DeepSeek V4 Pro runs on Huawei chips — and undercuts GPT-5.5 by 10x
DeepSeek launched V4 Pro and V4 Flash on April 24, 2026, marking the first time a Chinese frontier AI model has been built to run natively on Huawei's Ascend chips rather than Nvidia hardware. Both models are open-source under an MIT license with a one-million-token context window — and V4 Pro is currently 75% off until May 5, dropping the input price to roughly $0.036 per million tokens.
The hardware story
Previous DeepSeek models, including R1 and V3, were trained on Nvidia H800 GPUs. V4 is the first to formally support Huawei Ascend 950 chips for inference. Getting there wasn't straightforward: MIT Tech Review reports the delay of several months was driven by a deep rewrite of DeepSeek's software stack to squeeze performance from Ascend's different architecture. Experts caution that the training split between Nvidia and Huawei hardware remains unclear — this may be a partial migration rather than a clean break.
The significance is geopolitical as much as technical. US export controls have blocked China's access to Nvidia's most advanced chips since 2022. A frontier model that performs competitively on domestic hardware undermines one of the core assumptions behind that policy: that cutting off chips would slow Chinese AI development.
The price and the performance
V4 Pro is priced at $1.74 per million input tokens and $3.48 per million output tokens at full price — compared to GPT-5.5 at $5 input and $30 output. The promotional rate, valid until May 5, cuts input costs to around $0.036 per million tokens. The lighter V4 Flash model comes in at $0.14/$0.28, per The Next Web.
On world-knowledge benchmarks, V4 Pro leads every open-source model currently available. Among all models — closed and open — it trails only Google Gemini-3.1-Pro. That's a strong result for any model, and a remarkable one for a model optimized around non-Nvidia hardware.
What's available now
V4 Pro and V4 Flash are accessible immediately via the DeepSeek API at chat.deepseek.com, through third-party providers including OpenRouter, and as downloadable weights on Hugging Face. The 75% pricing promotion expires May 5, 2026, at 15:59 UTC. After that, the standard rates still undercut most US competitors by a wide margin.
For cost-sensitive developers and startups, this is a meaningful shift. Whether it represents a durable challenge to Nvidia's ecosystem dominance — or a single well-optimized model — depends on how broadly Huawei Ascend adoption spreads beyond China.
