2026년 2월 14일 · Unknown · financial · 출처 Yahoo Finance
This article first appeared on GuruFocus.
OpenAI has cautioned US lawmakers that its Chinese rival DeepSeek may be using increasingly sophisticated distillation tactics to extract outputs from leading American AI systems, potentially intensifying both competitive and national security concerns. In a memo sent Thursday to the House Select Committee on China and reviewed by Bloomberg, OpenAI said DeepSeek relied on distillation techniques as part of what it described as ongoing efforts to free-ride on capabilities developed by OpenAI and other US frontier labs. The company said it detected new, obfuscated methods designed to evade its safeguards, and it previously opened a probe with partner Microsoft Corp. (NASDAQ:MSFT) shortly after the R1 model's release last year to assess whether data had been obtained in an unauthorized manner.
Warning! GuruFocus has detected 2 Warning Sign with MSFT. Is MSFT fairly valued? Test your thesis with our free DCF calculator.
OpenAI's internal review suggests that accounts associated with DeepSeek employees may have attempted to bypass existing guardrails by accessing models through third-party routers to mask their source. The memo also said DeepSeek employees developed code to access US AI models and obtain outputs in programmatic ways, while pointing to networks of unauthorized resellers of OpenAI's services that could have been used to evade controls. Distillation, in which one AI model trains on the outputs of another, has persisted despite enforcement efforts, according to the company. Given that DeepSeek and many other Chinese models do not charge a monthly subscription fee, the spread of such techniques could pose a business risk to US firms that have invested billions in AI infrastructure and monetize through premium services, potentially narrowing what Washington views as a US advantage in artificial intelligence.
The warning comes amid broader scrutiny of semiconductor policy and export controls. US authorities previously opened a probe into whether DeepSeek circumvented export restrictions by purchasing chips via Singapore, and records obtained by the House China committee show that Nvidia Corp. (NASDAQ:NVDA) provided technical support to help improve and co-design the R1 model. The DeepSeek-V3 base model required 2.8 million H800 GPU hours for full training, and those processors were allowed to be sold to China for a few months in 2023 before a later rule halted sales. At the end of last year, President Donald Trump moved to ease certain chip restraints and allow Nvidia to sell its H200 processors, which are about 18 months behind the leading Blackwell versions, a policy shift that some lawmakers have framed as potentially accelerating China's AI progress.
View Comments