DeepSeek V4 to Run Entirely on Huawei Chips — A Major Win for China's AI Independence

Nvidia's Nightmare Scenario

DeepSeek's V4 model is on the verge of launching — and it will run entirely on homegrown Chinese hardware.

According to a report from The Information, DeepSeek spent months porting its next-gen model to run exclusively on Huawei chips, with chip designer Cambricon also involved. Nvidia, despite being the dominant global AI chip maker, didn't get early access. Only Chinese chip companies did.

The implications are significant. U.S. export controls were supposed to limit China's AI capabilities by cutting off access to advanced Western hardware. If DeepSeek V4 delivers competitive performance on domestic chips alone, those controls may have backfired: instead of slowing China down, they accelerated the development of an independent AI chip ecosystem.

The Hardware: Ascend 950PR

Huawei's Ascend 950PR is the new workhorse. Huawei claims it delivers roughly 2.8 times the computing power of Nvidia's H20 — though it still falls short of the H200.

The production challenge remains real. Huawei continues to face manufacturing bottlenecks tied to U.S. export controls. But the demand signal is loud enough to overcome capacity constraints for now.

The numbers tell the story:

• Alibaba, ByteDance, and Tencent have collectively ordered hundreds of thousands of Ascend 950PR units
• The surge in demand pushed chip prices up 20 percent
• Huawei is projected to control roughly 50% of China's AI chip share by end of 2026

Why This Changes Things

DeepSeek's V3 already shook up the AI world by delivering competitive performance at a fraction of the cost of OpenAI and Anthropic models. V4 running on domestic chips removes the last major dependency — the need for Nvidia hardware.

For the global AI landscape, this matters in three ways:

1. Supply chain independence: Chinese AI companies can now run frontier models without relying on U.S. chip manufacturers. Sanctions become less effective as leverage.
2. Cost pressure: If domestic chips are significantly cheaper than restricted Nvidia alternatives, Chinese AI labs gain a structural cost advantage.
3. Geopolitical signal: The U.S.-China tech decoupling is no longer theoretical. It's being encoded into the infrastructure itself.

Bigger Context

This story sits at the intersection of several major trends already reshaping AI:

• OpenAI is preparing for an $852B valuation IPO while losing key executives
• Google released Gemma 4 under Apache 2.0, making capable models freely available
• Jensen Huang (Nvidia CEO) recently told Lex Fridman "we've achieved AGI" — even as domestic competitors erode his moat

The irony is striking: Nvidia's own CEO declares victory in the AGI race while the company's customers in China rapidly adopt alternatives that don't need Nvidia at all.

What to Watch

DeepSeek V4 reportedly features 1 trillion parameters with pricing around $0.30 per million tokens — competitive with Claude and GPT-5 tier models. If the Huawei chip story holds up under real-world benchmarks, it won't just be a China story. It'll be a global one.

AI hardware independence matters as much as model capability. Whoever controls the silicon controls the ceiling. For the first time, China is signaling that its ceiling might be high enough.

---

Sources: Reuters, The Information, The Decoder