DeepSeek plans to shift a growing share of its AI training to Huawei chips, with the first deliveries expected in the fourth quarter of 2026 or the first quarter of 2027, according to comments CEO Liang Wenfeng made to investors and reported by The Information. The move marks one of the clearest examples yet of a Chinese lab routing around US export controls.
Liang called training on domestic chips one of the company biggest strategic moves, telling investors “we must succeed,” according to two people with direct knowledge of the meeting. DeepSeek has used Chinese chips for inference for some time, but large-scale training has remained on Nvidia hardware.
The chip at the center of the plan is Huawei Ascend 950DT, which has shown stable results in DeepSeek testing. The lab is considering a deployment of at least 160,000 of the processors at a new data center in Ulanqab, Inner Mongolia, a facility expected to draw roughly a gigawatt of power.
That would make it one of the largest known installations of Huawei AI accelerators anywhere. DeepSeek already completed full-parameter post-training of its V4 Pro model on about 1,000 Ascend 910C chips in June, a smaller proving run that informed the larger commitment.
The arithmetic of replacing Nvidia
Liang estimated that training a model comparable to OpenAI largest would require about 50,000 Nvidia GB300 chips or roughly 200,000 Huawei Ascend 950 chips. DeepSeek is investing about 30 billion yuan, roughly $4.2 billion, in expanding its computing capacity.
Huawei has pulled forward the launch of its next-generation Ascend 960DT training chip to the first quarter of 2027, about three quarters ahead of the original schedule, citing unusually high demand from Chinese AI developers. The company has faced production constraints from shortages of advanced memory and components.
DeepSeek has not committed to leaving Nvidia entirely. The lab continues to use Nvidia chips for frontier pre-training, and Liang acknowledged that fully replacing Nvidia with Huawei silicon will be difficult in the near term. Reuters reported Chinese government officials recommended the Ascend integration, framing the shift as partly state-directed.
Export controls and the two-stack market
US export controls on advanced chips, in place since late 2022 and tightened in successive rounds, pushed Chinese labs toward domestic alternatives. Huawei Ascend line filled the gap for inference first, where compute demands are lower, and is now moving into training workloads.
DeepSeek released its V4 family in April, and the V4-Pro model was engineered to run on Ascend 950PR inference hardware. The company validated its training scheme on both Nvidia GPUs and Ascend platforms, per its technical report, without specifying which hardware ran the main pre-training run.
The State Department has separately warned governments about what it calls industrial-scale extraction of US models through distillation, naming DeepSeek among the subjects of those concerns. DeepSeek parent company is raising an outside funding round at a $10 billion valuation.
For Huawei, guaranteed demand from China most prominent open-weight lab gives the kind of market signal that justifies continued capital spending. For Nvidia, each large-scale Ascend deployment removes revenue that would previously have been GPU sales, and China already accounts for a shrinking share of its data center business.
The economics still favor Nvidia on raw throughput per chip. DeepSeek bet is that the gap narrows fast enough, and that software work on Huawei CANN toolchain, which the lab has helped tune, closes the productivity difference for its specific workloads.
Chinese rivals face the same pressures. Alibaba, Moonshot AI and others have optimized recent releases for Ascend-class hardware, and benchmark results published with their model cards put several Chinese open-weight systems within striking distance of US frontier models on standard evaluations.
Whether the Ulanqab deployment proceeds at the reported scale depends on Huawei supply chain. The company has struggled to secure advanced memory in the volumes the chip line requires, and delivery slippage has been a recurring theme across the Ascend program.
If the schedule holds, DeepSeek would begin the transition while training its next generation of models, with the 8-trillion-parameter system it has discussed as a future target the first candidate to lean heavily on domestic silicon.
What Huawei still has to prove
Supply is the binding constraint. Huawei has repeatedly missed volume targets for Ascend production because of shortages in advanced memory and packaging capacity, and analysts who track the Chinese semiconductor supply chain say the 950DT line depends on domestic HBM output that is still ramping. A 160,000-chip order would consume a large share of whatever the line can produce in its first year.
Software is the second hurdle. Nvidia CUDA remains the default for most model developers, and Huawei CANN toolchain, while improved through joint work with DeepSeek, still requires custom optimization for each major model architecture. DeepSeek engineers have published work on adapting attention kernels for Ascend hardware, but the lab keeps that expertise in-house rather than contributing it upstream.
Power is the third. The Ulanqab facility would draw roughly a gigawatt, on the scale of the largest US data center campuses. Inner Mongolia offers cheap coal-backed power and cold winters that cut cooling costs, which is why several Chinese AI campuses cluster there. Grid connections at that scale still take time to build, and Chinese provinces have rationed power for data centers during demand spikes.
The view from Washington
The State Department cable that accompanied the April V4 release warned posts about model distillation, and the export control debate has shifted from stopping chip access to slowing model development. DeepSeek shipping frontier-tier weights on domestic silicon within months of GPT-class releases is exactly the outcome the controls were meant to prevent, and policy analysts on both sides now argue about whether the regime slowed Chinese progress or accelerated self-sufficiency.
Nvidia keeps a limited China presence under export licenses for downgraded chips, but its data center revenue from the country has fallen sharply since 2022. Company executives have said publicly that China revenue is unlikely to return to prior levels regardless of licensing outcomes, because domestic alternatives now exist for the workloads licenses would permit.
For DeepSeek, the calculation is straightforward even if the execution is not. Nvidia access can be cut off by a single rule change, and the lab has watched that risk materialize for peers. Building a working stack on Huawei silicon is insurance with a performance cost, and Liang told investors the company considers the cost worth paying.
The next test arrives with the 950DT deliveries in the fourth quarter. If the chips arrive on schedule and the Ulanqab deployment starts at scale, the two-stack structure of the AI hardware market hardens, with Chinese frontier development running on domestic accelerators and the rest of the industry on Nvidia. If deliveries slip, DeepSeek stays in the uncomfortable middle, training on restricted hardware while the alternative matures.
Additional context: Supercomputing News analysis of Ascend deployments; FourWeekMBA reporting on the investor meeting