Key Takeaways
- DeepSeek V4 will launch in the coming weeks running entirely on Huawei chips, marking a major milestone in China’s effort to eliminate dependency on foreign AI hardware.
- Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of units of Huawei’s Ascend 950PR chip, pushing prices up 20%.
- Nvidia was shut out of early testing for DeepSeek V4, with only Chinese chip companies receiving access.
- Huawei claims the Ascend 950PR delivers roughly 2.8 times the computing power of Nvidia’s H20, though it still falls short of the H200.
What Happened
DeepSeek, the Chinese AI lab, is expected to launch its V4 model in the coming weeks, and it will run entirely on Chinese-made chips. According to The Information, DeepSeek spent months working with Huawei and chip designer Cambricon to port the model to domestic hardware.
Notably, Nvidia did not receive early access to V4. Only Chinese chip companies were given access for testing and optimization. This is a deliberate strategic choice that signals China’s AI ecosystem is ready to operate without American semiconductor technology.
Chinese tech giants are already placing massive orders. Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of units of Huawei’s new Ascend 950PR processor, according to five people familiar with the matter. The surge in demand has driven chip prices up by 20%.
Why It Matters
DeepSeek V4 running exclusively on Huawei chips represents the most significant test yet of China’s ability to build frontier AI models without any reliance on American hardware. Previous Chinese AI models, including earlier DeepSeek versions, were trained or optimized on Nvidia GPUs. The V4 launch could prove that US export controls, while slowing China’s AI development, have not stopped it.
The strategic implications extend beyond technology. If DeepSeek V4 performs competitively with Western frontier models while running on domestic chips, it validates China’s multi-year, multi-billion-dollar investment in semiconductor self-sufficiency. It also undermines the US government’s leverage in using chip export controls as a geopolitical tool.
For Nvidia, losing access to the Chinese market for AI training and inference is a material business risk. China has historically been one of Nvidia’s largest revenue sources, and the shift to domestic chips could permanently redirect that spending to Huawei and Cambricon.
Technical Details
Huawei’s Ascend 950PR is the chip at the center of DeepSeek V4’s deployment. According to Huawei, the processor delivers approximately 2.8 times the computing power of Nvidia’s H20, the most advanced chip Nvidia can legally sell in China under current US export controls. However, the Ascend 950PR still falls short of Nvidia’s H200, which is restricted from export to China.
DeepSeek’s engineering effort to port V4 to domestic chips involved months of collaboration with both Huawei and Cambricon. Porting a frontier model to a different chip architecture requires significant optimization work, including rewriting kernels, adjusting memory management, and tuning inference performance for the specific hardware’s strengths and limitations.
Huawei continues to face production bottlenecks caused by US export controls on semiconductor manufacturing equipment. Despite these constraints, the company has scaled production enough to fill the large orders from Alibaba, ByteDance, and Tencent, though the 20% price increase suggests supply is tight relative to demand.
Who’s Affected
The most directly affected party is Nvidia, which faces exclusion from a major model deployment in the world’s second-largest AI market. Chinese cloud providers running DeepSeek V4 through Huawei chips represent lost revenue that Nvidia cannot recover under current export restrictions.
Chinese tech companies benefit from a proven domestic alternative to Nvidia hardware. Alibaba, ByteDance, and Tencent can now build AI infrastructure without the supply chain risk of depending on American chips that could be further restricted at any time.
Global AI researchers and companies outside China are watching closely. If DeepSeek V4 matches or exceeds the performance of Western frontier models, it demonstrates that hardware restrictions alone cannot maintain a technological lead in AI.
What’s Next
The launch of DeepSeek V4 in the coming weeks will provide the first real-world performance data on a frontier model running entirely on Chinese chips. Benchmark comparisons with models from OpenAI, Anthropic, and Google will determine whether the Huawei-based stack can compete at the highest level.
Huawei is likely to ramp up Ascend 950PR production to meet demand, though US export controls on manufacturing equipment remain a bottleneck. The 20% price increase may moderate as production scales, or it may persist if demand continues to outpace supply. The success or failure of DeepSeek V4 on domestic chips will shape China’s AI hardware strategy for years.
