-
Reuters, WASHINGTON
-
-
Chinese artificial intelligence (AI) start-up DeepSeek’s (深度求索) latest AI model, set to be released as soon as next week, was trained on Nvidia Corp’s most advanced AI chip, the Blackwell, a senior official of US President Donald Trump’s administration said on Monday, in what could represent a violation of US export controls.
The US believes DeepSeek will remove the technical indicators that might reveal its use of American AI chips, the official said, adding that the Blackwells are likely clustered at its data center in Inner Mongolia, an autonomous region of China.
The person declined to say how the US government received the information or how DeepSeek obtained the chips, but emphasized that US policy is :”we’re not shipping Blackwells to China.”
Nvidia Corp and Deepseek logos are pictured in this illustration taken on Jan. 27 last year.
Photo: Reuters
Nvidia declined to comment, while the Chinese Ministry of Commerce and DeepSeek did not respond to requests for comment.
The Chinese embassy in Washington said Beijing opposes “drawing ideological lines, overstretching the concept of national security, expansive use of export controls and politicizing economic, trade, and technological issues.”
Asked about the Reuters report at a regular Chinese Ministry of Foreign Affairs briefing on Tuesday, spokesperson Mao Ning (毛寧) said they were not aware of the circumstances and that China had earlier repeatedly clarified its stance on Washington’s treatment of US chip exports to China.
US government confirmation of DeepSeek obtaining the chips, first reported by Reuters, could further divide Washington policymakers as they struggle to determine where to draw the line on Chinese access to the crown jewels of American AI semiconductor chips.
White House AI Czar David Sacks and Nvidia CEO Jensen Huang (黃仁勳) argue that shipping advanced AI chips to China discourages Chinese competitors like Huawei Technologies Co (華為) from redoubling efforts to catch up with Nvidia’s and Advanced Micro Devices Inc’s technology.
But China hawks fear chips could easily be diverted from commercial uses to help supercharge China’s military and threaten US dominance in AI.
“This shows why exporting any AI chips to China is so dangerous,” said Chris McGuire, who served as a White House National Security Council official under former US president Joe Biden.
“Given China’s leading AI companies are brazenly violating US export controls, we obviously cannot expect that they will comply with US conditions that would prohibit them from using chips to support the Chinese military,” he added.
US export controls, overseen by the US Department of Commerce, currently bar Blackwell shipments to China.
In August last year, Trump opened the door to Nvidia selling a scaled-down version of the Blackwell in China. But he later reversed course, suggesting the firm’s most advanced chips should be reserved for US companies and kept out of China.
Trump’s decision in December last year to allow Chinese firms to buy Nvidia’s second most advanced chips, known as the H200, drew sharp criticism from China hawks, but shipments of the chips remain stalled over guardrails built into the approvals.
“Chinese AI companies’ reliance on smuggled Blackwells underscores their massive shortfall of domestically produced AI chips and why approvals of H200 chips would represent a lifeline,” said Saif Khan, who served as director of technology and national security at the White House’s National Security Council under Biden.
The official declined to comment on how the latest news would impact the Trump administration’s decision on whether to allow DeepSeek to buy H200s.
The model they helped train likely relied on the “distillation” of models made by leading-edge US AI companies, including Anthropic PBC, Google, OpenAI and xAI, echoing allegations made by OpenAI and Anthropic, the official added.
The technique known as distillation involves having an older, more established and powerful AI model evaluate the quality of the answers coming out of a newer model, effectively transferring the older model’s learnings.
