Huawei engineers move into DeepSeek's offices as China builds its own lab-plus-chipmaker answer to Nvidia
What's going on? DeepSeek is working closely with HuaweiDealroom has a profile for this one. Try Dealroom → to shape the design of Huawei's future Ascend AI chips, The Information reported on 19 October, citing two people familiar with the matter. Huawei engineers, some of them stationed at DeepSeek's offices for an extended period, are building the lab's requirements on how models compute and move data between processors into chips due next year and beyond. DeepSeek plans to start pre-training models on the Ascend 950DT as soon as the end of 2026, and Huawei plans to give it most of the supply of its next training chip, the Ascend 960DT, due in the first quarter of 2027.
The bigger picture: China builds its own lab-plus-chipmaker stack. In the US, the big labs have tied themselves to chip partners: Google DeepMind trains on Google's TPUs, Anthropic on TPUs and Amazon's Trainium, and OpenAI is designing custom chips with Broadcom. DeepSeek and Huawei are building the Chinese version, with hardware and software designed together. On 30 September DeepSeek released an Ascend 950 version of TileLang, an open-source programming language led by Peking University researchers that it uses as an alternative to Nvidia's CUDA. In September Bloomberg reported that DeepSeek plans to deploy at least 160,000 Ascend 950DT chips at a data centre in Inner Mongolia, to run models rather than train them. Huawei has pulled the 960DT forward by three quarters and now promises a new Ascend generation every year, with the 970 in 2028 and the 980 in 2029.
Squeezed from both sides: The report landed a day before the US Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products, including third-party products built on its models. Huawei has had no access to TSMC since 2020 and only limited access to high-bandwidth memory since 2024 because of US export controls. DeepSeek, valued at about $75B in its latest funding round, is now tying its training roadmap to a domestic supplier that says it cannot yet meet demand.
What's confirmed and what isn't: The co-design work, the training timetable and the 960DT allocation come from The Information's sources. Neither company has commented publicly, and the timetable is still being tested and could change. Huawei's Ascend roadmap and DeepSeek's TileLang release are public. The gaps are large: research group Epoch AI estimates Nvidia's B300 delivers roughly seven times the raw arithmetic performance of the Ascend 950, a full pre-training run could need more than 10,000 950DTs working together for weeks, and some of the cluster-management software is unfinished. DeepSeek also still wants Nvidia's Blackwell chips, and The Information has previously reported that it used smuggled ones to train its V4 model.
Why it matters: If DeepSeek manages to pre-train a frontier model on Huawei silicon, it would be the first proof that a top Chinese lab can work outside the Nvidia stack. That would give other Chinese labs a blueprint and cut further into Nvidia's China business. It would also turn the US-China AI contest from a fight over access to Nvidia chips into a race between two separate hardware-and-model ecosystems.
Sources: The Information (Qianer Liu, Juro Osawa, 19 Oct 2026); Investing.com (19 Oct 2026); PANews, citing Tencent Tech (20 Oct 2026); Updater News (19 Oct 2026, on TileLang, the Inner Mongolia order and Huawei's roadmap); Federal Register (20 Oct 2026, DHS DeepSeek order)
