39,000 AI startups, fewer than 12 real exits. Investors are picking sides.
Hg Capital says incumbents win the AI era. Tyler Cowen says the economy shows no productivity payoff. NFX says speed decides. Three arguments, one question: who gets paid?
- Hg Capital's Alan Cline says there have been 39,000 AI-specific companies started, but fewer than 12 exits of real consequence. Self-reported m:…0085
- Tyler Cowen and Alex Imas argue the US economy is not currently seeing productivity increases that would justify the revenue increases of AI companies. Self-reported m:…1459
- Hg Capital says incumbent software companies are best positioned to win in the AI era because they already have customer trust and installed bases. Self-reported m:…0085
- NFX's James Currier says speed is the single most important metric for startup success, rolling up 48 other founder attributes. Self-reported m:…2268
The exit drought
The numbers that frame this week's investor debate come from Hg Capital's Alan Cline. There have been 39,000 AI-specific companies started, Cline says, but fewer than 12 exits of real consequence. The current IPO market, he says, is waiting for Anthropic to debut before broader activity resumes. The imbalance is not new, but the scale of it is: the number of firms doing software buyouts has grown from 5 in 2000 to 450 in 2023, so the capital chasing software is abundant while the exits are scarce. The market, Cline says, is experiencing a K-shaped recovery, where companies with organic growth and AI traction are receiving higher valuations, and the rest are not. m:…0085
That gap between company formation and exits is the question every AI investor has to answer this year. If the exits do not come, the vintages now being raised will be judged on a pipeline of companies that never get sold. If they do come, the first few will set the price for everything behind them. Which is why everyone is watching the same debut. m:…0085
The incumbents' case
Cline's answer to the gap is that the winners are already winning. Incumbent software companies are best positioned to win in the AI era, he says, because they already have customer trust and installed bases. Simple software solutions that only store and represent data are facing a difficult business model, he says, because the value proposition is thin. The human capital efficiency gains enabled by AI can be 5 to 20 times larger than the original software value proposition, which is why incumbents with deep domain expertise are the ones Hg Capital is backing. The firm has over 150 people dedicated to AI across its portfolio, and it is pushing for a hybrid pricing model in software to provide predictability for customers while managing variable compute costs. m:…0085
One supporting number comes from OpenRouter data, Cline says: token consumption on open-weight models has risen to the low 60% range in 2026. The open models are taking share, in other words, but the capture of that value is still an open question. An incumbent with distribution can sell an open model; a startup with only a model cannot. m:…0085
The skeptics' case
Tyler Cowen and Alex Imas, speaking on the Andreessen Horowitz show, make the opposite case about the economy rather than the companies. Progress in AI is accelerating, they say, particularly in the transition from the chatbot era to the agentic era. But rapid technological progress often leads to hitting real-world bottlenecks more quickly, and current AI agents are primarily performing parallel processing rather than true coordination. Institutional inertia in government, non-profits, and healthcare is a major barrier to AI-driven economic transformation, they say, and the parameter for increasing returns to ideas in the current economy is decidedly less than one, suggesting a standard growth world rather than a singularity. m:…1459
Their sharpest claim is about the revenue itself. Extreme revenue projections for companies like Anthropic and OpenAI are likely false, they say, because no other sectors are growing fast enough to justify them. The US economy is not currently seeing evidence of productivity increases in the broader economy that would justify the revenue increases of AI companies. Frontier AI labs are currently engaging in extreme price competition for token usage, which squeezes the very margins those projections depend on. On regulation, they are calmer than the rhetoric suggests: the US political system's checks and balances will likely prevent extreme regulatory outcomes for AI. m:…1459
The founders' answer
NFX's James Currier offers the founder's version of the same argument, and it is about speed rather than valuation. Speed is the single most important metric for startup success, Currier says, rolling up 48 other founder attributes. Founders should set time constraints for projects rather than defining output goals to drive creativity, and founder psychology is the greatest point of leverage in a startup. Google did not invent search, maps, or email, he says, but succeeded through superior scaling and distribution. The context for all of this is a venture industry that has grown enormously: the number of venture capitalists in the US has grown from 150 in 1994 to 35,000 today. Thirty-five thousand investors, 39,000 AI companies, fewer than 12 exits. Speed is Currier's answer to that arithmetic. m:…2268 m:…0085
What follows
The three arguments do not cancel each other out; they describe three different bets. Hg Capital is betting the incumbents capture the value. Cowen and Imas are betting the economy absorbs AI slowly, whatever the labs' revenue says. Currier is betting the founders who move fastest survive the filter. The test for all three is the same, and it is coming: whether the debut everyone is waiting for opens the exit window, or confirms that the window is shut. Until then, the capital keeps forming companies faster than the market can absorb them. m:…0085 m:…1459 m:…2268
What is still unresolved
- Whether Anthropic's debut reopens the IPO window or confirms it is shut.
- Whether open-weight token share keeps rising, and who captures the value if it does.
- Whether agents move from parallel processing to true coordination, which Cowen and Imas treat as the real test.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.