DeepSeek’s Next Model Runs on 160,000 Huawei Chips. It Has Not Said So

DeepSeek’s next flagship is being trained on Huawei silicon. That is the reported plan, not the announced one.

On November 3, two Chinese financial outlets reported that DeepSeek V5 will pretrain on Huawei Ascend 950 chips rather than Nvidia hardware. Neither DeepSeek nor Huawei has confirmed it. Its own changelog names V4.1-Pro instead. There is no V5 announcement anywhere on its channels.

So this story has two halves. They weigh very differently.

The half that is documented

None of this arrives out of nowhere. The groundwork went public in September, and it is checkable.

  • September 30: DeepSeek open-sourced Ascend-optimised kernels, including FlashMLA attention code reported at 410 teraflops during prefill on Ascend 950.
  • Same release: DeepGEMM-Ascend for matrix operations, DeepEP-Ascend for inter-chip communication, and TileKernels, which runs identical Python across Nvidia GPUs and Huawei NPUs.
  • September, to investors: founder Liang Wenfeng described training on Huawei chips as one of the company’s biggest bets, and said it has to work.
  • Ongoing: Huawei engineers work inside DeepSeek’s offices.

Those libraries are not side projects. Attention kernels, matrix routines, and inter-chip communication are the three things you must own before a large training run on unfamiliar hardware is possible. Publishing them signals commitment.

The half that is still reporting

Certainty drops on the chip count and the timeline.

Claim Source Status
Next model pretrained on Ascend 950 Caiwen, STAR Market Daily, Nov 3 Reported, unconfirmed
160,000 Ascend 950DT at Ulanqab Bloomberg, Sept 4 Single source
Training end-2026, release 2027 The Information; Caiwen Unconfirmed

One point needs its own line. DeepSeek’s changelog does not name a V5 at all. It names V4.1-Pro, which contradicts the version label the reports use.

The version label is the weak point. The reports describe a next-generation model. DeepSeek’s own changelog describes a point release.

Both can be true. The reported model may be a separate, larger effort with no public name. That is what the more careful coverage suggests.

The hardware underneath

Ascend 950DT is Huawei’s third-generation DaVinci part, built at SMIC on a 7nm-class process. For training, memory and fabric matter. Raw throughput does not.

  • Memory: 144GB of HBM at roughly 4TB/s.
  • Inter-chip link: 2TB/s.
  • Formats: FP8, FP4, and Huawei’s HiF8.

Those figures put one chip near the memory bandwidth of a current Nvidia accelerator. They do not put it near the same class on power or on software maturity.

The DeepSeek Huawei arrangement involves no equity. Huawei holds none. The relationship is narrower than it sounds. It covers product supply, compute access, and joint optimisation. That is a commercial arrangement, not an alliance.

Why training is the hard part

DeepSeek has run Ascend chips for inference for over a year. Huawei deployed thousands of them to serve V3 and R1 on its cloud. So inference forgives a lot. Training does not.

Serving a model is predictable and parallel. Each request can fail and retry without harming the whole. Training a frontier model requires sustained numerical precision across thousands of chips. It also requires tight synchronisation on every gradient step.

Neither is optional. And it requires recovery when a node dies mid-run.

But the chip specifications matter less than the interconnect. An Atlas 950 SuperPoD links up to 8,192 chips. Whether that fabric holds for weeks under load is the real question. No spec sheet answers it.

Liang’s own estimate puts the gap in plain terms. Roughly four Ascend 950 chips match one Nvidia GB300. Efficiency at that ratio depends entirely on the interconnect and on how well the parallelism holds. That is why the software stack had to be published before the hardware claim made sense.

There is one more fact that cuts against the clean version of this story. When DeepSeek published its Ascend tooling in September, reporting at the time noted the company was still trying to buy more Nvidia chips. But porting software to a second platform is not the same as abandoning the first.

What changes if it works

A frontier-scale model trained end to end on Chinese silicon would be the first of its kind. Every major Chinese lab has built its training stack on Nvidia. Export controls were designed around that dependency.

It would also hand every Chinese lab a second full training platform to port toward. Huawei would get its flagship reference customer. The open-sourced TileLang and communication libraries would become the base layer for that migration, whether or not DeepSeek’s own run succeeds.

Huawei has already pulled the Ascend 960DT forward to the first quarter of 2027, three quarters ahead of the original schedule. Roadmaps move like that when a large customer is waiting.

Treat any parameter count you see for this model as noise. Figures from two trillion to eight trillion have circulated with no source behind them. The chip count has one source. The release window has none. Until DeepSeek publishes a model card, the only confirmed facts are the libraries and the statements.

The interesting question is not whether DeepSeek can afford the chips. It is whether a training run on Ascend survives contact with a cluster failure at week three. Nobody outside the company knows yet. That includes the people writing the reports.