The New Rule for Picking AI Winners | The a16z Show
Summary
The scale prior changed in November: George says Anthropic and OpenAI already add more monthly revenue than Meta, Google, or Microsoft while AI diffusion across the real economy remains below 5%. Clark says he would not be surprised if the pair reaches a combined $200 billion revenue run rate by year-end—before open source and other vendors—framing this as roughly 10% of the Fortune 500’s collective $2 trillion profit pool. That cost pressure means open-source and local models may matter sooner because “cost is going to hit us in the face.”
Enterprise AI remains an operational redesign story, not merely a software-adoption story. Clark says strong companies currently direct their best people toward new products rather than automating internal work; George characterizes recent layoffs as “trimming of previous fat,” not demonstrated AI efficiency. The most cutting-edge people Clark has spoken with are still documenting workflows and capturing context. Native-AI companies offer the preview: lean teams whisper instructions into “swarms of agents,” with tooling moving from reactive assistance toward proactive engagement.
AI’s power law is steepening even as the identity of the winners becomes less predictable. Clark cites a top-1% exit threshold of $10 billion for 2020-24 and $20 billion after the February update covering 2025 and the first two months of 2026. George says recently closed exits have pushed it to $32 billion, potentially north of $100 billion by September if OpenAI and Anthropic enter the sample. Yet George notes that 40% of last year’s Forbes AI 50 dropped off this year’s list; Clark says the half-life of these companies feels incredibly short.
George’s current rule for application-layer winners is blunt: “you have to be in the token path.” Legacy-software budgets cannot absorb rapidly rising AI costs, while value capture depends on unknowables including frontier-model competition, open source, local inference, and smaller models. Five frontier labs would likely mean cheaper tokens than two, supporting a broader application ecosystem.
The cost-performance battle could overturn today’s frontier concentration. George reports that leading Chinese LLMs appear roughly six months behind US models but 10x cheaper—an “innovator’s dilemma” in which 80% of capability costs 10% of the price. Clark says frontier demand remains voracious, though optimization may arrive sooner than expected; distillation reportedly costs around 2% of pretraining, while like-for-like token costs are falling more than 10x annually.
Today’s low AI loss rates are not evidence that venture risk disappeared. George says their early-stage funds historically have a roughly 60% loss ratio, versus probably single-digit losses recently in AI, and says the “laws of gravity” will reassert themselves. Clark rejects low loss ratios as the objective and calls never losing money “a horrible data point”; George adds, “That’s a PE firm.” The Khosla Ventures philosophy is to back the leader in a promising market, because massive winners—not unusually safe portfolios—drive returns.
Clark is pretty confident AI is not presently a bubble, chiefly because supply remains scarce rather than excessive. Data-center capacity at scale is unavailable until late 2028 or early 2029, and US construction may already be a year behind expectations; the main reversal risk is an algorithmic breakthrough enabling radically smaller models. Against a possible $5 trillion buildout, Clark considers $1-2 trillion of revenue a reasonable return expectation, especially if OpenAI and Anthropic alone approach $200 billion this year.
Deep dive
1. Revenue is arriving long before economy-wide adoption
George says he has changed his mind unusually quickly on both scale and value capture. He says Anthropic and OpenAI now add more revenue per month than Meta, Google, or Microsoft, even though economy-wide diffusion remains “less than 5%”; coding and technology-forward companies are important exceptions.
George’s upper-bound calculation starts with the Fortune 500 or S&P 500, which collectively generate about $2 trillion in annual profit. Clark says he would not be surprised if OpenAI and Anthropic reach a combined $200 billion revenue run rate by year-end, before open source and other vendors—roughly 10% of that profit pool.
The budget must come from somewhere, making local and open-source deployment important sooner than expected: “cost is going to hit us in the face.” George sees pricing increases or labor-force restructuring as more plausible funding sources than continued expansion of traditional software budgets.
George frames the current phase as skeuomorphic: companies mostly use AI to perform existing jobs faster. He calls layoffs “trimming of previous fat,” not proven efficiency. Clark says strong companies direct most resources toward products and new things rather than internal automation, while the most cutting-edge people he has spoken with are still documenting workflows into Markdown and capturing context. Native-AI teams already whisper instructions and run “swarms of agents”; the eventual shift is from reactive tools to proactive engagement.
2. Exit values are exploding while competitive half-lives shrink
Clark cites the top-1% exit threshold at $10 billion for 2020-24 and $20 billion after the February update covering 2025 and the first two months of 2026. George says recently closed exits have pushed it to $32 billion; with OpenAI and Anthropic included, it could be north of $100 billion by September—a potential 10x in roughly 24 months.
Clark says model companies are adding more revenue than the entire public software universe combined. He also says six years of venture-backed IPOs sum to just over $1 trillion, potentially less than any one of three anticipated large IPOs. The emerging giants may collectively exceed the Russell 2000, “if I’m not mistaken”; value creation is not only larger but markedly faster.
George notes that first movers do not necessarily capture category value—Google was not the first search engine, and Facebook was not the first social-media site—and that 40% of companies on last year’s Forbes AI 50 dropped off this year. Clark says the half-life of these companies feels incredibly short: the outcome ceiling is rising just as predicting who captures it becomes harder.
George has swung from “model companies are going to be everything,” to applications owning the opportunity while models become APIs, and now back toward labs expanding into applications for stickiness. Amid that pendulum, his operative test is whether a company sits “in the token path”; anything else faces the shifting technology beneath it and tightening buyer budgets.
3. Model-market structure will decide where the economics settle
George calls frontier market structure a central unknowable. A couple of leading labs would likely sustain higher token prices; five would likely drive them lower, easing pressure on customers and allowing a larger application economy. Clark says the current number is smaller—it is not five—and demand for the best intelligence is highly inelastic.
George relays colleagues’ view that leading LLMs in China are probably six months behind US capability but 10x cheaper. He frames this as an innovator’s dilemma: a system delivering 80% of frontier performance for 10% of the cost, then steadily closing the capability gap.
Clark says he has been surprised by the “voracious” appetite for absolute-frontier models, though optimization may arrive sooner than previously expected. Distillation may cost roughly 2% of a model’s pretraining expense, which would favor open source if it remains possible; meanwhile, like-for-like token costs are falling more than 10x year over year, but frontier consumption is growing even faster in dollar terms.
4. Picking the leader matters more than manufacturing a low loss rate
George compares the current apparent low-loss AI environment with 2021’s emerging-manager market. He says their early-stage funds historically have a roughly 60% loss ratio, versus probably single-digit losses recently in AI, and says the “laws of gravity” will eventually reassert themselves.
Clark rejects a low loss ratio as the objective. He jokes that a VC who has never lost money is presenting “a horrible data point”—“that’s a PE firm,” George adds. The Khosla Ventures philosophy is to back the best founder and market leader wherever talent and technological tailwinds converge; the true failure is choosing the wrong company in a market that succeeds.
George reports a conference poll in which 80% believed AI valuations were too high and about 6% too low. He thinks that distribution may be directionally right: perhaps 80% of companies are overvalued because most will fail, while a small subset is massively undervalued. From the LP perspective, he sees an advantage in diversified exposure to those possible outliers.
Clark says the firm therefore centers its business on early-stage access, then calibrates growth investments through “slugging percentage.” Platform services matter because winners encounter scale problems almost immediately: Cursor is already described as generating billions in revenue while still small and young, forcing early negotiations over suppliers, cloud capacity, pricing, international expansion, and major commercial deals.
5. Scarcity supports the cycle, and public markets need the growth
Clark is “pretty confident” AI is not in a bubble now, though less confident about three years hence. Capacity at scale cannot be secured until late 2028 or early 2029, the US buildout may be a year behind expectations, and shortages extend across data-center components. He expects supply constraints to persist for the next three years.
The principal bubble trigger would be an unexpected algorithmic leap producing massively smaller models; the human brain shows intelligence can learn with far greater efficiency and less context. Even so, Clark considers short-term oversupply unlikely. If roughly $5 trillion of CapEx can support $1-2 trillion of revenue, he considers the equation plausible, especially if OpenAI and Anthropic alone approach a $200 billion year-end revenue run rate.
Clark asks whether public markets can digest the coming IPOs. George thinks hypergrowth listings would be “an excellent thing” after the public-company count halved over roughly 20 years: excluding data-center suppliers, the Mag Seven and software companies grow below 30%, while Palantir is the rare exception at approximately 70%.
George’s five-year VC outlook hinges on labs, open source, and token competition. His admittedly “butchered” platform test is that companies built atop a platform should collectively exceed its value; he expects valuable labs and a large application ecosystem. The biggest outcomes may ultimately come from consumer businesses that redirect time and attention after a decade dominated by incumbent technology companies.