How DeepSeek Is Shaping Huawei's AI Chips in a Growing Challenge to Nvidia
Qianer Liu and Juro Osawa report that DeepSeek is working closely with Huawei to shape future AI chip designs (two people familiar; not previously reported). Huawei engineers, some stationed at DeepSeek's offices for an extended period, are building DeepSeek's requests on how models compute and move data between processors into chips due next year and beyond. DeepSeek plans to start pre-training on Huawei's Ascend 950DT as soon as the end of 2026, and Huawei plans to give DeepSeek most of the supply of its next training chip, the Ascend 960DT, due in Q1 2027 (roadmap: 970 in 2028, 980 in 2029). Founder Liang Wenfeng told investors last month that training on more Chinese chips is a major priority. DeepSeek's newer tests use 1,000+ processors, while a full pre-training run could need well over 10,000 950DTs running for weeks, and some cluster-management software is unfinished. DeepSeek also uses TileLang, an open-source CUDA alternative led by Peking University researchers, which now handles most kernel programs for V4 and has an Ascend 950 version. Caveats: the timetable is still being tested and may change; Huawei is constrained by its manufacturing partner's ramp-up and US export controls (no TSMC since 2020, limited HBM since 2024); Epoch AI puts Nvidia's B300 at roughly 7x the Ascend 950's raw arithmetic performance and sees Huawei 3–4 years behind on single-chip performance through 2030; DeepSeek still wants Nvidia Blackwell chips and trained V4 on smuggled ones, as previously reported.
Why it matters
It shows a frontier-lab and chipmaker co-design partnership forming outside the Nvidia stack, the pattern Google (TPUs) and Anthropic (TPUs, Amazon chips) have followed. That matters for how Dealroom tracks AI compute supply chains, China's AI ecosystem and Nvidia's addressable market, and for any management view on where frontier training capacity sits in 2027.
Executive takeaways
- Co-design, not just procurement: Huawei is building DeepSeek's requirements into chips due next year and beyond, with engineers on site at DeepSeek.
- Timeline: pre-training on Ascend 950DT could start as soon as end-2026; DeepSeek is set to get most of the Ascend 960DT supply from Q1 2027.
- Scale gap: tests now use 1,000+ chips, while a full run may need 10,000+ 950DTs for weeks, and cluster software isn't finished.
- Performance gap: Epoch AI sees Nvidia's B300 at ~7x the Ascend 950's raw arithmetic and Huawei 3–4 years behind per chip through 2030; Bernstein's Qingyuan Lin says DeepSeek is choosing to 'go a little slower but go a lot longer'.
- Software moat attack: TileLang now handles most of DeepSeek's V4 kernels and has an Ascend version, weakening CUDA lock-in.
- Supply limits: Huawei can't meet domestic demand, limits overseas sales, and depends on a manufacturing partner still ramping up.
- Hedge, not a switch: DeepSeek still wants Blackwell and trained V4 on smuggled Nvidia chips.
What The Information may be missing
No cost or pricing data for Ascend clusters versus smuggled or legal Nvidia supply, so the economics of the switch are unclear. No view from other Chinese labs (Moonshot, Zhipu, MiniMax, Alibaba, ByteDance) on whether they get similar access or are squeezed out if DeepSeek takes most 960DT supply. Huawei's manufacturing partner isn't named and its yields aren't quantified. It doesn't size what share of DeepSeek's compute runs on Huawei today. Nothing on how US policy might respond to a deeper DeepSeek–Huawei tie-up.
