DeepSeek V5 may train on Huawei Ascend 950, not Nvidia

Inference on Chinese silicon is routine now. Training is not. That's where Nvidia keeps the lock, and DeepSeek just walked up to it.

DigiTimes reports DeepSeek plans to use Huawei's Ascend 950 as the core training hardware for V5. Not fine-tuning, not serving — training, in 2027. Nothing official yet, and this is one of those leaks that says more than a press release would.

Why the stakes are different this time

Chinese labs have been careful. They run inference on domestic chips and train on whatever they can buy. Inference forgives: an error shows up in milliseconds, you rescan the batch, nothing burns.

Training forgives nothing. Weeks on a cluster, synchronized gradients, interconnect traffic, yields. If something drifts on day three, you find out on day three — and you lose a month of compute and the entire run.

Ascend 950 as primary training hardware is a durability test, not a speed test. And it's a software test: CANN against CUDA, an ecosystem that took fifteen years and an ocean of libraries nobody rewrites in one release.

The second signal in the same batch

Pandaily reports Huawei Cloud is pulling openPangu-2.0-Flash and Pro from MaaS. Sales stop on Nov 6, full retirement on Nov 30, and customers are pointed to DeepSeek-V4.1-Flash-1017 and GLM-5.3.

Two readings, both true at once. One: the stack is fusing — Huawei supplies chips and cloud, DeepSeek supplies the models, the customer gets a pipeline with no Nvidia inside. Two: Huawei's own models lost to someone else's and the company moved its users to what they actually buy. A strategic alliance and an admission of defeat look identical on a billing line.

What's actually being tested

It isn't whether Ascend beats top accelerators on peak numbers. It's whether it survives a full training run at V5 scale without half the cluster throwing errors and half the engineers rewriting kernels.

If it survives, Chinese labs get a second line in the budget and a second supplier at the table. If it doesn't, they go back — quietly. That's why news like this arrives as a leak, not a launch event.

For Nvidia this is not a fire. One provider, one model, a date well over a year out. But supplier switches always start the same way: one brave order, a stretch of silence in the filings, then an industry standard nobody bothers to debate anymore.

Prediction: If Ascend 950 survives a full V5 training run, Chinese labs will stop treating Nvidia as a mandatory budget line by 2027 — and the shift will be felt well beyond Asia.

Disclaimer: supply chain leak, not officially confirmed