U.S. official says DeepSeek trained latest AI on banned Nvidia Blackwell chips
A senior Trump administration official claims Chinese startup DeepSeek trained its upcoming model on banned Nvidia Blackwell chips, reportedly clustered in an Inner Mongolia data center. Officials also suspect the firm will scrub technical markers of chip usage. The leak further inflames Washington's bitter debate over export control enforcement.
No matter how tight export bans get, cutting-edge silicon always seems to find a way.
According to a senior Trump administration official cited by Reuters, Chinese AI startup DeepSeek trained its next flagship model on Nvidia’s state-of-the-art Blackwell processors. That comes despite strict U.S. export controls barring shipments of the chip to China.
The upcoming model could drop as soon as next week. U.S. officials believe the banned hardware is running inside DeepSeek’s data center in Inner Mongolia.
Ghost clusters and distilled weights
The report didn't detail how intelligence agencies confirmed the setup, nor how the Hangzhou-based startup acquired the chips in the first place. But officials noted they expect DeepSeek to scrub technical markers to hide any trace of American-made silicon.
Hardware isn't the only borrowed ingredient. The model reportedly relies heavily on distillation techniques drawn from the outputs of leading Western frontier models, including systems from Anthropic, Google, OpenAI, and xAI.
In response, the Chinese embassy in Washington criticized the U.S. for overstretching national security concerns and politicizing trade, reiterating its stance against broad export curbs.
The DC chip rift boils over
The revelation hits right in the middle of a messy debate in Washington over where to draw the hardware line.
Tech figures like Nvidia CEO Jensen Huang and David Sacks have long argued that selling controlled, capable chips to China keeps foreign developers locked into American tech stacks, ultimately slowing domestic competitors like Huawei.
Hawks see it completely differently. Chris McGuire, a former National Security Council official under Joe Biden, told Reuters that the leak proves why sending any AI chips to China is dangerous. He argued that if top startups are already bypassing restrictions, Washington cannot expect compliance on military guardrails. Another former official, Saif Khan, said reliance on smuggled Blackwells exposes weak domestic chip production, warning that selling Nvidia's H200 chips would just hand Beijing a lifeline.
Trump’s stance has ping-ponged for months. In August, he flirted with allowing a nerfed Blackwell variant for Chinese buyers, before reversing course to demand frontier chips stay domestic. In December, the administration approved sales of second-tier H200 chips, though actual shipments remain bogged down in regulatory checks.
Why it matters
For developers and founders, DeepSeek’s upcoming drop will be the benchmark to watch. If the rumors pan out, pairing Blackwell horsepower with aggressive distillation across Google, OpenAI, and Anthropic outputs could yield an exceptionally cheap, high-performing model. The catch? Western enterprise teams touching these weights face growing compliance headaches if U.S. regulators start policing tainted training runs.
For the broader industry, this blows up the hardware deterrence narrative. Tight physical embargoes are failing to stop high-profile labs from pooling frontier clusters. But politically, we think this almost certainly torches Nvidia’s lobbying push to ship H200s to China. Hawks now have a high-profile case study to argue that any opening simply fuels circumvention.
Tactically, DeepSeek is pulling off impressive technical gymnastics under extreme pressure. But building a foundation on smuggled clusters and distilled data is fundamentally fragile. It gives you a burst of performance today, but makes long-term, industrial-scale infrastructure nearly impossible to sustain under the tightening microscope of Washington and rival frontier labs.
Sources (2)
- reuters.com(opens in a new tab)Originalreuters.com
- firstpost.com(opens in a new tab)Reportfirstpost.com