The Anthropic accusation against Alibaba, Moonshot and DeepSeek is not really a story about fraud. It is a story about price. If you want frontier behavior, you have two options: spend nine figures building a cluster, or spend a few million renting one through an API. Distillation is the cheapest capex arbitrage in tech. Anthropic says it banned 24,000 fraudulent accounts tied to scraping campaigns. No exploit. No smuggled GPU. A credit card and a loop pointed at a paid endpoint. That is the whole trick, and it is why compute controls leak by design.
Export controls monitor Nvidia, TSMC and ASML. But capability also moves as tokens. A lab that cannot buy H100s in volume can still rent outputs from a model trained on tens of thousands of them. A $300 million pretraining run can be partly replicated for a few million in API fees. That shifts the enforcement frontier to API terms, identity checks and KYC. Washington now wants AI firms to fingerprint Chinese users and quietly degrade their models. I get the instinct. It poisons the training signal, but it also means paying customers get a worse product and are never told. Contracts, not silicon, are the chokepoint.
If distillation is arbitrage, then export controls are a leaky dam. You can seize a GPU at a port. You cannot seize a token. Proving distillation means inferring intent from traffic patterns, and the labs doing the detecting are the ones losing share. Meanwhile US labs sell the product that erodes their lead. That is structural, not accidental. The fix isn't banning 24,000 accounts after the fact. It is pricing, rate limits, verification, maybe watermarking outputs. Every fix adds friction to legitimate users. The cheapest capex arbitrage in tech stays cheap until renting capability costs more than building it. Compute controls alone will not do that.