New Report Alleges DeepSeek Illegally Accessed OpenAI Data

Bloomberg: DeepSeek Accused of Stealing OpenAI Data to Train Its AI Models

A new report from Bloomberg claims that Chinese AI company DeepSeek illegally obtained data from OpenAI to train its language models. The report cites findings from a bipartisan committee in the U.S. House of Representatives, which has labeled DeepSeek a serious threat to U.S. national security.

According to the committee, DeepSeek allegedly used unauthorized means to gather OpenAI’s data and even circumvented safeguards by using employees to reduce development costs. OpenAI confirmed portions of the report, stating that DeepSeek used “illegal methods” in developing its models and stole proprietary data from the company.

The committee also raised concerns about DeepSeek’s close ties to the Chinese government. It claims that Liang Wenfeng, DeepSeek’s founder, operates the company as part of an “integrated ecosystem” in partnership with a government-linked hardware distributor and the research institute Zhejiang Lab.

The report further alleges that DeepSeek did not only rely on stolen data but also made massive hardware investments to train its flagship R1 model. While early reports estimated R1’s training costs at around $6 million using 2,048 AI chips, Bloomberg’s sources suggest the true figure is closer to $1.6 billion, including around 50,000 of Nvidia’s high-end Hopper GPUs.

In light of these concerns, the House committee has requested Nvidia to provide details about its chip sales that may have supported DeepSeek’s model training efforts.

The allegations come amid growing U.S.–China tensions over leadership in artificial intelligence. The rise of generative AI has intensified this tech rivalry, prompting the Biden administration to implement strict export controls on advanced chips bound for China. While the White House insists the restrictions are intended to prevent military misuse of AI—not to cripple China’s economy—Beijing has made significant advances regardless.

Notably, DeepSeek’s R1 model reportedly outperformed OpenAI’s o1 reasoning model in several benchmarks—and did so at a fraction of the cost, fueling both intrigue and alarm in the global AI race.