DeepSeek V4 Is Coming — and It Won't Run on Nvidia Hardware

DeepSeek is expected to release its next major model, V4, before the end of April, and the detail drawing the most attention isn't the model itself. It's the hardware. V4 will run on Huawei's Ascend 950PR chips, not Nvidia's.

That's a bigger deal than it might sound. Nvidia CEO Jensen Huang called it a "horrible outcome" for the United States in a recent podcast appearance. If DeepSeek demonstrates that a frontier model can run on Chinese-made silicon, it weakens one of the more durable advantages the US has held in AI: near-monopoly control of the CUDA software ecosystem.

The CUDA angle

CUDA is Nvidia's software framework for running computations on its GPUs. Over the past two decades, it became the standard foundation for AI development globally. Export controls can limit what hardware China can buy, but CUDA dependence is a separate constraint: even if you have the chips, most AI code is written to run on Nvidia's stack.

DeepSeek has spent months rewriting its core code to work with Huawei's CANN framework, according to TrendForce. If V4 performs well in production, it validates a parallel compute ecosystem. Export controls become harder to enforce when your target has built a working alternative.

What V4 is supposed to be

On the technical side, V4 is a one-trillion-parameter Mixture-of-Experts model, though it activates only around 37 billion parameters during inference. That architecture is designed to achieve frontier-level performance at much lower per-query compute cost. It also targets a 1-million-token context window, putting it alongside GPT-5.4 and Claude Opus 4.7 at the top end of what's currently available.

The model has been delayed at least three times, which suggests the Huawei optimization hasn't been straightforward. Alibaba, ByteDance, and Tencent have been placing bulk orders for Huawei's Ascend chips in anticipation, with prices up roughly 20% in recent weeks, according to Dataconomy.

Why this matters for export controls

The US government has used chip export controls as its primary tool for slowing China's AI development. Those controls have been effective at limiting access to Nvidia's H100 and H200 hardware. But if Huawei's Ascend chips reach production-grade performance with a high-profile model like V4, the policy rationale starts to look less solid. The controls assumed there was no viable alternative.

Huang's framing is, in this context, fairly literal: a working DeepSeek V4 on Huawei hardware would demonstrate that China can develop and deploy frontier AI independently of American technology. He said on the Dwarkesh Podcast that if the development of AI shifted to a different tech stack in China, it would slowly change global standards. Whether V4 delivers on its specifications is still an open question, per Gizmochina. The late-April window is approaching.