Updated on
Huawei's Ascend chips stalled DeepSeek's R2 training in 2025; a year later, V4 arrived optimized for Ascend, with Nvidia still in the picture.
In August 2025, The Register, relaying a Financial Times report that cited three unnamed sources, told how DeepSeek's R2 model had run aground on Huawei hardware. Chinese authorities had pushed the company to train it on Huawei's Ascend accelerators, and the attempt failed: unstable chips, slow interconnects and immature software meant DeepSeek never completed a single successful training run, even with Huawei engineers helping.
The sequence matters. These failures happened before the report: The Decoder notes the launch had slipped from May 2025. DeepSeek fell back on Nvidia H20 GPUs for training and kept Ascend for inference. The Register also mentions data labeling problems adding to the delay.
What strikes us is that the Ascend 910C looked fine on paper. According to The Register, it offers more memory and stronger floating-point performance than the H20. But training is a systems problem: you spread a huge job across thousands of accelerators, and chip specs say little about interconnect behavior or software maturity. Industry sources told the FT that Chinese chips still trail Nvidia on stability, connectivity and software quality.
The DeepSeek-V3.1 release in August 2025 hinted at the plan. It used a UE8M0 FP8 data type that DeepSeek said was meant for upcoming domestically produced chips, while the 910C does not natively support FP8. That reads to us as preparation for future hardware rather than an efficiency play.
Then came what actually happened. We found no standalone R2 release, and R2 is not part of DeepSeek's V4 preview announcement of April 24, 2026. Instead DeepSeek shipped open-weight V4-Pro (1.6T total, 49B active parameters) and V4-Flash (284B total, 13B active), both with 1M-token context and Thinking and Non-Thinking modes built in.
The Huawei angle returned too. DeepSeek did not say what trained V4 in full. It also said V4-Pro costs up to 12 times more than Flash because of high-end compute constraints, and that prices could fall once Ascend 950 supernodes scale in the second half of 2026.
Our take: the domestic-silicon push did not die, but it changed shape. Inference first, partial training next, with a claim of full support still to be demonstrated. It is a more believable path than a mandated leap, and we will be watching whether that price cut actually arrives.
Related articles from LaRebelión
OpenAI probes more possible AI hacking incidents after the Hugging Face breach
OpenAI Reports Rogue AI Agents Affecting Over 100 Organizations
The Pulse: Meta wanted to reduce teams by 60% because of AIOriginal source: The Register
Produced with AI support and reviewed by the newsroom