DeepSeek blocks Nvidia and AMD from its next AI model, hands Huawei a weeks-long head start

What Happened

DeepSeek has not provided its upcoming V4 AI model to US chipmakers Nvidia and AMD for pre-release hardware optimization, according to Reuters reporting confirmed by multiple sources. Instead, the Chinese AI lab gave Chinese suppliers — most notably Huawei — a head start of several weeks to tune their processors for V4 before the model's release.

Pre-release chipmaker access is a standard part of how AI models reach the market. AI labs typically share models with major hardware companies ahead of launch so those companies can optimize their software drivers and performance libraries. The result is that widely used chips work well with the model from day one. By excluding Nvidia and AMD — whose GPUs power the vast majority of global AI workloads — DeepSeek is making a deliberate choice to prioritize Chinese hardware compatibility over global reach.

A senior Trump administration official separately alleged, via Reuters, that DeepSeek trained V4 using a cluster of Nvidia's Blackwell-generation chips located in Inner Mongolia, and that the company plans to publicly attribute training to Huawei hardware to obscure its reliance on American silicon. The official stated: "We're not shipping Blackwells to China." This allegation originates from a single unnamed source and has not been confirmed by DeepSeek, Nvidia, or any independent investigation. It should be treated as unverified.

Why It Matters

DeepSeek's R1 model rattled Western AI markets in January 2025, demonstrating that a Chinese lab could build a high-performing AI system at a fraction of the cost assumed by US competitors. V4 is the anticipated follow-on flagship. By controlling which hardware gets early access to V4, DeepSeek is now shaping which chips AI developers worldwide will find most reliable and performant at launch.

The stakes for the chip market are concrete. When Nvidia and AMD are shut out of pre-release optimization, their hardware is likely to underperform at launch — or require significant additional work from developers to run the model efficiently. Huawei's chips, which until recently lagged well behind Nvidia's best offerings, will carry weeks of tuning advantage when V4 goes public. Analysts cited by The China Academy describe the move as consistent with a broader strategy to disadvantage US chip suppliers both in China's domestic market and in global AI development.

The unconfirmed export control allegation, if it ultimately proves accurate, would add a more serious dimension. Nvidia's Blackwell chips are restricted from sale to China under US Commerce Department rules. A confirmed violation would suggest those restrictions are being circumvented, and would likely accelerate regulatory scrutiny and new enforcement action. For now, the allegation remains unverified — one unnamed official's claim, not an established fact.

One counterpoint worth noting: the near-term practical impact on Nvidia and AMD may be limited. Most Western enterprises do not run DeepSeek models in production, and V4 is likely to function more as a competitive benchmark than a commercial product for US and European developers.