It took OpenAI and Anthropic seven years to raise a combined $5.3 billion before ChatGPT changed the industry. According to new data from Radical Ventures, a new crop of research-first AI startups has raised about $24 billion in just the past two quarters, nearly five times that sum. Most of them have no product, no market and no revenue.
The figures, first reported by the Financial Times' Louise Lucas and widely picked up on Sept. 25 and 26, describe the rise of the "neolab." Radical partners Rich Kotite and Aaron Rosenberg define it as a startup chasing long-term technical breakthroughs, usually founded by scientists who left frontier labs or top universities. Radical's June essay on the category, updated this week, says the wave has raised "roughly 7x" OpenAI and Anthropic's pre-ChatGPT total in the last year alone. It adds that financing is "doubling every two quarters."
A Cap Table Built on Pedigree
The pattern traces back to June 2024, when Ilya Sutskever launched Safe Superintelligence with what he called "one goal and one product." SSI raised $1 billion within four months and was later valued at $32 billion, still without a public roadmap. Thinking Machines, World Labs and Reflection AI followed. Radical says more than 40 neolabs have now raised over $40 billion, often with billion-dollar first rounds. Since the start of 2025, 13 model providers in North America and Europe have raised at least $1 billion in a single round, and eight of them were neolabs.
The recent deals are large. Reflection AI raised $2 billion at an $8 billion valuation to build open-source models. Humans& raised $480 million at a $4.48 billion valuation in January. Flapping Airplanes, which is betting on data-efficient training, raised $180 million at a $1.5 billion valuation from GV, Sequoia, Index and Menlo, and was later reported to be in talks at $5 billion. Jeff Bezos-backed Project Prometheus raised $12 billion at a $41 billion valuation. Jeff Dean's Discovery Loop has reportedly discussed raising $1 billion at around $10 billion.
"It's a bet that this will be very commercially valuable," Flapping Airplanes co-founder Ben Spector said, describing the company's focus on the data efficiency problem. Pete Sonsini, co-founder of Laude Ventures, put the shift more bluntly: "It's not traditional venture capital."
Counting the Money, and the Revenue
Different trackers count neolabs in different ways, and the totals vary. One widely shared tally puts the category at 102 neolabs, $70 billion raised and a $320 billion combined valuation, against less than $1 billion in combined revenue. A September analysis by Leonis Capital is broader: it counts $94.3 billion invested in 119 research labs since 2024, and only 16 of those labs disclose any revenue at all. Leonis gives an example of the mismatch. Mistral reports about $400 million in annual recurring revenue at a $14 billion valuation, while robotics lab Skild discloses $30 million in revenue at the same $14 billion price. Investors quoted in the FT's follow-up coverage warned that headline figures can obscure how much of this capital is real and how much is valuation.
Nvidia connects much of this activity. Radical ranks the chipmaker as the most active strategic investor in model providers, ahead of any institutional venture firm. Leonis finds Nvidia on 34 of the 119 cap tables it tracked. Radical says compute typically accounts for "well over half" of a neolab's spending, which means a large share of each round flows back to the companies selling GPUs.
Why It Matters
The neolab boom tests whether venture capital can price pure research. SaaS companies are revalued on revenue and retention, and biotech on clinical trial results. Neolabs have no such public milestones. Their valuations move on founder pedigree, compute contracts and private evaluations shown selectively to investors. Leonis notes that Thinking Machines reportedly sought a $50 billion valuation in late 2025, about four times its last price, but the round never closed. That gap reflected investor sentiment more than any public change in the research.
The skeptics are specific about the risk. "The vast majority of them will not cross that at all. They will end up with something that is just incrementally better," Foundation Capital general partner Ashu Garg said of the technical gap each lab must close. Radical's own essay describes the bet as "the ultimate power law": most neolabs will not reach the frontier, but one that does could pay for the whole portfolio. Its bear case includes wind-downs, fire sales and "zombie" labs, which becomes more likely as the pool of acquirers shrinks and regulators look harder at acquihire deals. Those deals have already set a floor for failed labs. Inflection, Adept and Character.AI were absorbed by big tech companies, and SAP bought Prior Labs this year.
The compute side is also under pressure. On Sept. 21, Rothschild & Co Redburn started coverage of CoreWeave and Nebius with Sell ratings, citing falling GPU rental prices and rising financing costs. If neolab rounds slow, demand at the neoclouds would slow too.
What to Watch
Watch the next wave of up-rounds, especially whether Thinking Machines, Flapping Airplanes and Discovery Loop close at the valuations reported. Also watch for the first neolabs from the 2024–25 cohort to show public, comparable results. Leonis expects a "re-sorting of capital" rather than a single crash, with venture funds pulling back from decade-long research programs and strategics, sovereign funds and wealthy backers taking their place. The timing of that shift will depend on whether the next rounds are priced on research results or on the pedigree of the founders.
“The vast majority of them will not cross that at all. They will end up with something that is just incrementally better.”— Ashu Garg, General Partner, Foundation Capital