DeepSeek plans to pretrain its next large model, DeepSeek V5, on Huawei’s Ascend chips, STAR Market Daily reported on 3 November, citing several people in the supply chain. Until now, according to the report, Ascend chips mainly ran inference for DeepSeek’s models. Neither DeepSeek nor Huawei has confirmed the plan in any source we read.
Huawei’s own figures for its Ascend 950DT put its memory bandwidth at 4 TB/s, about 83% of the 4.8TB/s Nvidia lists for its H200, by our arithmetic. The report as relayed does not say which Ascend chip V5 would use.
Business Pill · THE MEMORY WALL
A short explainer on the memory wall, why memory capacity and bandwidth limit computing. It teaches the general idea only and says nothing about DeepSeek, Huawei or this report.
The key insight: STAR Market Daily reported that DeepSeek plans to pretrain V5 on Huawei’s Ascend chips, citing supply-chain sources; on Huawei’s and Nvidia’s stated figures, the Ascend 950DT has about 83% of the H200’s memory bandwidth and about 2% more memory, by our arithmetic.
STAR Market Daily wrote that its reporters learned from multiple industry-chain sources that DeepSeek V5 plans to be pretrained on Ascend chips, according to the text as republished by Huxiu. The report gives no date for the training run and no chip count. A Cailian Press (cls.cn) page that carried the story returned a not-found message when we checked on 4 November.
The report traces the relationship back to January 2025, when, it says, Huawei and DeepSeek 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. It also says DeepSeek announced on 30 September that it was open-sourcing infrastructure components for Huawei’s Ascend platform, including the TileLang compiler, compute libraries and the DeepEP communication library, matching the components it had published for Nvidia’s platform.
According to the report, DeepSeek’s API page said after the April launch of V4 that V4-Pro throughput was limited by high-end compute, and that DeepSeek expected to cut the Pro price sharply once Ascend 950 SuperPoDs ship in volume in the second half of the year.
Per-chip memory capacity and bandwidth as each maker states them: Huawei’s Ascend 950DT (144 GB, 4 TB/s) against Nvidia’s H200 (141GB, 4.8TB/s). The percentages are our arithmetic; compute is not compared because the two makers state FP8 on different bases. STAR Market Daily’s report that DeepSeek V5 will pretrain on Ascend chips is unconfirmed and names no chip model.
What Was Posted on X
BREAKING: DeepSeek plans to pretrain its next big model, DeepSeek V5, on Huawei Ascend chips, according to STAR Market Daily citing multiple supply chain sources.
Until now, Ascend mostly handled inference for DeepSeek, running models after they were trained on other hardware.
Inference was the easy test. Training is the real one.
If a top Chinese lab can pretrain a frontier model without Nvidia, export controls stop being a ceiling and start being Huawei’s best sales pitch. Neither company has confirmed it yet, so V5 itself will be the proof.
Do you think a model trained on Huawei silicon can keep up with the US frontier labs
Guillermo Flor’s post at 13:16Z on 3 November relayed STAR Market Daily’s report, noted that neither company had confirmed it and asked whether a model trained on Huawei silicon can keep up with the US frontier labs; it had about 1,300 views by 17:00Z on 4 November.
What Bloomberg Reported in September
On 4 September Bloomberg reported that DeepSeek planned to deploy “at least 160,000” of Huawei’s top accelerators at a data centre it is building in Inner Mongolia. It said DeepSeek “intends to use Huawei’s next-generation Ascend 950DT chips for operating its models”.
The same report said: “The startup doesn’t currently plan to use the 950DT chips for training, one of the people said”. By our count, STAR Market Daily’s report came 60 days later. The two reports do not name the same chip: Bloomberg wrote about the 950DT, while STAR Market Daily, as republished, refers to Ascend chips without a model number.
The Chip on Huawei’s Own Figures
At Huawei Connect in September 2025, Huawei said the Ascend 950DT would pair with its own high-bandwidth memory, “delivering 144 GB of memory and 4 TB/s memory access bandwidth”, and that it “will be available in the fourth quarter of 2026”. It said the Ascend 950 series “will provide 1 PFLOPS in FP8 and 2 PFLOPS in FP4”.
Nvidia’s product page lists the H200 at 141GB of memory and 4.8TB/s of memory bandwidth. By our arithmetic, the 950DT’s stated memory is about 2% more than the H200’s and its stated bandwidth about 83% of it. We do not compare compute: Nvidia lists 3,958 TFLOPS of FP8 for the H200 SXM with sparsity, and Huawei’s figure does not say whether sparsity is counted.
The H200 is the Nvidia chip that Washington moved to a case-by-case licence policy for China on 15 January 2026, as our Nvidia and China timeline records.
What Was Posted on X
Guillermo Flor’s account (@guilleflorvs) relayed the report on X on 3 November, writing 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”.
As we read the two reports, the change is in the job the chips are reported to do: Bloomberg wrote in September that DeepSeek did not then plan to train on the 950DT, and STAR Market Daily now reports a plan to pretrain V5 on Ascend chips.
As we read the stated specifications, the 950DT matches the H200 on memory capacity and trails it on memory bandwidth; neither figure says how a training run on Ascend would perform.
Bloomberg, 4 Sep 2026
“The startup doesn’t currently plan to use the 950DT chips for training, one of the people said”
Three Implications
FOR HUAWEI STAR Market Daily reports that Ascend chips mainly handled DeepSeek’s inference until now; a pretraining run would be a new use, if the report is borne out.
FOR NVIDIA The H200 is the chip Washington moved to case-by-case licensing for China on 15 January 2026; its stated bandwidth is 4.8TB/s against the 950DT’s 4 TB/s.
FOR DEEPSEEK’S PRICING According to STAR Market Daily, DeepSeek’s API page tied a sharp cut in the V4-Pro price to volume shipments of Ascend 950 SuperPoDs.
The framework says: “Memory is becoming the new compute.”
As we read it, the stated figures put the Ascend 950DT level with the H200 on memory capacity and behind it on bandwidth, which is where the framework places the constraint. The framework predates this report and does not discuss DeepSeek or Huawei.
DeepSeek and Huawei have not confirmed the plan in any source we read, and STAR Market Daily’s sources are unnamed. We read the report as republished by Huxiu and PConline; the cls.cn page was not available to us. The report as relayed gives no chip model, chip count or timetable for V5; some later relays name the Ascend 950 and 2027, which we could not trace to the original.
Bloomberg’s September report is paywalled beyond its opening paragraphs. The chip comparison uses each maker’s stated per-chip figures, not measured performance, and says nothing about how a training run would perform. We make no forecast about V5.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
STAR Market Daily reported on 3 November that DeepSeek plans to pretrain V5 on Huawei’s Ascend chips, 60 days after Bloomberg reported that DeepSeek did not then plan to train on the Ascend 950DT. On Huawei’s and Nvidia’s stated figures, the 950DT carries about 2% more memory than the H200 and about 83% of its bandwidth, by our arithmetic.
A note on sourcing. The DeepSeek V5 report is STAR Market Daily’s, read as republished by Huxiu and PConline on 3 November 2026; it cites unnamed supply-chain sources and has not been confirmed by DeepSeek or Huawei. Bloomberg’s 4 September report was read in its free preview. Chip figures are Huawei’s and Nvidia’s stated specifications; the ~83%, ~2% and 60-day figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
