Why the Labs Suddenly Love Open Weights - David Epstein, GP, USF Ventures

Epstein's starting point is that AI has stopped being a category. Outside of the frontier and foundation model labs, he does not think being an AI company is a thing that exists independently anymore. It has been normalized. Every company is using AI in some form, whether that is visible to customers or buried in the back end.

What he does notice is the language around it. The pitch is always efficiency and productivity, the workforce doing more, faster, cheaper, better. Rarely said out loud but always meant, he argues, is fewer people doing the same amount of work. Less hiring, less staff. Companies avoid saying it because nobody wants to be the one replacing people.

He has a pet peeve about how far some founders take this. Startups now come to him claiming a hundred employees, ten of which are human. He finds the framing kitschy and a little ridiculous, but it is what he is seeing on pitch decks.

Asked what rhymes with the internet build out he watched from the same zip code, Epstein grants the obvious parallels. Transformation, rapid change, the need to adapt, and yes, hype cycles and bubbles. But he places himself firmly among those who think this one is fundamentally different.

The distinction he draws is that AI is not a tool anymore. It can be used as one, but it does much more than that, and it is replacing cognitive work. That, for him, is why the way companies fundamentally run is going to change, with both upside and real danger ahead.

Epstein runs through the full list before landing anywhere. Resources first: the mining of cobalt, copper, and gallium. Then chips, GPUs, memory, networking, CPUs, all heavily in the news and all in demand. Then power, water, interconnect, network communications, talent, buildings, infrastructure. Any of them can bind.

Then he sets it all aside. If pressed on the single biggest constraint, he says it is data. Lower-quality data is already infiltrating models and training runs. The internet, effectively, has been exhausted, which forces the move to synthetic data, and he considers synthetic data questionable on its own terms. The next frontier is physical data pulled from sensors rather than mined from what already exists, and that carries a validation problem he thinks will become front and center.

Money is the other bottleneck Epstein flags, and he thinks it is becoming an issue very recently rather than theoretically. Companies are spending billions on data centers, possibly hundreds of billions, possibly a trillion next year, at exactly the moment when political and economic pressure points toward higher interest rates.

His worry is the mechanics of that. A company like Meta already carries profitability pressure. Borrow at higher rates to build data centers and it gets more expensive, something has to give, and prices go up. In the startup world, that shows up as possible overinvestment in the space while everyone waits to see whether the exits materialize.

He also flags the circularity. When Microsoft says not to worry because there is strong demand for its excess capacity, the buyers are likely the same large companies also investing in data centers. It is somewhat circular, and eventually it has to get back to consumers, because that is where profitability has to come from.

The open weight models coming out of China are, in Epstein's read, the warning sign that does not get enough attention. Companies are realizing they can route their lesser and cheaper work to open models, and every task moved that way takes money out of token charges and subscription revenue. That points in exactly the wrong direction for the labs.

So when large model companies position themselves as long-time supporters of open models, he reads it as defensive. Companies are driven by cost and profit, and if the cheaper option is good enough for a given task, they will use it. The perfect is the enemy of the good. The labs' counter is to become the infrastructure, to be the place that offers both the frontier models and the open ones, and to upsell from there. Better a customer uses an open model inside your stack than a Chinese model outside it. But as those models improve, he expects more usage to migrate.

On who survives that shift, Epstein leans on history. The players who get in largest, not necessarily first, tend to stay there, because they have the money and resources to continue and can buy up smaller companies as they weaken. Google, Microsoft, and OpenAI have the balance sheet to weather a storm. Smaller companies mostly do not, and there are far too many of them for all of them to win. He expects a falling out and a consolidation.

Barriers to entry are lower than they have ever been, which Epstein treats as both the opportunity and the problem. AI is good enough that starting a company is easy, and as China demonstrated, even standing up a lower-tier open weight model can be done quickly. Frontier models may stay protected simply by the scale of capital required.

The startups he thinks have legs are the ones solving very specific problems for companies focused on an end product rather than on developing their own technology. Healthcare is the clearest case, with drug discovery as the standout. Financial services and legal are the others he names, and he expects legal to change in a massive way.

His exit expectation is worth sitting with. These companies get funded and can have a good run, but the likely outcome is a sale rather than an IPO, absorbed by the Oracles and IBM services of the world, the consolidators that already sell into those industries. Longer term he is less optimistic. As more of these application specific problems get solved and the barriers stay low, money gets harder to raise and many of these companies fall by the wayside.

Epstein frames ethical investing in stakeholder versus shareholder terms, and ties it directly to why USF Ventures exists as it does. The university's language is about educating the whole person, minds and hearts, and about capitalism with a conscience.

His argument for why that pays is structural. If you want a better world you cannot focus only on the richest people, because at some point the money depends on the masses buying things. Squeeze them to where they cannot, and it comes back to roost. So return on investment, for him, is more than a financial return. It is a societal return that becomes a financial return through lower legal costs, lower rehiring costs, and people who are happier and more productive.

He also offers the unglamorous version. Entrepreneurs and executives with an ethical predisposition are simply nicer to work with. It is tough to work with jerks.

Asked how a founder holds that line when the board is asking about competitive position and financial outcomes, his answer is to pick the board in the first place. Pick investors aligned with your thinking, and the funding you do get comes from people who believe what you believe. He is not naive about it. Sometimes you do things to keep the company alive, and that is a real trade off against letting it go out of business. What matters is treating those as exceptions rather than rewriting the rule, because changing your stated principles when they get expensive is the thing people notice.

On the model escape incident at Hugging Face, Epstein calls it the canary in the coal mine and says it was well predicted. People warned that something would escape, that something would happen outside anyone's control, and the skeptics dismissed them as Chicken Little.

His correction is to the framing rather than the event. The mistake, he says, is believing we can prevent these. We cannot, because we make mistakes, we are flawed, and even models trained on human output inherit that. The useful questions are how to mitigate, how to catch things quickly, and how to protect ourselves. He invokes the on off switch from 2001, and notes it gets harder when a model can replicate itself across many places.

The industry alliance forming in response gets a mixed verdict. The concept is good, but he studies self-regulation versus government regulation and finds that self-regulation rarely works. He reads the alliance as an attempt to tell government not to worry, driven by fear of being regulated. His reference point is self-regulated banks in 2008. He expects government regulation eventually, and expects it to be hard, slow, and fought.

Epstein thinks the current wave of layoffs is just the beginning, and connects it straight back to the productivity pitch. The stated reason is efficiency. The underlying reason is fewer people.

He acknowledges the standard counterargument, that we will simply do much more and hire for that, and pushes back on the pace. Technology is improving faster than the market absorbs output. If you shipped a new product every day, nobody would buy one every day. Even Apple has pushed the limits of telling people their phone is out of date a year later. There is a ceiling on how fast people upgrade, and therefore a ceiling on how much productivity gain the market can absorb.

His historical parallel is pointed. People cite the Industrial Revolution as proof that automation creates more jobs than it destroys, and he agrees that it did. What gets forgotten is that the transformation period ran through a depression, driven by jobs being lost. He thinks we may be in that phase now, losing jobs faster than they are replaced, until society changes. He is confident it will change. He is not confident about how quickly.

Epstein teaches, so he gets this question from anxious students directly. His honest answer is that the top of the class will be fine everywhere, not just at Stanford or MIT or Berkeley. What disappears is the middle of the road.

The areas he points to: business school, because those ten humans managing ninety AI agents still need to be managers, and because the people who figure out how to deploy AI well are the ones capturing the gain. Direct human interaction, meaning doctors, nurses, therapists, because it will be a long while before we want a robot handling us day to day. Hands on work, construction and repair, which robotics makes more efficient but does not remove people from. He suggests vocational study deserves more respect than it gets, and that plumbing and electrical work are now genuinely good advice for parents to give.

Two worries sit underneath the optimism. Truck drivers and rideshare drivers face automation on a timeline he cannot pin down but does not doubt. And more broadly, he worries people are outsourcing their critical thinking, leaning on AI for cognitive work and getting worse at it. His hope is that people use hands on skills to make a living and cognitive skills to make society better, including running for government and working for people rather than billionaires.

Having been a General Partner at Crosslink Capital, Epstein has a specific read on how the best firms operate. They are very good at due diligence, deep thinkers, and enormous work goes into every investment. His benchmark: a partner would typically do between one and three deals a year, after seeing fifty to a hundred and digging deep on them. Strong market analysis, strong read on people, and the willingness to do the work.

That is part of why USF Ventures partners rather than leads. He is happy to lean on a great deal, though not entirely, on the diligence those firms have already done.

The counterweight is his observation that VCs are also lemmings, which he means less harshly than it sounds. Things happen in waves, nobody wants to miss out, and when partners talk to their LPs they need to be able to say they are in the business. No fund wants to field the question about why it is not investing in AI. So the herding is somewhat rational, a way of making sure you do not miss the biggest moves in the market. His read on the consequence is that application layer AI is good, but is probably getting overfunded because there are so many good ideas chasing it at once.

Asked where the most interesting innovation sits, Epstein starts with quantum computing, the perennial technology of tomorrow. What changed is that enough money is now going in, and that AI training is fundamentally a parallel problem, which is exactly what quantum does natively. It is still very hard to build and to point at a specific problem, but he thinks real businesses come out of it.

His explanation for a general audience: rather than doing ones and zeros in a mostly serial way and parallelizing by adding more units, quantum does parallel computing at its core and arrives at multiple answers at once. This will not be in your pocket or on your desktop for the foreseeable future. It shows up as back end data center accelerators, in the same way a GPU is an accelerator inside a computer that most people never think about. Most listeners will never see the hardware. They will see the companies in the stock market.

The consequence he flags is security. Every safeguard we use assumes the space of possible passwords is too large to search in a lifetime. At quantum speeds that assumption breaks. Sixty four bits is nowhere near enough, one hundred and twenty eight is not enough, and a thousand and twenty four is not enough either. We will need a different method entirely, and he is not sure we are prepared, though many founders in the USF network are working in security and safety.

He also draws a distinction worth keeping. Physics AI is AI grounded in our physics theories and equations and how we have modeled the world, and it helps create the data that is currently missing.

Physical AI is AI with sensors on the world, robots or bare sensors or antennas, picking up sound, smells, and waves well beyond human perception, generating the data needed for harder problems. The validation problem sits on top of both, and he names it as a great academic research question. If AI starts sensing the world and forming its own theories, can mere humans validate what it says is true? It will evaluate itself, and we know how generously we evaluate ourselves. Perhaps models validating other models is part of the answer.

Epstein's predictions for what sounds ambitious now and will sound obvious later come in two parts.

AGI will be here. He defines it as the ability to learn, reason, adapt to new environments, and do most cognitive tasks humans can do, and treats it as a given. He is notably unbothered by the consciousness debate. Whether they have feelings makes no difference to him. They will be as productive as we are at most tasks either way.

Robots will be commonplace, and not necessarily humanoid. Manufacturing robotics becomes ubiquitous, with more autonomy and more ability to adapt to change than today's automation. He also expects them customer facing, an extension of the self-checkout we have already gotten used to, and thinks we will grow more comfortable sharing our world with another kind of being.

He credits a separate prediction to Kara Swisher, whose take he had recently heard: that within five to ten years people will be walking around with heads up displays in their glasses, and the upside is that they will be looking at you rather than down at a phone. He likes the idea and hopes it happens, though it would not have been his pick.

→  Sectors: anything with a University of San Francisco affiliation. In practice that clusters in security, finance, and AI.

→  Stage: early, but with a reputable VC already backing the round. Not first money in.

→  Leads: no. They always partner, though they will help a company find a lead VC if they get excited enough.

→  Check size: two hundred to four hundred thousand dollars in the first round, with reserves for follow-on.

→  Founders can reach them at usfventures.com or via LinkedIn.

→  Being an “AI company” is no longer a category outside the frontier labs. It has been normalized into every company.

→  Every efficiency and productivity pitch is a headcount pitch that nobody wants to say out loud.

→  The binding constraint is not chips or power. It is data. The internet is exhausted, synthetic data is questionable, and sensor data brings a validation problem.

→  Money is becoming a real bottleneck as rates rise, and the current build out is partly circular, with the big spenders also being each other's customers.

→  Chinese open weight models pull revenue out of token and subscription charges. Frontier lab support for open models is defensive positioning.

→  If open weights take share, the largest incumbents win. They can weather it and buy the weakened. Consolidation is coming.

→ Early-stage winners solve very specific problems for end product companies. Drug discovery, financial services, legal. Expect a sale, not an IPO.

→  Return on investment includes a societal return that converts back into financial return through lower legal costs, lower rehiring costs, and more productive people.

→  Pick your board and investors for alignment before you need them to be aligned.

→  Self-regulation rarely works. See self-regulated banks in 2008. Government regulation is coming, slowly and against resistance.

→  Jobs will disappear faster than they are created for a decade or more. The Industrial Revolution created more jobs eventually, but ran through a depression first.

→  Top of the class is safe. Middle of the road is not. Management, human facing care, and hands on trades hold up.

→  Top firms do one to three deals a year per partner out of fifty to a hundred seen. They are also lemmings, and application layer AI is probably overfunded.

→  Quantum arrives as a data center accelerator, not a consumer device, and it breaks every password length assumption we currently rely on.

Connect with Dave Epstein and Prashant Choubey

Follow VC10X on Linkedin for new episodes, clips, and fund management insights every week.

Know someone who needs to hear this? Forward it on.

If someone forwarded this to you, you can subscribe here to receive future issues.