DeepSeek has broken from industry practice by denying Nvidia and AMD pre-release access to its upcoming V4 flagship model. Instead, the Chinese AI lab granted Huawei and other domestic chipmakers a multi-week head start to optimize their processors for the new model.
The move comes as the Trump administration alleges DeepSeek trained V4 on Nvidia’s most advanced Blackwell chips - which are banned from export to China - at a data center in Inner Mongolia.
What’s Happening
According to Reuters, DeepSeek has not provided US chipmakers with pre-release access to V4, which could launch within days. This breaks a long-standing industry convention where AI labs work closely with chip vendors during model development to optimize performance.
Huawei’s Ascend chips are receiving priority attention instead. Chinese domestic suppliers now have several weeks to tune their hardware and software stack before the model goes public.
Ben Bajarin, CEO of Creative Strategies, told Reuters this reflects a broader Chinese government strategy “to try to keep US hardware and models disadvantaged” in the Chinese market.
The Blackwell Allegations
A senior Trump administration official told Reuters that DeepSeek trained V4 on Nvidia’s Blackwell chips, potentially violating US export controls. The official claimed the chips are “clustered at DeepSeek’s data center in Inner Mongolia.”
The administration also expects DeepSeek to “strip out technical indicators” that could reveal American chip usage, and to publicly claim the model was trained on Huawei hardware.
Additionally, the official alleged DeepSeek’s training “likely relied on distillation from leading U.S. AI models,” naming Anthropic, Google, OpenAI, and xAI as sources.
Neither the administration nor Reuters provided details on how DeepSeek allegedly obtained Blackwell chips or how US intelligence gathered this information.
Why This Matters
DeepSeek’s January 2025 release of its open-source reasoning model sent the Nasdaq tumbling 3% and wiped nearly 17% off Nvidia’s stock price. Markets are watching V4 closely for similar disruption.
But the strategic implications run deeper than stock movements. By cutting Nvidia and AMD out of the optimization loop, DeepSeek is accelerating China’s indigenous AI chip ecosystem. Software optimized for Huawei’s Ascend processors makes those chips more competitive - and reduces the leverage US export controls were designed to provide.
The allegations about Blackwell usage, if true, would demonstrate the limits of export controls in practice. Advanced chips find their way to restricted destinations through complex supply chains, and proving sanctions violations after the fact does little to prevent the training runs.
What Nvidia Says
Nvidia declined to comment when contacted by reporters. The company’s stock was trading slightly up in premarket activity following the news.
JPMorgan analyst Harlan Sur noted that changes to export approval processes could “unlock billions in additional revenue” for Nvidia, positioning China as a significant catalyst for the stock.
The Broader Pattern
This isn’t happening in isolation. DeepSeek’s models have been downloaded more than 75 million times on Hugging Face, with Chinese AI releases now surpassing every other country on the platform.
The V4 lockout signals that Chinese AI labs may increasingly treat American hardware vendors as adversaries rather than partners - the exact opposite of what export controls intended to achieve.
The Bottom Line
US export controls aimed to slow China’s AI development by restricting access to advanced chips. DeepSeek’s response - building domestic alternatives while allegedly using banned hardware anyway - suggests the controls may be accelerating Chinese self-sufficiency rather than preventing it.
For Nvidia, being frozen out of DeepSeek’s optimization process is a symbolic loss. The real damage comes if this becomes the new normal across China’s AI ecosystem.