Models
DeepSeek is reportedly helping Huawei design its next AI chips: what is sourced and what is not
According to The Information, Huawei engineers work inside DeepSeek's offices to fit future Ascend chips to its models. The report rests on two anonymous sources.
On October 19, 2026, US outlet The Information reported that DeepSeek is working closely with Huawei to shape the design of its future AI chips, in a collaboration that could help China train frontier models without Nvidia technology (The Information, paywalled article by Qianer Liu and Juro Osawa).
Anonymous sources. The report rests on "two people familiar with the matter," according to the article's summary. Neither DeepSeek nor Huawei has commented publicly (Dealroom). We found no official confirmation and no denial.
What The Information reports
According to write-ups by Investing.com and Dealroom, which cite the article:
- Engineers on site: Huawei engineers are stationed at DeepSeek's offices, some for an extended period, building its requests on how models compute and move data between processors into chips due from next year.
- Training: DeepSeek plans to start pre-training models on the Ascend 950DT as soon as the end of 2026. The timeline is still being tested and could change.
- Supply: Huawei plans to give DeepSeek most of the supply of its Ascend 960DT, due in the first quarter of 2027.
- Limits: DeepSeek's latest tests use more than 1,000 processors, while a full pre-training run could need more than 10,000 Ascend 950DTs for weeks; some of the cluster-management software is unfinished. DeepSeek still wants Nvidia chips.
Some relays published diverging versions (950DT launch dates, the direction of the collaboration); we stick to the points above, which are reported consistently.
What is public and verifiable
- Huawei's roadmap: at Huawei Connect on September 17, Huawei announced the Ascend 960DT for the first quarter of 2027, three quarters ahead of plan, the 960PR for the third quarter of 2027, then the Ascend 970 and 980 in 2028 and 2029 (Huawei).
- Software: on September 30, DeepSeek released an Ascend 950 version of TileLang, the open-source language it uses as an alternative to CUDA (Dealroom). Its DeepInfer inference framework, open-sourced on October 20, already runs on Nvidia GPUs and Ascend 950 NPUs (see our article).
- The performance gap: according to Epoch AI, cited by Investing.com, Nvidia's B300 delivers roughly seven times the raw arithmetic performance of an Ascend 950.
Context: US export controls
Huawei has had no access to TSMC since 2020 and only limited access to high-bandwidth memory since 2024 because of US controls (Dealroom). On the Nvidia side, the January 2026 easing only covers the H200, AMD's MI325X and similar chips, whose export licenses to China are now reviewed case by case, under conditions (BIS). On October 20, the US Department of Homeland Security also ordered DeepSeek products, including third-party products built on its models, removed from federal civilian agencies' IT systems (Federal Register).
What it means for people who code with DeepSeek
- No short-term change in prices: the API bills
deepseek-flash(DeepSeek V4.1-Flash-1017) at $0.30 input (cache miss) and $1.20 output per million tokens at peak hours, half price off-peak;deepseek-v4-procosts $1.32 and $3.96 (DeepSeek). See the DeepSeek V4.1-Flash-1017 page and the model comparison. - Less dependence on Nvidia, more on Huawei: if DeepSeek really pre-trains on Ascend, its next models will depend less on access to US chips. But Huawei says it cannot yet meet all demand, according to Dealroom: available compute remains the factor to watch for usage limits and the pace of model releases.
- The weights stay downloadable: V4.1-Flash-1017 is released under the MIT license (Hugging Face), whatever chip it was trained on. A team can host it itself (see our Ollama guide).
- For US public-sector customers, the October 20 order already closes the door to products built on DeepSeek, regardless of this collaboration.
To compare DeepSeek with its rivals on price and coding scores, see also the DeepSeek V4.1 Flash, MiniMax M3.1 Flash, Gemini 4 Argon comparison.
Sources
- The Information: How DeepSeek Is Shaping Huawei's AI Chips in a Growing Challenge to Nvidia (Oct 19, 2026, paywalled, summary only read), verified 2026-10-21
- Investing.com: DeepSeek teams up with Huawei to challenge Nvidia (Oct 19, 2026), verified 2026-10-21
- Dealroom: Huawei engineers move into DeepSeek's offices (Oct 20, 2026), verified 2026-10-21
- Huawei: David Wang's keynote at Huawei Connect 2026 (Sep 17, 2026), verified 2026-10-21
- BIS: revised license review policy for semiconductors exported to China (Jan 13, 2026), verified 2026-10-21
- Federal Register: DHS order on DeepSeek products (Oct 20, 2026), verified 2026-10-21
- DeepSeek: API models and pricing, verified 2026-10-21
Key takeaway: neither company has confirmed the DeepSeek-Huawei co-design; the Ascend roadmap and DeepSeek's API prices, however, are public.
Spotted an error in this article? Report it · Our corrections policy