The Information reports that DeepSeek plans to pre-train models on Huawei's Ascend 950DT as early as late 2026. DeepSeek has not confirmed it.
The workload matters more than the chip. Plenty of Chinese deployments already run inference on domestic accelerators, which is comparatively forgiving - a single request, a single device, software that can route around gaps. Pre-training is the opposite: thousands of chips held in lockstep for weeks, where interconnect bandwidth, collective communication libraries and failure recovery all have to work together. It is the workload where Nvidia's software stack, not its silicon, is the real moat. Moving it is a far stronger claim about Ascend maturity than any inference benchmark.
Context for why this is credible rather than aspirational: DeepSeek shipped V4.1-Flash-1017 on 17 October, a 552B mixture-of-experts model under an MIT licence at $0.30/$1.20 per million tokens at peak. A lab shipping frontier-adjacent open models at that price is not short of engineering capacity. It arrives the same week Taiwan confirmed plans to criminalise AI chip smuggling to China by adding AI servers and advanced chips to its controlled commodities list.