DeepSeek V5 Said to Shift Pretraining Onto Huawei Ascend Silicon

DeepSeek plans to pretrain its next flagship model, V5, on Huawei's Ascend chips — a shift from inference duty to the heaviest compute job in AI, according to supply-chain sources cited by STAR Market Daily on 3 November. Neither DeepSeek nor Huawei has confirmed it. The report, read via republications by Huxiu and PConline, gives no chip model, no chip count and no timetable.

The claim matters because it reverses what Bloomberg reported 60 days earlier. On 4 September, Bloomberg wrote that DeepSeek planned to deploy "at least 160,000" Huawei accelerators at a data centre it is building in Inner Mongolia — but that "the startup doesn't currently plan to use the 950DT chips for training, one of the people said." If STAR Market Daily's account holds, the job description for those chips has changed.

What do the chip numbers actually show?

On each vendor's stated specifications, Huawei's Ascend 950DT pairs 144 GB of memory with 4 TB/s of memory bandwidth. Nvidia's H200 carries 141 GB at 4.8 TB/s. Run the arithmetic: the 950DT offers roughly 2% more memory capacity and about 83% of the H200's bandwidth. Neither figure predicts how a full pretraining run on Ascend would actually perform.

Huawei announced those figures at its Connect event in September 2025, saying the 950 series "will provide 1 PFLOPS in FP8 and 2 PFLOPS in FP4" and that the 950DT "will be available in the fourth quarter of 2026." Compute comparisons get murky fast: Nvidia lists 3,958 TFLOPS of FP8 for the H200 SXM with sparsity, while Huawei's figure doesn't say whether sparsity is counted.

The bandwidth gap lands exactly where the AI industry's sharpest constraint sits. Memory bandwidth, not raw FLOPS, increasingly gates large-model training — the "memory wall." The 950DT trails the H200 by 17% on stated bandwidth even while edging it on capacity.

How did DeepSeek and Huawei get here?

The relationship has a documented timeline. In January 2025, according to the report, the two companies deployed more than 3,000 Ascend chips over the Lunar New Year holiday to serve DeepSeek-V3 and R1 from Huawei Cloud's Gui'an data centre. Inference work, in other words. That was the pattern until now — Ascend ran models trained on other hardware.

Then came September. On 30 September, DeepSeek open-sourced infrastructure components for the Ascend platform: the TileLang compiler, compute libraries and the DeepEP communication library, matching what it had already published for Nvidia's platform. Analyst Guillermo Flor noted on X that "Until now, Ascend mostly handled inference for DeepSeek, running models after they were trained on other hardware" — and that "Neither company has confirmed it yet, so V5 itself will be the proof."

Why does this matter commercially?

Three threads run through the story. For Huawei, pretraining would be a new, flagship-grade use case for Ascend, moving it up the value chain from serving models to building them. The Information has separately reported Huawei engineers working on chip design with DeepSeek, with the Ascend 960DT dated to the first quarter of 2027.

For Nvidia, the backdrop is Washington's 15 January 2026 move of the H200 to case-by-case licensing for China — a policy that has already pushed Chinese buyers toward domestic silicon regardless of performance gaps. Chinese clients reportedly expect roughly 1% of $70 billion in Nvidia RTX chip orders to materialize.

For DeepSeek's pricing, the report ties money directly to capacity. According to STAR Market Daily, DeepSeek's API page said after April's V4 launch that V4-Pro throughput was limited by high-end compute, and that the company expected to cut the Pro price sharply once Ascend 950 SuperPoDs ship in volume in the second half of the year. Silicon supply becomes list price.

What remains unverified?

Plenty. The sources are unnamed. A Cailian Press page carrying the story returned a not-found message when checked on 4 November. Some later relays of the report name the Ascend 950 and 2027 — details that cannot be traced to the original. Bloomberg's September piece sits behind a paywall beyond its opening paragraphs.

The verifiable core: a single unconfirmed report, published 3 November, says DeepSeek will do on Ascend what it told Bloomberg in September it wouldn't. Until Huawei ships 950DTs in volume in Q4 2026 and V5 arrives, the strongest evidence either way will be DeepSeek's own pricing page — and whether the promised V4-Pro price cut lands on schedule.