On September 23, 2026, multiple outlets citing The Information reported that DeepSeek founder Liang Wenfeng told a closed-door investor meeting the company is shifting the center of gravity of its model training to domestic chips, planning to adopt Huawei's Ascend series at scale for training. At the same time, DeepSeek is training a 2-trillion-parameter model and has an 8-trillion-parameter version on the roadmap. Liang called "training AI models on domestic chips" one of the company's biggest strategic moves and said it "must succeed."
One caveat up front: this is not an official DeepSeek announcement — it is secondhand reporting from a closed-door meeting. The core information comes from The Information, picked up by several tech outlets, but neither DeepSeek nor Huawei has responded publicly. Read it with "rumor pending confirmation" weighting.
What was said in the closed-door meeting
Across the overlapping reports, four points stand out:
First, training is shifting to domestic chips. DeepSeek previously trained mainly on Nvidia's high-end GPUs, with domestic chips used mostly on the inference side. The new claim is that Huawei Ascend is entering training — the most compute-hungry stage — with the first new chips deliverable as early as Q4 2026.
Second, model scale keeps climbing. The 2-trillion-parameter model in training is widely guessed to be the next flagship, V4.1 Pro; the 8-trillion-parameter version is a longer-term plan. For reference, Alibaba's Qwen 3.8 Max sits at 2.4 trillion parameters and Moonshot's Kimi K3 at 2.8 trillion — 8 trillion would be roughly three times the largest existing open model.
Third, money is flowing into compute. DeepSeek reportedly plans to invest about 30 billion yuan to expand domestic AI compute, and reports say it has placed an order for 160,000 Ascend 950DT chips from Huawei for its 1GW data center in Inner Mongolia.
Fourth, Liang's bet: Huawei chips will catch up with Nvidia on performance within a few years. That is both the source of confidence for the whole move and its biggest uncertainty.
The compute math: a 4-to-1 price
The shift is not free. The reported conversion: training an OpenAI-scale largest model takes roughly 50,000 Nvidia GB300 chips, or at least 200,000 Huawei Ascend 950s — about a 4:1 ratio. Another figure: training the 2-trillion-parameter model needs about 60,000 Ascend chips versus 30,000 Nvidia H100/H200s.
That means the domestic route costs several times the physical footprint, power, and operational complexity for the same training job. DeepSeek is betting on two things: that export controls make Nvidia's top cards ever harder to get, and that Huawei's iteration speed can outrun the gap.
Why now
The timing is no accident. On one hand, US chip export controls keep tightening, Nvidia's sales in China keep hitting snags, and the window for stockpiling cards is narrowing. On the other, Huawei disclosed in April that its chips had already been partially used to train DeepSeek's lightweight V4-Flash model — so this move turns a "trial on the inference side" into strategy.
Coincidentally, just a day earlier, Alibaba unveiled its in-house AI chip, the Zhenwu V900 (Alibaba Launches Zhenwu V900: Triple the Performance, Aiming at 10-Trillion-Parameter Models (/article/2050-alibaba-launches-zhenwu-v900-triple-the-performance-aiming-at-10-trillion-parame)), also targeting training and inference for trillion-parameter models. Domestic AI chips are moving from "backup option" to "main track" — DeepSeek's bet is simply the heaviest one.
How to read it
The symbolic weight outweighs the technical details: this would be the first time a top Chinese model lab has publicly staked training on domestic chips. If DeepSeek can genuinely train a 2-trillion-parameter model on Ascend, it validates the whole domestic compute chain — chips, interconnect, training frameworks, operations, every link has to hold.
But the risks are real: the 4:1 compute cost is tangible, Huawei's supply is still constrained by shortages of advanced memory and components, and Liang's own "must succeed" admits this is a bet with no way back. For Nvidia, losing a flagship customer like DeepSeek is another signal of its AI chip share loosening in China.
For ordinary developers, this won't change API prices in the short term. Longer term, if domestic compute really brings training costs down, open models could iterate faster and open up more. Watch Q4 2026, when Huawei's first deliveries are due — that is the first checkpoint for this gamble.
