A Chinese AI Lab Is Betting Its Future on Chips America Won't Sell It

Keypoints:

• DeepSeek CEO Liang Wenfeng told investors that training on domestic Huawei chips is now a top company priority
• He called it one of DeepSeek's biggest bets, saying it "must succeed"
• Huawei is expected to deliver a new batch of training chips as early as Q4 2026
• An earlier training attempt on Huawei's Ascend 910C reportedly failed, forcing a return to Nvidia hardware
• DeepSeek is training a 2-trillion parameter model and reportedly planning an 8-trillion parameter successor

DeepSeek founder Liang Wenfeng told investors at a closed-door meeting that shifting AI model training onto Huawei and other domestic Chinese chips is now one of the company's biggest strategic priorities, according to reporting from The Information's Juro Osawa and Qianer Liu. 

Huawei is expected to deliver a fresh batch of training chips as early as the fourth quarter of 2026, continuing into early 2027, though the specific models and quantities involved haven't been disclosed.

This isn't DeepSeek's first attempt at the switch. An earlier training run on Huawei's Ascend 910C reportedly failed, forcing the company to revert back to Nvidia hardware, and Liang reportedly estimated that a frontier-scale training run needs roughly 50,000 Nvidia GB300 chips or the equivalent of 200,000 Huawei Ascend 950 processors, a rough measure of how far Chinese domestic chips still trail Nvidia on a per-unit basis. 

Huawei's chip supply is also constrained by shortages of memory and other components elsewhere in the chain, meaning DeepSeek plans to keep using Nvidia hardware in parallel rather than switching over all at once.

The stakes go beyond one company's hardware roadmap. US export controls have pushed Chinese labs to move domestic chips from simple inference work into the far harder job of large-scale model training, and DeepSeek is reportedly working toward an 8-trillion parameter model as part of that push, alongside a nearer-term 2-trillion parameter model likely to become its V4.1 Pro release. 

The company is also said to be raising a new funding round targeting a $74 billion valuation, adding financial weight behind the hardware pivot. Liang reportedly told investors he expects Huawei chips to close the gap with Nvidia within a few years, framing the current transition as preparation, adapting software, training workflows, and infrastructure now, so DeepSeek isn't caught flat-footed if access to Nvidia hardware tightens further down the line.