September 12, 2026
Anthropic published a report Thursday accusing Alibaba, Moonshot AI, and DeepSeek of running sustained distillation campaigns against its models, per TechCrunch. Distillation -- training a smaller model to mimic a bigger one's outputs -- isn't new, and it isn't obviously illegal or even against most labs' terms in every jurisdiction. What's notable is the timing and the tone: this lands the same week Moonshot AI said it's targeting $2 billion in annual revenue, with OpenRouter data showing K3 models generating as much as 300 billion tokens a day even as usage has softened slightly. Anthropic isn't just flagging a security concern here -- it's also making the case that Chinese labs are catching up by copying rather than innovating, which conveniently reinforces Anthropic's own pitch that its frontier work deserves a premium. I think both things can be true at once: distillation is a real competitive threat to labs that spend hundreds of millions training from scratch, and it's also a predictable response to how good and how cheap open, distillable models have become. Garry Tan's separate call for U.S. open-weight labs to distill American frontier models too is the tell here -- if distillation is fine when it keeps capability domestic, the objection was never really about the technique. For businesses picking a model provider, the practical takeaway is that price and openness are becoming real differentiators alongside raw benchmark scores, and it's worth understanding where your vendor sits on that spectrum before you build a workflow you can't easily move.
An Anthropic researcher resigned this week and posted on X that the company is racing toward self-improving superintelligence and gambling with lives -- and, per TechCrunch's podcast coverage, the company's own alignment lead co-signed rather than distanced from the message. The industry has flirted with doomer rhetoric before, but a co-sign from someone still inside the building is different from an outside critic's op-ed. I'm genuinely split on how much weight to give this. Safety researchers leaving loudly is sometimes career theater, and sometimes it's the clearest signal a company gives you before something goes wrong. What I don't think business leaders should do is treat it as background noise. If the people closest to frontier model development are uneasy enough to say so publicly, that's worth factoring into any plan that assumes ever-larger, ever-more-autonomous models are an unqualified good for your operations. It's also a useful reminder that the vendors selling you AI capability and the people building it don't always agree on the pace, which is exactly the kind of tension worth watching before you deepen a dependency.
Jensen Huang told TechCrunch Nvidia expects roughly 70% growth next year, insisting the surge isn't the product of circular deals propping up demand. Pair that with Mecka AI closing in on a $500 million valuation for robot training data, and Nscale adding former OpenAI president Fidji Simo to its board ahead of a possible IPO, and you get a clear picture: the money is still flowing hardest into infrastructure and picks-and-shovels plays, not into any single chatbot or app. That's consistent with what we've argued before about deployment, not models, being where AI money is going. For a business owner, the lesson isn't to go buy Nvidia stock -- it's that the tools you'll actually touch, the ones layered on top of this infrastructure boom, are where the real decision-making happens. Meanwhile OpenAI pausing new Pro subscriptions because Astra demand is straining its systems is a small but telling data point: even the best-funded labs are still capacity-constrained, which should temper anyone's assumption that frontier AI access is a solved, always-on utility. If you're building workflows that depend on a single provider's uptime, that's a real risk worth planning around, whether through redundancy or through platforms like Internal Tools & Admin that let you own more of the logic yourself.
Do you see Anthropic's distillation report as a legitimate security concern, a competitive complaint, or both -- and does a researcher's resignation change how you'd evaluate a frontier AI vendor for your own business?
Sources