AI Venture Winners Will Be Larger, Faster, and Harder to Identify

David George said the scale assumption that governed much of venture’s thinking about AI changed in November. Before then, he said, AI in the enterprise could still be contextualized as “a nebulous promise” around cloud, software, and productivity enhancement. Consumer AI could be modeled more like a consumer internet business: users, price, and eventual market size.

That framing no longer fits the numbers he is seeing. George said Anthropic and OpenAI are adding more revenue per month than Meta, Google, or Microsoft, even though “actual diffusion of this technology into the real economy is tiny” — less than 5% by his estimate. In coding and tech-forward companies, adoption is much further along. Across most other enterprise functions, he said, companies are “nowhere” near full utilization of the capabilities.

The implication, in George’s view, is that the current revenue curve is not a late-stage saturation story. It is happening before broad enterprise penetration. He said he would not be surprised if OpenAI and Anthropic together reached a $200 billion revenue run rate by the end of the year.

George framed that number against the profit pool of large enterprises. The Fortune 500 or S&P 500, he said, collectively generate roughly $2 trillion of profit per year. If two AI model companies alone are approaching $200 billion of revenue run rate, that is already on the order of 10% of that profit base, before adding open-source usage or other vendors.

That creates a practical question rather than just a valuation question: where will the dollars come from? George said enterprises will have to pay for AI somehow. That pressure is one reason he now expects open source and local deployment to matter sooner than previously assumed. Earlier arguments for local and open-source AI may have been strategic or architectural. In George’s telling, cost will force the issue faster.

The usage pattern he expects is not confined to software engineering. Coding is the clearest current example of what happens when models become good enough and products around them become good enough: usage takes off. George said he is beginning to see similar dynamics in legal work, even though legal is much smaller than coding. He expects takeoff patterns across “a bunch of different functions and organizations and verticals” over the next 12 months.

David Clark used Chris Dixon’s framing of early technology cycles: the first applications are often skeuomorphic, carrying over the forms of the old medium before native applications emerge. In AI, Clark said, most users are still applying the technology to do their existing jobs faster, cheaper, or more efficiently. The more native phase is beginning to appear around agentic AI.

George’s answer was that the biggest enterprise change has barely started: companies are not yet being run differently. He distinguished between two kinds of AI adoption inside companies. The first is product work — building new AI-enabled products or capabilities for customers. The second is internal operating transformation — automating or reorganizing how the company itself runs.

At the best companies, George said, the scarce resources are mostly going to product and new things, not internal automation. The best engineers want to work on product, and management understands that the prize from getting the product right is large. More mature companies may be better suited to reworking internal operations, but they are also slower adopters.

Where internal transformation is happening, George described it as early and somewhat unglamorous. The most cutting-edge people he has spoken with are in a “documentation phase”: turning everything into Markdown files, capturing as much context as possible, and then looking for ways to manage the business without degrading customer experience while driving efficiency.

That is different from the way native AI companies operate. George said the founders of those companies are “built different”: leaner, more aggressive, and working constantly. He contrasted them with the previous SaaS generation, whose companies often had strong business models but were not especially tightly run. In hindsight, he said, investors did not realize how inefficiently many SaaS companies operated until much later.

The difference is visible in the daily work of the most advanced AI companies. George described researchers who are not even typing, but whispering into systems while running swarms of agents. Clark summarized the shift bluntly: “they’re not typing anymore.” George said that kind of workflow is likely part of the future, but still very early.

For George, the skeuomorphic AI phase is defined by reactive software. The native phase will involve proactive engagement, both in consumer and enterprise. He said some early-stage companies are beginning to show that pattern, but the category is still at the beginning.

Clark said one of his prior beliefs has been reinforced: the largest companies in this cycle are likely to be an order of magnitude larger than prior-cycle winners. The surprising part is not just the size of the outcomes, but the speed at which the threshold is moving.

He cited VenCap’s work on top 1% venture exits. Between 2020 and 2024, Clark said, a top 1% exit started at $10 billion. When VenCap updated the numbers in February, the threshold for 2025 and the first two months of 2026 was $20 billion. After another update the day before the discussion, looking only at exits that had closed, the threshold had moved to $32 billion, with Wiz as the cutoff. If OpenAI and Anthropic are included later, Clark said, the threshold could be north of $100 billion by September.

Clark summarized the change as a 10x increase in what a top 1% exit looks like over roughly 24 months. George agreed with the direction and added his own comparison: a16z analyzed all VC-backed IPOs over the previous six years and found that, summed together, they were a little over $1 trillion. George said that total may be smaller than any one of three large IPOs he expects.

The point, for both investors, is that venture outcomes are not merely rising with inflation or public-market multiples. The pace of value creation itself has changed. Clark pointed to companies such as Wiz and Cursor moving from nothing to tens of billions of dollars of value in four to six years. George said the current AI wave is less than four years from the ChatGPT moment, and he believes the companies being formed now will include the generational companies of the next decade.

That belief does not make the investing task easier. It makes it more asymmetric. Outcomes are larger, but concentration among winners is also greater. George said a16z has built its firm around the view that subsequent generations of technology winners will be larger than their predecessors, but the identity of those winners is becoming harder to predict.

Clark introduced the concern through a simple observation: first movers often do not capture the economics of a market. Google was not the first search engine, and Facebook was not the first social media site. In AI, the turnover may be even faster. Clark said Forbes’ AI 50 startup list changed by 40% year over year, with 40% of the prior year’s companies dropping off.

For Clark, that short half-life means the largest outcomes may be much larger while the ability to identify durable winners becomes much harder. George agreed. He said the “shifting sands” under founders are real, and a16z’s priors about where value will be captured have changed repeatedly.

George described three phases in that thinking. Before ChatGPT, after a16z had invested in OpenAI, there were moments when the firm thought model companies would be everything and application companies would disappear. Then the pendulum swung toward the idea that application companies would exist everywhere and model companies would become APIs. Now, George said, model companies are “legging their way up into the application” because applications are their largest path to stickiness.

The rule George emphasized for evaluating companies now is whether they are “in the token path.” That means being part of the flow of AI usage and spending rather than being a prior-generation software product adjacent to it. Buyers are already feeling cost pressure from AI, he said, and they will not keep increasing budgets for older software categories in the same way. Those budgets cannot cover the growth in AI costs. The money may have to come from higher prices businesses can charge or from restructuring labor costs.

The biggest unknown, George said, is the market structure of the model companies. If only a couple of companies remain at the frontier, token prices are likely to be higher. If five companies are at the frontier, prices are likely to be lower. Lower token prices would likely be better for the overall economy because they reduce the pressure to restructure labor as quickly.

George said the current number of frontier competitors is “smaller” than five, and there is significant inelasticity for frontier intelligence: customers want the best models and are willing to pay. But he also left open several unresolved questions. How much work can be done by previous-generation models? What role will open source play? How much can run locally? How much can be done with small models?

Clark added an international comparison from colleagues in China. He said leading Chinese LLMs appear to be about six months behind U.S. models in capability, but 10x cheaper. That raises an innovator’s dilemma: if a next-level-down model can do 80% of what the frontier model can do at 10% of the cost, it may capture meaningful demand over time as capabilities improve.

George said he has been surprised by how voracious demand remains for the absolute frontier. That may be because the market is not yet in the optimization phase. But he now expects optimization to arrive sooner than he would have thought. Per-token costs, like for like, are falling more than 10x year over year, he said, but demand for frontier tokens is exceeding that decline in dollar terms.

He also flagged model distillation as a key open question for open source. The large model companies do not want their models distilled, but George said it may cost on the order of 2% of a model’s pre-training cost to distill it. If that continues to hold and remains possible, it would bode well for open source. If not, the outlook is weaker.

Clark compared today’s AI investing environment to a dynamic he saw in 2021. Then, he said, emerging managers often led seed rounds, established firms marked them up six months later, and the loss ratio appeared close to zero. That is not how venture works. Historically, he said, VenCap’s early-stage funds have had about a 60% loss ratio, meaning 60% of deals do not return the invested capital. In AI over the past couple of years, he said, the loss ratio is not zero but is probably in single-digit percentages — and that is not sustainable.

George agreed with the premise and said a16z does not want to target a low loss ratio at the early stage. If the firm had a low loss ratio, it would mean it was not taking enough risk. He said a venture investor who prides himself on never losing money on a deal is not describing good venture performance; Clark joked that this is a private equity firm.

George described a16z’s early-stage philosophy as inherited from Chris Dixon: in any major space with multiple talented entrepreneurs, strong tailwinds, and a favorable view on the technology, the firm should try to back the best founders and the early market leaders. If the space works and the firm has backed the leader, that is excellent. If the space does not work but the firm backed the leader, that is acceptable and part of the business.

The failure case that matters, George said, is when the space works and the firm picked the wrong company. Those are the decisions a16z scrutinizes most closely.

This matters because AI valuations can be simultaneously too high in aggregate and too low for the eventual winners. Clark cited a UK Venture Capital Association conference survey in which 80% of attendees said AI valuations were too high and about 6% said too low. He thought that balance was roughly right: perhaps 80% of AI companies are overvalued because most will not work, while a small subset may be massively undervalued because they will become the leaders and compound far beyond current prices.

George said that is why a16z’s business must be centered on early stage. The firm needs to own the early investments in the companies that work, knowing many will not. At growth stage, the question becomes not just which companies to back, but how much to invest in a given company in a given situation. “Slugging percentage” is heavily discussed in venture, he said, but for a16z the sizing question is equally critical because of the risk dynamic Clark described.

George also argued that the structure of a venture firm has to change because AI companies are encountering big-company problems much earlier. Cursor, he said, has billions of dollars of revenue while remaining very small and early in its life. Companies like that need help with major business deals, supplier relationships, cloud deals, pricing, salesforce scaling, international expansion, and channel development far sooner than previous technology companies did. George said a16z’s platform expansion reflects entrepreneur demand for that kind of support.

Clark framed the bubble question around supply. Classic bubbles, he said, tend to involve excess supply that destroys economics. AI currently looks different: there is not enough compute, memory, data-center capacity, or power. The market is supply-constrained, not demand-constrained.

George said that condition makes him “pretty confident” AI is not in a bubble right now, though he was less confident about three years from now. At present, he said, large-scale data-center capacity cannot be obtained until late 2028 or early 2029. He also believes the U.S. is probably a year behind where people would expect on data-center buildout.

The constraints extend across the data-center supply chain. George mentioned TSMC showing restraint and trying to be balanced, along with other hardware components that are hard to manufacture and ramp. He also criticized local resistance to data centers, arguing that the best operators are offering communities nature preserves, high-speed internet for schools, attractive facilities, jobs, and tax revenue, while still facing objections about issues such as water use.

In George’s view, the more likely near-term scenario is continued supply constraint rather than bubble-like oversupply. The main thing that could change that would be a major algorithmic breakthrough producing massively smaller or more efficient models. He said some companies are working on that kind of efficiency. The human brain is far more efficient than current models at learning and using context, so he expects some shift over time toward less token-consumptive systems. But he does not expect a sudden enough breakthrough to create oversupply in the short term.

George also tested the AI infrastructure buildout against possible revenue. If the industry spends $5 trillion of capex over four or five years, he said, getting $1 trillion to $2 trillion of revenue as a return may be reasonable. If OpenAI and Anthropic alone reach $200 billion of revenue run rate by year-end, he argued, investors should feel comfortable with that equation over the next few years.

The caveat is supply. George said supply, not current demand, is what would drive a bubble. Right now, he believes the system is too constrained for that to be the present condition.

Clark asked what would happen if SpaceX, OpenAI, and Anthropic entered public markets and together represented $4 trillion to $5 trillion of value. Would public markets have enough capacity to absorb them, and would later companies face indigestion after those IPOs?

George’s answer was that bringing those companies public while they are still in hypergrowth would be “an excellent thing” for the investor community. He wants broader public ownership, including through index funds. He gave the example of his parents’ retirement funds being in index accounts and said he would like them to be able to directly own SpaceX, OpenAI, and Anthropic through index inclusion.

He connected that to a longer-running decline in the number of public companies, which he said has shrunk by half over roughly 20 years. Adding high-growth companies to the public market would be a “shot in the arm,” especially because, outside the data-center supply chain, there are few fast-growing public companies available to buy. George said the Magnificent Seven are all growing below 30%, as are public software companies. Clark pointed to Palantir as an exception, and George agreed it is the main public company growing at roughly that faster pace.

George expects some ownership shifting as investors make room for these companies, but he believes the market can bear it. The reason is the same one underlying the rest of his argument: the absolute values may be large, but these companies may continue hypergrowth for many years. He compared the possibility to how investors now look back at the rise of the largest technology companies and accept that $4 trillion or $5 trillion market capitalizations became possible.

Asked what venture capital looks like in five years if the optimistic AI case is right, George returned to the market structure of the model industry. The role of open source, the number of competitive labs, and the cost of tokens will determine how much value is created above the model layer.

He invoked a Bill Gates quote, paraphrasing it as the idea that for a platform to be valuable, the companies built on top of it need to exceed the value of the platform itself. If that is the future of AI, George said, he is very optimistic: the labs can become extraordinarily valuable, and a massive ecosystem of companies built on tokens, AI, and intelligence can also become very valuable.

That is the venture opportunity he wants a16z positioned for: seeing and backing the best early-stage companies, then continuing to support the same founders through later rounds. He said the health of the business is measured by whether the firm is seeing and doing the best early-stage companies and following on successfully.

George also argued that much of the discussion had focused on B2B, while some of the biggest outcomes may come from consumer. Over the past decade before AI, he said, consumer time spent was captured by large technology companies, making competition against them extremely hard. He is optimistic that AI-driven technology shifts will change consumer attention and time spent, creating major new outcomes.

Clark, reflecting on 34 years investing in venture funds, described the moment as the most exciting and scary period of his career. The pace of change creates opportunity, but also raises the cost of being wrong. George ended on the more optimistic side: he believes AI will make the way people live and work “a lot better,” and that a great deal of value will be created as those patterns change.