The US-China AI race continues to escalate, with OpenAI recently submitting a memorandum to the US Congress accusing Chinese AI company DeepSeek of improperly extracting the output of top US models through "distillation" technology to train its own R1 chatbot. OpenAI claims this is tantamount to "free-riding," threatening not only the commercial interests of US companies but also potentially impacting national security and technological privacy.
OpenAI: DeepSeek uses "distillation technology" to bypass defenses.
According to a Bloomberg report, OpenAI stated in a memo submitted to the U.S. House Select Committee on the CCP that DeepSeek uses a so-called "distillation" technique to learn from the output of cutting-edge U.S. AI models, thereby enhancing the capabilities of its R1 chatbot.
Distillation is a common training method in the AI field, used to improve performance by having smaller models learn the output of larger models. However, OpenAI states that DeepSeek used a "new and obfuscated approach" to circumvent platform protection mechanisms, including hiding the source through third-party routers, accessing services through unauthorized resellers, and programmatically obtaining large amounts of output from US models.
( Can OpenAI turn ChatGPT into a cash cow before burning through hundreds of billions of dollars? )
Free models are impacting the paid business model of AI in the United States.
In a memo, OpenAI warned that if distillation continues to spread, it could erode the competitive advantage of US AI companies. Companies including OpenAI and Anthropic have invested billions of dollars in recent years to build computing infrastructure, charging for it through subscriptions or enterprise services.
In contrast, DeepSeek and most Chinese models adopt a low-cost or even near-free pricing model. If cutting-edge capabilities can be replicated at a lower cost through distillation technology, it could alter the competitive structure of AI business models and further shrink the market space for American companies.
National security and chip export issues come to the forefront
In addition to commercial considerations, OpenAI also mentioned national security concerns in the memo, including the DeepSeek chatbot's content censorship on topics such as Taiwan and Tiananmen Square, and the potential weakening of security mechanisms during the distillation process, making high-risk areas such as biological or chemical applications more susceptible to abuse.
The U.S. political circles reacted strongly to this, with John Moolenaar, chairman of the aforementioned committee, saying it was a typical tactic of China's "steal, copy, and stifle" strategy.
Furthermore, Nvidia's H800 chips, previously sold to China, have also become a focus of discussion. The US investigation into whether DeepSeek obtained related hardware resources through third parties has intertwined this technological dispute with the issue of semiconductor export controls.
( One year after DeepSeek's meteoric rise, how has the development of AI models in China progressed? )
The AI race is heating up.
Michael McCaul, former head of the House Export Controls Oversight Subcommittee, said, "DeepSeek should serve as a wake-up call, reminding people of the dangers of selling advanced semiconductor chips to the Chinese Communist Party."
China has developed the world's most advanced open-source model using relatively weak Nvidia chips. The thought of what they could do if they had more advanced hardware like the H200 chip sends chills down my spine.
It's clear that this controversy is no longer just a technical dispute between a single company, but a microcosm of competition within the AI industry and geopolitical maneuvering. As generative AI technology accelerates its evolution, the rivalry between China and the US over model capabilities, chip supply, and regulatory frameworks is likely to intensify.
This article, "OpenAI Accuses DeepSeek of 'Hitchhiking'! Stealing US AI Technology to Train Its Own Models," first appeared on ABMedia ABMedia .
