China's DeepSeek Unveils V4 Built on Huawei Chips, Challenging Nvidia's AI Dominance

DeepSeek, the Chinese AI startup that shocked the world earlier this year with its cost-efficient R1 model, has unveiled V4 — and this time, the story isn't just about performance. It's about independence.

The new model runs entirely on Huawei's Ascend 910C processors, marking the first time a top-tier Chinese AI model has been trained and deployed without relying on Nvidia GPUs. The achievement represents a significant milestone in China's accelerating push for technological self-sufficiency and raises uncomfortable questions about the long-term effectiveness of US export controls.

DeepSeek V4 demonstrates performance competitive with the latest offerings from OpenAI and Anthropic on standard benchmarks, scoring particularly well on mathematical reasoning, code generation, and multilingual tasks. The model was trained on a cluster of 20,000 Ascend chips — a fraction of the hardware typically required for models of this size.

The efficiency gains come from a combination of architectural innovations. DeepSeek's mixture-of-experts design activates only a subset of the model's parameters for any given task, dramatically reducing computational requirements. The company has also developed proprietary training techniques that squeeze more performance out of each chip.

For the US semiconductor industry, the implications are significant. The export controls implemented over the past three years were designed to slow China's AI development by restricting access to advanced Nvidia GPUs. But DeepSeek's success with domestic hardware suggests that these controls may be accelerating the very outcome they were intended to prevent: the development of a viable Chinese alternative to Nvidia's ecosystem.

Huawei's Ascend chips have improved rapidly. The 910C, while still lagging behind Nvidia's latest H200 in raw performance, offers competitive throughput when paired with the right software stack. DeepSeek's engineers have clearly optimized for the hardware they have, producing results that few in the industry expected so soon.

The White House responded to the announcement by reiterating its commitment to export controls, with a spokesperson stating that "industrial-scale theft of American AI models remains a serious concern." But the statement underscored a growing tension in US policy: between the desire to maintain technological superiority and the reality that restrictions are driving Chinese companies to develop alternatives that may ultimately prove equally capable.

For the global AI industry, DeepSeek V4 signals that the competitive landscape is shifting. The assumption that Nvidia's dominance in training infrastructure was unassailable is no longer a safe bet. And for AI developers worldwide, the availability of a non-Nvidia training pipeline — even a limited one — opens new possibilities for cost reduction and supply chain diversification.

The next chapter of the AI race won't just be about who has the most chips. It will be about who can do the most with what they have.