DeepSeek's next-generation AI model, V4, will run entirely on Huawei's Ascend chips — a strategic decision that marks the most significant milestone yet in China's push for AI hardware self-sufficiency. The model will not use Nvidia or any other US-made semiconductor. [1]

The decision was reported by The Information and confirmed by Reuters. DeepSeek, the Chinese AI lab that stunned the industry with its cost-efficient V3 model last year, is now building its successor on domestic hardware. [2]

The Chip

Huawei's Ascend 950 is the chip in question. It is widely described as launching in 2026, with availability expected in the fourth quarter. The chip represents Huawei's most advanced AI accelerator and the culmination of years of development under US export controls that have cut the company off from leading-edge semiconductor manufacturing. [3]

DeepSeek's decision to build V4 on the Ascend 950 is not just a technical choice. It is a political one. It signals that China's leading AI lab believes domestic hardware is ready for frontier model training — a claim that US chipmakers have disputed. [4]

The Strategic Context

US export controls have restricted China's access to advanced AI chips since 2022. Nvidia's most powerful GPUs cannot be sold to Chinese customers. The controls were designed to slow China's AI development by denying it the hardware necessary to train large models. [5]

DeepSeek's V4 on Huawei chips is the first major test of whether those controls are working. If V4 performs competitively on Ascend hardware, the controls have failed to prevent China from developing frontier AI capabilities. They have only forced China to build its own supply chain. [6]

The Orders

Chinese tech giants have already placed large orders for Huawei's Ascend chips. Alibaba and Tencent are among the customers, according to reports. The demand is driven by the same export controls that pushed DeepSeek toward Huawei — if you cannot buy Nvidia, you buy what is available. [7]

The question is whether the Ascend 950 can deliver the performance necessary to train a model that competes with GPT-5, Claude and other frontier systems. DeepSeek's track record suggests it can. The company's V3 model achieved competitive results at a fraction of the cost of Western alternatives. V4 on Ascend would extend that advantage to the hardware layer. [8]

The Implications

A successful V4 launch on Huawei chips would reshape the global AI landscape. It would prove that China can train frontier models without US hardware. It would accelerate the bifurcation of the AI industry into two parallel ecosystems — one built on American silicon, one built on Chinese. [9]

The bifurcation is already underway. DeepSeek's decision makes it visible. [10]

-- DAVID CHEN, Beijing