Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek

There is something almost nostalgic about the latest accusation in the AI wars. Model distillation — using a big model's outputs to train a smaller or alternative model — is one of the oldest tricks in the machine learning book, a practice as old as ChatGPT itself. But the report indicates that the practice has evolved into something far more systemic, weaponized even. Anthropic is engaging in a business model of bandwidth: every question you sell through an API is also a possible donor for a lexicon of someone else's training data. Distillation has an undeniable cold logic. Building a frontier-class model requires years of compute and billions. A highly capable model's responses, meanwhile, contain a dense amount of implicit knowledge, curated by heavyweights, aligned through layers of RLHF. For a lab like DeepSeek or Alibaba, that is an irresistible shortcut: rather than reverse engineering the weights, you can effectively channel the teacher's intelligence through a learner of the same kind. Anthropic's report trying this as not just piracy but the need to cut the cost in the billion-dollar cost of agility, thus making the "must" faster. But the claims in this article (TechCrunch's coverage is available at https://techcrunch.com/2026/09/10/anthropic-detals-distillation-campaigns-from-alibaba-moonshot-ai-and-deepseek/) also highlight a deep fragility in the "frontier labs" business model. A company can lock its weights, watermark its text, and put complex rate limits into place, yet the outputs of the API are so fundamental to the product that you cannot fully close them. Ultimately, the line between normal use and distillation attack will always be a fuzzy line. Policy and geopolitics turn the story on its head. I expect this report to be cited in future debates over export control, AI regulation, and the increasingly messy trust between US and Chinese ecosystems. On the one hand, it's legitimate: these Chinese labs have powerful open-source counterparts; they could argue that they're simply participating in an ecosystem that houses them. On the other hand, the "we'll shampoo, you harvest" dynamic describes a moment where the sunshine ideal of open AI is getting worn. What will actually be built is up, but Anthropic has put a flag in the ground: the output is their asset, and they are drawing the line.