DeepSeek plans to train its next model, V5, primarily on Huawei Ascend 950 accelerators rather than NVIDIA chips, with the project expected to be completed in 2027, according to TrendForce, citing Chinese outlets Caiwen News and Star Market Daily. According to Crypto Briefing, a new data center in Ulanqab, Inner Mongolia, is being set up with at least 160,000 Ascend 950DT accelerators for training. There is no official confirmation from either company yet, but the engineering foundation for the transition already exists: in late September 2026, DeepSeek open-sourced its software stack for Ascend.

image
image

What Happened

On November 3, 2026, TrendForce reported on statements from Chinese outlets Caiwen News and Star Market Daily: DeepSeek intends to train model V5 primarily on Huawei Ascend 950 accelerators. The project is expected in 2027, and the model is set to become the first LLM natively trained on this chip. The day before, on November 2, Crypto Briefing reported that DeepSeek would deploy at least 160,000 Ascend 950DT accelerators, featuring 144 GB of HBM memory and a bandwidth of approximately 4 TB/s, in a new data center in Ulanqab. According to these reports, pre-training could begin as early as late 2026, while Huawei's commercial release of the chip is targeted for the fourth quarter of 2026. The main change compared to previous practices is that Ascend chips at DeepSeek have been used primarily for inference (i.e., generating responses) until now, and it is only now that pre-training—the heaviest stage of the full cycle—is planned to be moved to them.

Context

Engineering-wise, this is the riskiest part of the claim: pre-training brings inter-chip communications and the stability of long runs to the forefront, including loss stability and cluster behavior during node failures, whereas inference mode forgives many stack weaknesses. There are no publicly confirmed precedents of frontier-scale pre-training on Ascend to date. The groundwork that makes the plan plausible has already been opened for verification: in late September 2026, DeepSeek published the source code of its software stack for Ascend—the TileLang compiler, computational libraries, and distributed communication libraries—mirroring its NVIDIA components. The relationship between the parties remains pragmatic: Huawei does not hold a stake in DeepSeek, cooperation is limited to the product, computational resources, and technical optimization, and DeepSeek remains the only known major customer for Ascend chips—Moore Threads and other Chinese accelerator manufacturers are not mentioned in reports about the project.

Why This Matters for the Industry

For the AI chip market, this would be the first real transition of a Chinese frontier lab from NVIDIA to domestic hardware specifically at the pre-training stage: Huawei gains a flagship client and validation of the Ascend stack for training, while NVIDIA loses its largest Chinese buyer. A completed project would create a second full-fledged training platform and weaken the industry's dependence on the CUDA ecosystem, while the open-source TileLang and distributed communication libraries would become a ready-made base for porting tools and multi-hardware portability—a niche for migration and tooling for startups is forming around them. The economics of the claim are complex: according to an estimate by Liang Wenfeng, approximately four Ascend 950 chips are equivalent to one NVIDIA GB300, so the efficiency of the stack with a fourfold number of chips is limited by interconnect and parallelism quality, and training costs become sensitive to chip prices, energy costs, and data center area. A signal of Huawei's confidence in production is the accelerated roadmap: Ascend 960DT has been moved up to the first quarter of 2027, three quarters earlier than the original plan; Ascend 960PR is expected in the third quarter of 2027; Ascend 970 in 2028; and Ascend 980 in 2029.

Why This Matters for Users

There is no direct effect for users at the moment: model V5 does not yet exist, pre-training has not started, and the Ascend 950DT has not hit the market, so neither the availability nor the quality of DeepSeek models is changed by this news. Practical significance only emerges if the plan works: DeepSeek will be able to release frontier models without critical dependence on US export restrictions and continue supplying new versions even if they are tightened. Developers have an actionable step available today: the open-source stack—TileLang compiler, computational and communication libraries—can be downloaded, studied, and tested with microbenchmarks to check the insides of the Chinese training stack in practice. The timeline is tight, however: the start of pre-training in late 2026 almost coincides with the commercial release of the chip, so tangible results for readers will not arrive before the release of V5 in 2027—and only if the model quality remains at the frontier level.

What Is Still Unknown / Limitations

The entire story of the pre-training shift remains a plan reported by secondary sources—Caiwen News and Star Market Daily via TrendForce, as well as Crypto Briefing—without direct statements from DeepSeek or Huawei, and the architecture of V5 and the methodology for comparing training quality with the NVIDIA variant have not yet been disclosed. The claimed metric of over 95% of peak hardware performance for the open stack is not accompanied by a description of the workload and metric: it is not specified whether this refers to MFU, HBM bandwidth, synthetic tests, or a real training workload. The plan requires at least 160,000 chips immediately, and actual supply volumes are not yet confirmed by anything—the acceleration of the roadmap reflects the manufacturer's confidence but does not guarantee timelines. The key milestone is the fourth quarter of 2026: the mass production of Ascend 950DT and the start of pre-training in Ulanqab will show whether the plan is being realized, and engineers should watch for the first independent measurements of the stack on real training runs.

Sources

Author

Look at AI, editorial team