Alibaba and DeepSeek Distill US Models to Close Frontier Gap

For a couple of years the loud argument about US-China AI competition was mostly about compute, and whether export controls on high-end chips would keep Beijing a step behind. The New York Times reports that the more consequential story is quieter and mostly about training data. Chinese labs, notably DeepSeek and Alibaba's Qwen group, have leaned on distillation, collecting question-and-answer pairs from leading US models and using them to train their own, to approximate frontier capability at a fraction of the cost.

The technique itself is not new, but the scale being alleged is. Anthropic told Capitol Hill that Alibaba's Qwen lab ran what it called the largest known distillation attack against Claude, allegedly generating 28.8 million exchanges through more than 25,000 fake accounts between April 22 and June 5, according to Forbes. In February the company had already accused DeepSeek, Moonshot AI and MiniMax of a similar pattern. In April, the White House Office of Science and Technology Policy publicly described the activity as industrial-scale distillation of US frontier models and characterized it as theft of American AI intellectual property.

Why any of this matters commercially: the traffic has already moved. OpenRouter data cited by Forbes shows Chinese-developed models now account for roughly 61% of token consumption among the top 10 models on the platform, up from under 1.2% of weekly token consumption in late 2024. That is Western developers, on a Western routing layer, choosing Chinese models. If cheap distillation is what closed the capability gap, chip export controls address only one leg of the stool.

The honest caveat is that 'distillation' is doing a lot of work in this reporting. Anthropic's account is a company complaint, not a court finding, and the exact contribution of distilled outputs versus open-weight practices versus original research inside DeepSeek and Qwen is not something the coverage can cleanly separate. Take the fake-account counts as reported, not settled. What the reporting does not give you is a clear read on how US policy will police API abuse without breaking legitimate research access, or how model builders detect distillation at inference time without hobbling their own products.

The forward-looking piece is who benefits from here. If a US developer can rent a Qwen or DeepSeek model that behaves closely like a frontier US system at a fraction of the price, cost curves for downstream startups improve regardless of who wins the policy fight. The question worth watching over the next couple of quarters is whether US labs respond with tighter API terms and detection, or concede the middle of the market and defend only the very top.