DeepSeek V4 Will Run Entirely on Huawei Chips — China's AI Just Cut the Last American Wire

DeepSeek's upcoming V4 model — a one-trillion-parameter open-source system and the most anticipated AI release of early 2026 — will run entirely on Huawei's Ascend 950PR chips, according to a report from The Information published on April 4.

This is the first time a frontier-class AI model has been trained and deployed without any American silicon.

Chinese tech giants Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of Ascend 950PR units to run DeepSeek V4 through their cloud services and integrate it into their own AI applications. The surge in demand has pushed chip prices up by 20 percent.

The Technical Picture

DeepSeek V4 uses a Mixture-of-Experts (MoE) architecture with approximately one trillion total parameters but only 37 billion active per token — the same activation footprint as the previous V3 model. This means inference costs remain manageable despite the 50 percent increase in total model size.

The model is expected to include native multimodal capabilities — text, image, and video generation — and a one-million-token context window. Internal testing reportedly shows V4 outperforming both Claude and ChatGPT on long-context coding tasks.

Huawei claims the Ascend 950PR delivers roughly 2.8 times the computing power of Nvidia's H20 (the restricted chip still legally sold to China), though it falls short of the H200. Huawei continues to face production bottlenecks caused by US export controls on advanced lithography equipment.

Why It Matters

The geopolitical significance is difficult to overstate. US export controls on advanced AI chips were designed to slow China's AI development by cutting off access to Nvidia's most powerful hardware. DeepSeek V4 running successfully on Huawei chips would demonstrate that those controls have failed in their primary objective.

DeepSeek has been deliberate about this positioning. The company reportedly withheld V4 from US chipmakers Nvidia and AMD for optimisation, instead granting early access exclusively to domestic suppliers Huawei and Cambricon. This is not a workaround — it is a strategic choice to build China's domestic AI hardware ecosystem.

The cost differential adds another dimension. DeepSeek V3 was trained for an estimated $5.2 million — a fraction of the $100 million-plus budgets typical for comparable US models. If V4 achieves frontier performance at similar cost efficiency, it challenges the assumption that AI leadership requires massive capital expenditure on American hardware.

Mubboo's Take

For consumers, the DeepSeek V4 story matters for one practical reason: competition drives down prices. If a Chinese open-source model with free weights delivers performance comparable to proprietary US systems that charge $20-200 per month, it creates downward pressure on what every AI company can charge for consumer-facing products.

The censorship trade-off remains real — DeepSeek's models refuse to engage with topics sensitive to the Chinese government, and data processed through DeepSeek's own app is subject to Chinese law. But for developers building consumer applications on top of open-weight models, V4 could offer dramatically lower costs without those restrictions, since the weights can be run on any infrastructure.

The AI industry is splitting into two supply chains. That split will eventually reach consumer products — and when it does, the question of which AI powers your shopping assistant, travel planner, or financial advisor will carry geopolitical weight that most consumers never see.