Key Takeaways

  • Thinking Machines raised $5 to $6 billion at a $40 billion valuation (down from $50 billion), with Accel leading and Nvidia taking half the round.
  • Poolside was forced into an asset licensing exit after admitting in an internal memo that it simply ran out of capital to compete.
  • Rory O'Driscoll argues that US-based, open-weight foundation models with custom fine-tuning tooling fill an urgent corporate enterprise demand.
  • Jason Lemkin predicts that frontier model jumps like OpenAI's Astra will wipe out the commercial need for tier-two neolabs within 24 months.

The Disagreement

Capital requirements in AI have split the venture world into two camps. On one side stands the belief that specialized, enterprise-focused foundation model builders can carve out defensible positions. On the other stands the reality that training runs cost billions, and anyone without a sovereign balance sheet will get crushed.

Stebbings laid out the sheer scale of the survivors: “Thinking Machines? $40 billion new price. It's down from the $50 billion last year. The round is 5 to 6 billion. Accel leading with Nvidia doing half, couple hundred million bucks in revenue.”

O'Driscoll sees a path forward for select players. Corporate buyers distrust closed ecosystems and want US-based, open-weight foundation models with private fine-tuning layers. But even he acknowledges that the window to fund these efforts has slammed shut. Looking at Poolside's forced asset sale, O'Driscoll explained: “The memo that Jason is referencing is the note they wrote at the time basically saying, we were right, but we couldn't access enough capital to continue to play. Just because that's true doesn't mean the other 100 Neolab bets that you can bet on today will also turn out to be amazing venture bets because now you have other companies with the capital already.”

Lemkin takes a harder line. For him, tier-two labs are living on borrowed time. “It is a reminder that the music will end for a lot of the neolabs, and so be it as it should. It should be a thinning of the herd,” Lemkin said. As frontier labs expand their multimodal capabilities, Lemkin believes the need for separate foundation model providers will evaporate: “The world just changed after Astra, we will be talking about in 24 months. And that was the end of the need for Neolabs.”

Who's Right (and When They're Wrong)

Lemkin is right about raw foundation model research. If your startup's core value proposition is training an alternative base LLM to compete with OpenAI or Anthropic, you are running a capital-raising treadmill you cannot win. When frontier labs drop major capability upgrades, generic secondary models lose their pricing power overnight. Poolside proved that being technically correct about an architecture does not save you when your compute bill outpaces your balance sheet.

O'Driscoll is right about enterprise workflows. Fortune 500 buyers will not send proprietary financial data or trade secrets through closed public APIs without strict isolation guarantees. Companies that combine open-weight weights with sovereign deployment and custom fine-tuning infrastructure will survive. But they must run like enterprise software businesses with high gross margins, not like academic research labs burning billions on unconstrained training runs.

What to Do With This

Audit your model dependencies before you sign another compute contract. If your product relies on a tier-two base model provider, build an abstraction layer this week that allows you to swap model weights within 24 hours. Price your product on workflow utility and proprietary business logic rather than wrapper access to generic model intelligence.