Chinese AI Firm Reportedly Circumvented US Chip Ban
China’s DeepSeek reportedly trained advanced AI models using Nvidia’s top-tier chips, despite a U.S. export ban designed to prevent such access. An official's statement suggests the circumvention highlights the difficulty of enforcing tech controls. The development underscores the persistent challenge for U.S. firms in ensuring supply chain integrity and regulatory compliance amid the ongoing tech rivalry.
- The specific hardware allegedly used was Nvidia's "Blackwell" chip, the company's most advanced and powerful AI processor, which is explicitly barred from shipment to China under current U.S. export controls. - A senior U.S. official stated the restricted Blackwell chips are likely operating in a DeepSeek data center located in Inner Mongolia, and suggested the company may be trying to erase technical data that would reveal the chips' origin. - This incident follows a consistent pattern of escalating restrictions; the U.S. first banned top-tier chips like the A100 and H100 in 2022, leading Nvidia to create compliant, lower-spec versions (A800, H800) for China, which were subsequently also banned in October 2023. - DeepSeek is a relatively new AI startup, founded in July 2023 by Liang Wenfeng, and is fully owned and funded by his Chinese hedge fund, High-Flyer, which previously specialized in using AI for quantitative trading. - Beyond direct acquisition, Chinese firms have also accessed banned high-performance chips by renting computing power through U.S.-based cloud service providers, a method that exploits a loophole in export laws focused on the transfer of physical hardware. - The alleged circumvention occurs as China actively works to reduce its reliance on foreign technology through a "whole-of-nation" approach, including developing domestic alternatives like Huawei's Ascend 910C chip, which is reportedly comparable to some of Nvidia's banned products. - DeepSeek is also reported to have used a technique called "distillation," which transfers knowledge from established Western AI models (from OpenAI, Google, etc.) to accelerate the training of its own models.
