first iteration of ChatGPT, for example, was directly competitive (and fairly destructive to) Jasper, a major early corporate partner. And this is just one example of many. It is sadly a common conversation topic within VC circles to discuss the number of startups that will likely fail with each new OpenAI application release.

Unlike OpenAI/Microsoft, the connection between Google Deepmind and Google has been obvious from day one (hint: they’re literally the same company). It is then no surprise that Google has released many applications powered by the foundational models created by its research teams. While Google Deepmind’s multi-layer strategy is more obvious than OpenAI’s, there are a few nuanced items worth noting:

• First, Google Deepmind was initially perceived as slow-moving, and it was believed they were stalling the release of their technology due to (1) reputational risks and (2) potential cannibalization of core revenue streams such as search. However, Google has recently taken a more aggressive approach to model and application releases – signaling a competitive showdown with OpenAI and others.

• Second, Google has started to build LLM-enabled applications beyond its existing core application suite. Vertex AI (ML workflow platform) and SynthID (image watermark tool) are two examples where Google has built new AI applications/tools vs. upgrading existing Google products. If this trend continues, entrepreneurs may struggle to identify categories with limited competitive risk from the model-layer.

• Third, after merging Deepmind and Brain, Google has been internally pushing its researchers to spend more time on commercial research areas vs. non-commercial research areas – a shift that emphasizes Google’s commitment to becoming a power player in the multi-layer.

Anthropic has consistently portrayed itself as an application-first AI company with its own models. Relative to OpenAI and Google Deepmind, Anthropic directly markets Claude for specific end markets such as customer service, legal, coaching, search, back-office, and sales. Amazon’s recent multi-billion-dollar investment in Anthropic is remarkably similar to Microsoft’s investment in OpenAI, and I anticipate the relationship to evolve in a similar way, further accelerating Anthropic’s application suite.

Less needs to be said about Cohere, Adept, and AI21 Labs – their model and application suite, while smaller, is similar to that of their larger peers.

Multi-Layer Thoughts and Predictions

• Model-layer companies are well into the multi-layer transition: The model-layer is not content to sell access to its models. Instead, there is an accelerating shift towards other layers in the stack. Most model-layer companies are open about their application aspirations, though others (e.g., OpenAI) demonstrate inconsistencies between their messaging and actions. So far, this has yet to elicit pushback from the community.

• Resources + time correlate with application density: The model-layer companies that have gone the most multi-layer are those with the most funding, research resources, time (i.e., were founded earlier), and ties with big tech. It is safe to assume application density will continue to increase as time goes on (and assuming capital continues to fund these businesses).

• Model-layer companies have significant application parity: Despite having limited product suites, there is a surprising amount of application parity in the model-layer (e.g., the chatbot/knowledge assistant is nearly ubiquitous). This is likely driven by (1) nascent and limited customer demand, (2) unproven and untested use cases, and (3) limited moats among this cohort.

• Prediction #1 – Every application that logically fits into the Big Tech product suite will be built by the model layer: If you’re a founder, don’t just ask yourself whether a model-layer company will compete with you, ask yourself whether your product deserves to exist within the context of the Google, Microsoft, or Amazon product suites. These are the most obvious applications that will be built in the next 1-2 years.

• Prediction #2 – Model layer companies will build more products in the tooling layer: Many view the tooling-layer as a place where model-agnostic newcomer solutions will thrive. Model-layer companies, however, may view the tooling-layer as a means to improve user experience. Currently, Google is the main player currently innovating in this space. I believe all model-layer companies will build their own tools down the line. It’s also possible those tools may be model agnostic.

For the sake of playing devil’s advocate, it’s also worth noting a few potential obstacles to the multi-layer:

• How valuable is model performance? Earlier, I described the concept of the “threshold of commercial viability,” which I argued would eventually limit demand (and thus monetization potential) for more powerful models. There are a couple ways in which this plays out differently:

  • If it turns out the threshold of commercial viability is years away for most industries, the model-layer may not need to go multi-layer for a long time, allowing newcomers to scale in the meantime.

  • The value of AGI remains very unclear. If AGI proves reachable and valuable, the model-layer may allocate fewer resources to applications and tools. The applications and tools they build may also be easy to compete with.

  • The “threshold of commercial viability” may be irrelevant for certain industries. Some industries may follow the trajectory of the call center example I provided. But we can imagine other categories (e.g., cybersecurity, finance, etc.) where incremental performance gains, no matter how small or hard won, provide value into perpetuity.

• Will technical achievements rewrite the cost equation? High costs are a critical consideration for model ROI and thus a potential driver for the multi-layer. If more performant models are no longer gatekept by high-cost requirements, model-layer companies may no longer be incentivized to go multi-layer.

Why the Multi-Layer Matters

People currently perceive a distinction between the model-layer and companies building applications and tools around it. Because of the lack of differentiation model builders and economic pressures, I believe this distinction will disappear over time and would argue we’re already witnessing its effects.

I generally believe people draw too many parallels between the future of LLM’s and the history of the cloud. However, in this case, I’ll allow a brief comparison. The cloud providers, in their early days, provided storage and compute. Today, the cloud providers provide hundreds of products and managed services around storage and compute. This isn’t to say that startups haven’t succeeded in any of these categories or that the cloud providers are clear market leaders in everything they offer. But it tells a story of rapid horizontal and vertical expansion that will likely repeat itself within artificial intelligence.

I previously wrote that incumbents are well-positioned to own the AI landscape. Founders need to forget the model-layer and think in terms of the multi-layer. Find the platforms that deserve to exist outside the multi-layer or that have a reasonable strategy to compete with the multi-layer. Leverage your enablers but treat them as competitors.

All founders, either explicitly or implicitly, believe the following to be true:

1. Customers will use my product instead of building it themselves

2. Customers will use my product instead of my competitors’ products

Fortunately, innovation has historically favored the founder. Despite deeper resources, customers and competitors are often too constrained to outcompete incredible new products. This cycle of innovation has historically enabled entrepreneurs to realize their visions and build valuable (and, at times, generational) companies. Today, entrepreneurs and venture capitalists believe artificial intelligence will drive the next wave of innovation. Unfortunately, most entrepreneurs and VCs are missing the bigger picture.

AI is poised to drastically lower the constraints that have previously enabled startups to succeed. Without these constraints, it will become harder for founders to build generational companies, especially in software.

This post explores the future of software – what some call “Software 3.0.” My goal is to illustrate why Software 3.0 poses unforeseen risks to founders. I believe that by keeping these risks in mind, founders can direct their incredible talents towards solutions and markets that deserve to be won by newcomers in the current age of AI.

What is Software 3.0?

Software 1.0 is how we typically think about software. The Software 1.0 programmer writes a set of instructions in a coding language, those instructions are converted into machine code by a compiler, and a computer reads and executes the machine code.

Software 2.0 is a term coined in a 2017 blog post by Andrej Karpathy. If code is the primitive for Software 1.0, then neural networks are the primitive for Software 2.0. To paraphrase Karpathy: The Software 2.0 programmer specifies some goal on the behavior of a desirable program, writes a rough skeleton of the code (i.e., a neural net architecture) that identifies a subset of program space to search, and uses the computational resources at their disposal to search this space for a program that works.

Software 3.0 is the future of software. The Software 3.0 programmer provides a set of instructions and a dataset, and an AI agent uses these inputs to generate a program. Though it will take a while to realize the full potential of Software 3.0, the building blocks are already here. Today, you can query ChatGPT for a piece of code, use GitHub Copilot to suggest code in real-time, or automatically debug your code with Codium. You can also give autonomous agents a goal written in natural language, and they will attempt to achieve it by breaking it down into subtasks. Sometime in the future, we may be able to prompt an agent to create an entire tool, workflow, application, and even a software platform with a high degree of accuracy and reliability.

• For further reading on Software 3.0, I recommend reading this post by Itamar Friedman, co-founder and CEO of Codium.

For the purposes of this post, we won’t debate the details of Software 3.0 – the timeline, its precise capabilities, etc.

Instead, we will try to understand how Software 3.0 can impact the future of entrepreneurship. As mentioned, all founders share the same two beliefs: (1) customers will use my product instead of building it themselves and (2) customers will use my product instead of my competitors’ products. But how well will these beliefs hold up in the era of Software 3.0?

To Buy or To Build?

I believe customers of startups will increasingly become competitors to startups. Founders who have started companies in the past year or are in the process of starting companies need to be especially aware of this threat. 

To understand why, let’s look at the primary reasons why companies have historically purchased software instead of building it themselves:

Figure I: Nine reasons why companies buy third-party software rather than build it in-house

To assess the impact of Software 3.0, it’s important to understand which of the above nine reasons are impacted by artificial intelligence. It stands to reason that if artificial intelligence can reduce some barriers to building software (without raising others), then companies will build more software than buy it in the future.

Figure II: Predicting how Software 3.0 will impact the build vs. buy decision tree

As we can see, the outcomes are mixed. Building internally with Software 3.0 isn’t necessarily better than buying third-party software. Most likely, companies will in-house simple software tools that can easily be built with Software 3.0 but will not drastically in-house more complex platforms that are purchased for other reasons beyond resource limitations.

But software founders should not start celebrating just yet. Figure II only measures a customer’s willingness to replace a mature software solution with an internal version. In reality, early-stage companies rarely demonstrate these nine characteristics. Early-stage startups don’t have network effects, they have nothing to bundle, their security and privacy standards are worse than mature alternatives, they may raise regulatory concerns, they certainly don’t have a brand, and they have very few contracts. Instead, they are usually more willing or able to direct resources to a specific problem. And in some cases, they might have better customer service and integrations.

Figure III: Predicting how Software 3.0 will impact the build vs. buy decision tree for early-stage startups

In other words, early-stage software companies have far fewer characteristics that drive customers to purchase their product. And of those characteristics, the majority are likely to lose out to Software 3.0.

For early-stage entrepreneurs, you need to be thinking several steps ahead. It’s not enough to have a good idea and the resources (people, knowledge, and time) to devote to it. Instead, you need to build a product that lends itself to some other form of defensibility to prevent your potential customers from building it internally.

The Rising Threat of Competition

Preventing your product from being in-housed is only half the battle. Once customers decide they need to purchase solutions like yours, you need to convince them to purchase yours vs. someone else’s.

Unfortunately, competitors will become far more dangerous in the Software 3.0 paradigm. This includes both early-stage and mature competitors:

• Early-Stage Competitors:

  • Code generation will reduce IP and unique insights: LLMs are really good at distilling the corpus of human knowledge and making it available to a broad set of people. As artificial intelligence capabilities improve, it will be more difficult for founders to identify unique insights and then build unique products on top of those insights. Furthermore, unique insights, once identified, will become more easily replicated with fewer resources.  

  • Venture capital is oversaturating the market: VC investment is flowing freely in AI, and that capital is rapidly funding many competitors in many categories. This trend is always present in tech but becomes particularly exacerbated during hype cycles.

• Mature Competitors:

  • Code generation also enables incumbents: Code gen doesn’t just make the early-stage market more competitive; it also unlocks massive efficiency gains among mature companies with large engineering teams. Every early-stage founder should imagine a world where their mature competitors become at least twice as efficient/innovative as they have been in the past.

  • Customers are seeking vendor consolidation: Customers are tired of software bloat. In a high interest rate environment, they’re also prioritizing cost savings and don’t want to pay for too many solutions. This pressures markets to consolidate around mature market leaders.

  • Incumbents are seeking the next wave of growth: Public growth companies (or growth companies seeking to go public) are seeking ways to maintain strong top-line growth to support a premium valuation, particularly in an environment where valuation multiples have fallen from prior highs. Expect mature companies to use artificial intelligence to re-imagine their core product or move horizontally into tangential product lines.

  • Stakeholder pressure to invest in AI capabilities is high: The hype around AI has impacted all companies, not only early-stage startups. Mature companies are being pressured by investors, board directors, public shareholders, and customers to implement AI technologies. In Q1 2023, more companies cited “artificial intelligence” in earnings calls than ever (nearly double the average in 2022).

  • Not all mature competitors look the same: The definition of incumbent has changed over time. In the late 20^th century, an incumbent may have been a non-internet enabled competitor. In the early 21^st century, an incumbent may have been a competitor not utilizing the cloud, or not utilizing mobile. Today, an incumbent may be technologically illiterate (this is rare), or a modern public technology company that has scaled in the last decade, or a younger technology company that has raised billions of dollars in private capital. Most early-stage entrepreneurs and VCs don’t give enough credit to the innovation capabilities of the latter two groups. Here’s a simple test – determine the revenue growth rate of your mature competitors in the market. If that number is >20%, you should treat them as extremely formidable adversaries.

Parting Thoughts

As a techno-optimist, I believe technology creates value and generally makes the world a better place. But value is not always dispersed evenly.

Software 3.0 – a new paradigm where code is written through natural language prompts – is poised to deliver meaningful revenue and efficiency gains to both software users and mature software companies. However, the job of the founder, which has never been easy, will become even more difficult. Most technologies have been driving forces for innovation. AI, on the other hand, will often be a driving force for industry consolidation and market share concentration.

What is my edge? Who are my competitors? How easily can someone copy me? These are questions from a bygone era.

Instead, founders should consider these questions in the context of the future. Is my edge defensible in the Software 3.0 paradigm? How dangerous can my competitors become using Software 3.0? Will Software 3.0 give companies the ability to copy me more easily?

Successful founders will be the ones who can answer these questions and navigate these obstacles. If you want to chat about Software 3.0 and defensibility in AI, don’t hesitate to reach out!

2023 has been a year full of milestones, and not particularly happy ones at that. In a little over four months, we've witnessed an unprecedented global banking crisis, whose causes we will be analyzing for decades and whose consequences are still not fully known. Nobody, except perhaps for quick-fingered Wikipedia editors, were excited to see three new additions to the list of the largest bank failures in U.S. history - First Republic Bank, Silicon Valley Bank, and Signature Bank - which have claimed the ignominious positions of second, third, and fourth place. Unfortunately, the turmoil and fear seem far from over.

It has been a long time since the last major banking crisis during the Great Recession. Consequently, everything about 2023 feels “new,” especially to a younger generation of professionals who came of age in a world subsidized by low interest rates and free capital. According to consensus, this crisis is new because it has occurred in the digital age, an era where people can derail a financial institution from the comfort of their couch, rather than physically running to their local neighborhood branch and demanding cash.

But just because something feels new does not mean it is new. That's the big question for this post. Is the 2023 banking crisis actually new? Is it actually unprecedented? Are the failures of First Republic, SVB, and Signature so unique in the context of our most infamous financial institutions?

Five (In)Famous Case Studies

To answer these questions, we'll start by looking at five major bank runs in U.S. history - each from a different time period and under different macroeconomic and microeconomic constraints. Apart from SVB, each bank in this list set the record for the largest bank failure in U.S. history at the time.

The Bank of the United States (1930)

What was it? The Bank of the United States was the largest commercial bank in the U.S. by deposits and customers in 1930. At the time of its failure, the bank held $268M in deposits (~$5bn today) and served over 440,000 customers, who were primarily working-class immigrants living in New York City. The bank run on the Bank of United States' Bronx branch is often cited as the trigger that initiated the collapse of the banking system during the Great Depression.

Why did it fail? After undergoing rapid expansion in the late 1920's fueled by a handful of mergers, the Bank of the United States faced a handful of legal difficulties. In conjunction, a series of smaller bank runs in the southeast of the U.S. as well as the collapse of the stock market set the stage for an atmosphere of fear. According to historians, when a Bronx businessman was unable to sell his Bank of United States stock back to a bank manager in the Bronx branch, the businessman spread a (false) rumor that the bank could not afford to repurchase the stock. The rumor spread like wildfire, prompting a massive bank run with over 20,000 customers queued to withdraw funds at the Bronx branch alone. The following day, the bank was temporarily closed, and then permanently shuttered when support financing and a merger offer failed to materialize. Ultimately, the bank was illiquid, not insolvent - and depositors eventually received 93 cents on the dollar for their deposits when all was said and done.

What was the timeline? On the morning of December 10th, 1930, a Bronx businessman spread the rumor that the Bank of the United States lacked liquidity. That afternoon, over 20,000 individuals lined up to withdraw from the Bronx branch. The next day, December 11th, the bank directors closed the bank in anticipation of a massive run and gave the New York Superintendent of Banks control over its assets. When a merger was unable to be negotiated, the Bank of the United States was suspended on December 13th, 1930.

Franklin National Bank (1974)

What was it? Franklin National Bank was a bank based in Long Island, New York that failed in 1974. At the time of failure, the bank held ~$3.7bn in assets (worth $21.bn today).

Why did it fail? It is impossible to do justice to the lurid failure of Franklin National Bank in a single paragraph. The long version of the story involves the mafia, the free masons, the Nixon administration, drug cartels, a cyanide pill, and an incredible number of broken laws. In short, however, an extremely shady individual named Michele Sindona purchased a controlling stake in Franklin and, through a series of poor and often illegal decisions was responsible for many losses at the bank, prompting a bank run and ultimately insolvency.

What was the timeline? In early May 1974, Franklin National Bank disclosed to the public significant losses in foreign exchange markets in excess of $60M. Over the next 14 days, deposits declined by ~$325M, or ~11% of total deposits. To cover withdrawals, the bank borrowed over $1bn from the Federal Reserve Bank in May. However, deposit outflows continued for the next several months, reaching a ~50% decline from peak at the start of September 1974. On October 9th, 1974, Franklin National Bank was declared insolvent.

Continental Illinois (1984)

What was it? Continental Illinois is the fifth largest bank to fail in U.S. history. Continental Illinois was a commercial bank headquartered in Chicago that was the seventh largest commercial bank in the U.S at its peak. At the time of its failure, the bank held $40bn in assets (worth $116bn today).

Why did it fail? Following a decade of rapid growth in the 1970's, Continental Illinois made a series of risky investments that proved its downfall. Due to an old regulation that prevented banks from expanding beyond state lines, banks often purchased loans from other banks across states. In this case, Continental Illinois purchased $1bn in risky energy loans from Penn Square Bank, which was based in Oklahoma. Continental Illinois also owned risky assets in several developing countries. Soon after, Penn Square Bank failed, and there was an international debt crisis spurred by Mexico's default. The bank attempted to rightsize its balance sheet, but fear sparked a bank run on the bank's deposits, the majority of which were uninsured. It’s also worth noting that ~40% of Continental Illinois' deposits were foreign leading up to the failure – and historians believe these creditors were more likely to withdraw relative to domestic creditors.

What was the timeline? Continental Illinois' troubles came to light in a poorly received earnings release in April 1982. This release, combined with the failure of Penn Square Bank (July) and Mexico's default (August), resulted in a slew of credit and stock downgrades. In late April 1984 (two years later), the bank publicly announced an increase in nonperforming loans (i.e., a loan that is unlikely to be repaid in full). On May 7th, 1984, rumors arose that Continental Illinois would fail or would be forced to seek a merger. A 9-day bank run commenced on May 10th, 1984 in which 30% of Continental Illinois' existing funding fled, and was replaced by a significant amount of federal assistance from the FDIC as well as support from a coalition of private banks. The government effectively owned and ran Continental Illinois until it finally disposed of its shares in 1991, and the bank was purchased by BofA in 1994.

Washington Mutual (2008)

What was it? Today, Washington Mutual (aka WaMu) is remembered as the largest bank failure in U.S. history - both on a nominal and inflation-adjusted basis. Prior to its collapse in 2008, WaMu was the sixth largest bank and the largest Savings and Loan Association in the U.S. Unlike a traditional commercial bank, Savings and Loan Associations primarily focus on residential mortgages (in addition to other traditional banking services). At time of failure, WaMu held $307bn in assets (equivalent to ~$422bn today).

Why did it fail? To some extent, the failure is self-explanatory - it was 2008 and WaMu specialized in residential mortgages. But to be more specific, WaMu had significant exposures to subprime mortgages because they catered to higher risk consumers and approved variable rate loans despite a lack of provable income and assets. Additionally, WaMu had a lot of business in California (where housing fared worse than average) and had rapidly expanded its physical presence in the years leading to the crash. When housing prices fell, WaMu wrote down billions in defaulted mortgages and struggled to raise cash, which was exacerbated by the collapse of the secondary market for mortgage-backed securities. The final nail in the coffin was a bank run led by depositors, who withdrew $16.7bn over the course of ten days. The FDIC determined that WaMu lacked sufficient funds and seized the bank. JPM eventually stepped in and assumed WaMu's assets.

What was the timeline? Housing prices peaked in early 2006. By December 2007, average home values had declined 6.5% from the prior high. The same month, WaMu wrote down its home lending unit by $1.6bn, cut 6% of its workforce, and closed ~50% of its home-loan offices. In 2008, the Bear Sterns takeover was negotiated in March and IndyMac Bank failed in July. In the first two weeks of September, Fannie Mae and Freddie Mac were taken over by the government, Merill Lynch was sold to BofA amidst a liquidity crisis, and Lehman declared bankruptcy. The latter, which occurred on September 15th, 2008, kickstarted the ten-day long bank run on WaMu, culminating in its failure. Most of those deposit withdrawals were retrieved via internet-based electronic banking and wire transfer.

Silicon Valley Bank (2023)

What was it? SVB was a commercial bank founded in 1983 in Santa Clara, California that specialized in banking technology companies. As of 2021, SVB claimed to bank ~50% of all US venture-backed startups and had relationships with a meaningful number of venture capital firms. At the time of its failure, SVB held ~$209bn in assets, positioning it as the third largest bank failure in U.S. history as of writing.

Why did it fail? Prior to the technology downturn that started in November 2021, SVB had both grown rapidly and catered to a specific clientele. Like any bank, SVB invested client deposits in return-producing securities, like bonds. When the federal reserve rapidly raised interest rates to combat inflation, the current value of those bonds declined. Why? Because if you want to sell a bond that pays a low interest rate relative to market, you will need to sell that bond at a discount to convince someone to buy it. That doesn’t really matter if you are planning to hold the bond until maturity. Unfortunately, SVB’s customer deposits were dwindling because startups were not raising more capital and were continuing to burn through existing cash. To cover a decline in deposits, SVB sold securities from their portfolio for a $2bn loss and would undertake a share sale to raise further cash. Instead of rightsizing their balance sheet, SVB instead spooked the market – its share price collapsed, and existing customers rushed to pull out their deposits. Soon after, the California Department of Financial Protection and Innovation took possession of SVB.

What was the timeline? In mid-2021, a surge in inflation impacted the global economy. On March 16^th, 2022, the Federal Open Market Committee enacted the first of nine rate hikes, increasing the federal funds rate from 0.25% to 5.25% as of writing. By December 31^st, 2022, SVB had unrealized losses of ~$15bn for securities held to maturity. On March 8^th, 2023, SVB announced it had sold $21bn of securities at a ~$2bn loss and was further planning to raise capital from a stock sale. On March 9^th, SVB stock declined 60% and customers ran on the bank, initiating ~$42bn in withdrawals. On March 10^th, regulators took control of SVB, shutting it down.

Drawing Conclusions

• Bank runs are as much a function of perception and psychology as risk: Playing “what if” is a fool’s game in history – that being said, it’s not immediately obvious that all failed banks would have failed without a bank run on deposits. Customer confidence may have saved the Bank of the United States, Continental Illinois, and SVB (Franklin Illinois and WaMu, on the other hand, had serious balance sheet problems). Skeptics scoffed at the initial SVB fears, pointing towards recent balance sheet data demonstrating solvency. Nonetheless, liquidity disappeared when enough customers demanded their money back. Legislation aimed at strengthening our financial institutions must not only focus on how banks are run, but also must incentivize customer confidence and trust.

• Running on your bank is optimal according to game theory: A bank run is a variation of the most famous problem stated in game theory – the Prisoner’s Dilemma. Imagine a scenario where a bank has two depositors and if either depositor attempts to withdraw, the bank fails. If one depositor withdraws, the other ends up with nothing, and if both depositors withdraw, they only get a portion of their money. In this scenario, if the depositors have trust in the system, it is optimal for them to stay at the bank, but as soon as they think the other will leave, it becomes optimal to withdraw. This is how bank runs start, and this is why they are so hard to stop. It’s just game theory. Depositors are never to blame in a bank run, despite a few recent efforts to paint them so.

• Bank runs are fast-moving events: The notion that bank runs occurred over prolonged periods prior to the internet is misleading. The closest parallel to SVB’s collapse, from a timing perspective, is the Bank of the United States, which failed almost a century ago. Money matters and people will rush to protect their cash when they lose confidence in financial institutions, regardless of whether they’re clicking a few buttons on a website or waiting for hours on the street to enter their local branch.

Figure I: Five of the most well-known bank failures in the last century experienced runs that lasted less than two weeks

• SVB’s collapse was notable for its magnitude, not its speed: In a single day, $42bn (~24% of existing deposits) was withdrawn from SVB. According to regulators, if SVB had not been shuttered the following day, another $100bn of deposits (a staggering 81% of existing deposits) was scheduled to be withdrawn. This far exceeds deposit withdrawals in prior bank runs.

Figure II: While its length was not without historical precedent, SVB’s bank run witnessed an unparalleled magnitude of deposit withdrawals

• It’s too early to assess the impact of digitization on bank run timing: We don’t have enough data points to make a clear-cut assessment of the impact of mobile and internet on bank runs. While 2008 was very different than 2023, it’s worth noting that the WaMu run primarily occurred over the internet during a macro environment experiencing extreme dislocation for over a year. Despite these dynamics, WaMu’s failure took time, and its bank run was noticeably slower than other historical examples. Similarly, we could argue the speed and magnitude of SVB’s failure was not only a function of digitization, but also a function of overconcentration and the size of customer accounts, which makes it hard to isolate variables.

• Three common drivers of bank failures are overconcentration, rapid expansion, and bad investment decisions: It is remarkable how most U.S. failed banks demonstrate one (or more) of these characteristics. Of course, just because a bank demonstrates these characteristics, doesn’t mean it will fail. However, it certainly increases risk profile, which increases susceptibility to runs.

Figure III: The largest failed banks in the U.S. share a common set of characteristics

• Post-facto government assistance can slow a collapse, but often fails to prevent it: In the case of Franklin International and Continental Illinois, federal assistance allowed the banks to continue operations for longer periods of time. Though this assistance may have enabled more customers to relocate deposits, it did little to prevent the eventual failure of these institutions. Instead, regulators should be focused on aiding banks before bank runs happen. Insured deposit protection is one example of an ex-ante regulation that can build customer confidence. Of course, if these regulations are not comprehensive enough (e.g., the majority of SVB deposits exceeded the insured deposit limits put in place by regulators), they can be rendered ineffective.

• Regulators should be worried about the growing role of misinformation and its potential impact on our financial ecosystem: If a single Bronx businessman in 1930 can spread a local rumor and derail a bank, we should be highly concerned about misinformation in the modern era, which can be (and has been) amplified rapidly across social media and the internet. This issue strikes at the heart of so many unsolved, controversial topics today. Should social media platforms take content moderation stances on misinformation that could spark a financial crisis? Will adversarial governments begin to use bank runs as a novel attack vector and geopolitical bargaining chip? Will the next bank run be sparked by an erroneous hallucination from an AI model?

• Despite these concerns, smaller-scale banks play an incredibly important role in our economy: Smaller-scale banks include regional banks, community banks, specialized banks, and minority-owned banks. If their risk profile is elevated, why should anyone deposit capital in them? Well, there are numerous benefits to smaller-scale banks, including more personalized customer attention, flexible lending options, community support and investment, and sometimes lower fees. Community banks represent over 95% of all bank charters, over 40% of branches, and roughly 15% of deposits, loans, and assets. For a significant portion of the population, the pros of a smaller-scale bank outweigh the cons. While I have focused this post on large bank failures, it is usually the smaller-scale banks that fail during a crisis, and that is precisely the time to remind ourselves about the importance of these institutions. Which brings us to a final, comforting point…

• In major financial crises, the government has your back: To end on a positive note, it’s worth pointing out that the government has historically done a good job at protecting deposits when financial institutions fail. Whether it’s through direct bailouts or regulator-assisted asset sales, regulators have prioritized maintaining confidence in the financial system during times of distress. While this confidence boost has not stopped bank runs from happening in the first place, we can confidently say that U.S. banks have been, and will continue to be, trusted holders of capital for consumers and corporations alike.

It has been ten months since the Nasdaq reached its prior peak, seven months since Vladimir Putin invaded Ukraine, and one month since the U.S. formally entered a recession after two quarters of negative GDP growth. Much has already been written about the market downturn and its impact on the startup ecosystem. Founders are taking steps to extend cash runway, and investors have slowed their pace of deployment. In times of uncertainty, the name of the game is survival.

This post takes a step back and attempts to answer the following question:

How fragile is the startup ecosystem exactly?

Part of the answer is rooted in a fundamental misunderstanding that is currently accepted as fact. This misunderstanding relates to dry powder – its definition, size, and implications. By defining a new metric, the Dry Powder Ratio, we can recast a misunderstood term to help us measure the fragility of our industry.

Defining Dry Powder

In venture capital, dry powder refers to the amount of committed, but unallocated, capital available to a firm. These are the funds a VC firm has raised from its limited partners but not yet deployed into startups.

During bull runs, companies are marked at higher valuations, fund returns look increasingly attractive, and limited partners desire more exposure to startups and growth companies. As we can see in Figure I below, these dynamics cause dry powder to explode.

Today, VC funds are sitting on a record haul of $290bn, more than double the amount of dry powder from five years ago and almost three times the ~$100bn of dry powder that existed at the peak of the Dot-Com Bubble. Of course, dry powder does not grow unchecked. As new funds are raised, capital is also deployed, and deal velocity (both size and frequency of deals) follows a similar trajectory.

Dry Powder and Downturns

So what role does dry powder play during a downturn?

Downturns affect different parts of the market differently. In the startup ecosystem, the impact of recession is real, but can also be relatively muted. Sure – valuations decline, fundraising activity shrinks, exit opportunities dry up, and maybe selling a product or service becomes more challenging. Despite these consequences, however, startups also have two things going for them:

1. The cash on their balance sheet

2. The cash sitting in VC funds

The former is runway, and it allows cash flow negative startups to continue operations for a finite period. The latter is dry powder, and it allows startups to tap into additional pools of capital to prevent shutting down. With the extra time, startups can figure out ways to adjust strategy, boost revenues, and improve margins.

This is why dry powder is so important. The more dry powder that exists, the more protection startups have from failing. VCs can continue to back their portfolio companies, or they can invest in companies in other VC portfolios, or they can invest in brand new companies.  The point is…even during a recession, VCs still have money to invest, even when money has become scarce everywhere else.

The Dry Powder Ratio

As an investor, it’s easy to look back at Figure I and conclude, “There’s $290bn of dry powder in the U.S. venture capital market! My portfolio companies should be able to raise money and weather the storm!” In fact, this sentiment is fairly widespread – see here, here, here, here, here, and here.

Unfortunately, nominal dry powder is not the correct metric to measure whether markets are well-capitalized.

The correct metric is the amount of dry powder relative to the amount of capital existing startups need to raise to fund operations. The latter is quite difficult to measure – it’s a function of the cash burn and cash runway of all U.S. startups. As a proxy, however, we can use the total amount of venture investment each year. Let’s call this the Dry Powder Ratio.

Dry Powder Ratio = Dry Powder / Total Annual U.S. VC Investment

The Dry Powder Ratio represents the number of years of capital reserves given the current investment pace. Whereas nominal dry powder demonstrates the total dollars VCs need to deploy, the Dry Powder Ratio estimates how quickly those dollars will be deployed. When the ratio is high, the ecosystem is flush with capital. When the ratio is low, venture funds must continue raising money to support existing and new investments.

Analyzing the Dry Powder Ratio

There is no magic threshold for what the Dry Powder Ratio should or should not be. Instead, the Dry Powder Ratio illustrates where we are today relative to prior periods.

As is evident in Figure III, the amount of dry powder relative to annual deal value has significantly contracted in the last two decades. Looking even further back, the current Dry Powder Ratio of sub-1.0x stand in stark contrast against the 1981 – 2001 average of 2.3x as well as the prior trough at the height of the Dot-Com Bubble.

The decline in the Dry Powder Ratio represents an unprecedented shift in the U.S. startup ecosystem. If the capital needs of startups remain constant, the $290bn surplus of dry powder will not last longer than a year. In other words, U.S. VC funds are more reliant on new capital entering the system than ever before.

Of course, the capital needs of startups do not remain constant. This is why, for example, founders are being advised to cuts costs and extend runway. Furthermore, it’s possible that average startup runway is longer today than during prior downturns, due to the sheer amount of capital that was raised by companies prior to the downturn.

However, while there are a few caveats to the recent trends in the Dry Powder Ratio, those caveats are likely limited in scope. First, extending runway is not something that can be completed overnight. The overwhelming majority of startups are cash flow negative (even in the years preceding and following IPO), so being conservative with cash can only take you so far, especially if your company already has little cash in the bank or your business model has unprofitable or unscalable unit economics. And second, for many companies, costs have scaled with the cash they’ve raised, resulting in negligible impact to runway.

It should also be noted that there is another way in which startups, in aggregate, become less dependent on cash. Companies fail. The math is simple – when there are less companies to invest in, less cash is needed to support those companies. This dynamic played out during the Dot-Com Bubble and was one of the driving factors for the rebound in the Dry Powder Ratio within a few years following the crash.

We do not know how severe the current downturn will be. We do not even know to what extent VC firms will struggle to raise new funds. Nonetheless, the Dry Powder Ratio indicates that we have entered the current downturn with a limited amount of cash, and the notion that existing dry powder will provide a meaningful buffer to startups is misleading at best.

Other Considerations

• The Dry Powder Ratio varies fund-by-fund: Funds have dry powder on an individual basis and can calculate a fund-specific Dry Powder Ratio (a fund’s dry powder divided by the capital needed to sustain their portfolio). By design, most VC funds will not be able to support their entire portfolio without the help of other investors; however, having a high fund-specific Dry Powder Ratio provides greater flexibility to managers and the founders they have chosen to back. Furthermore, founders can estimate the quality of their investors’ dry powder by comparing their cash needs to the amount of undeployed capital in their VC’s funds.

• Excess dry powder carries its own risks: Most VC funds have a lifecycle, typically ten years, in which to invest their capital and then return profits. Excess dry powder, particularly in bull markets, places managers in a difficult position where too much money is chasing too few deals, which drives unreasonable valuations. Ironically, unhealthy excess dry powder can create the precise bear market conditions where dry powder then becomes desirable.

• In extreme downturns, existing dry powder is often overstated due to LP defaults: LP defaults – when limited partners are unable, or unwilling, to meet their capital commitment obligations – are very rare. However, when there are significant shocks to the economy, the rate of LP default increases, and the estimated amount of Dry Powder does not fully materialize.

Conclusion

The Dry Powder Ratio, at a historic low, demonstrates that current dry powder levels may not be enough to support U.S. startups if VC firms struggle to raise new funds.

Of course, there is no substitute for a good business model. Efficiently run companies that sell useful products should continue to raise capital. Companies lacking those characteristics will fail in the long run, regardless of how much dry powder exists in the system. And yet, it takes time to become an efficiently run company with a good product. For some startups, that time is short, and for other startups, that time is lengthy with many bumps along the way. The Dry Powder Ratio is a proxy for the fragility of the startup ecosystem insofar as it is a proxy for time. Not all companies that fail during a downturn would have otherwise. This is why the Dry Powder Ratio matters.