DeepSeek and Huawei Co-Design AI Chips: Claimed vs Verified

A new report says DeepSeek is shaping the design of Huawei's next AI chips and plans to begin using the Ascend 950DT for pre-training as soon as the end of 2026. The source is The Information, as summarized on Techmeme, which cites people familiar with the matter. That is a bigger claim than the one Bloomberg made seven weeks ago, and the two accounts do not line up on the most important detail: whether DeepSeek will train on Huawei silicon at all.

This post lays out who reported what, what is confirmed, what is still unverified, and why supply, not design, is the gating factor. If you follow China's AI stack, DeepSeek's model cadence, or Nvidia's China exposure, the distinction matters.

TL;DR: claimed vs verified

What The Information reported

The Techmeme headline summarizes the story in one sentence: DeepSeek and Huawei are "collaborating on future AI chip designs," and DeepSeek plans to start using the Ascend 950DT for pre-training as soon as 2026's end. We could read only that summary, not the paywalled article, so we are attributing each point to The Information and not adding details we cannot see. Two things follow from the wording:

1. Co-design, not just purchase. The report describes DeepSeek influencing Huawei's roadmap, which would make the relationship closer than customer and vendor. That fits a pattern in which the model builder tells the chip designer what its next architecture needs (memory capacity, interconnect, low-precision formats).
2. Pre-training is the hard part. Inference is the easier workload to move to new silicon. Pre-training stresses interconnect, memory bandwidth and software stack maturity far more, which is why it is the real test.

What Bloomberg reported in September

On September 4, 2026, Bloomberg reported that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Inner Mongolia. Syndicated coverage of that report adds the details that matter here:

• DeepSeek does not currently plan to use the 950DT for training, even though Huawei designed and marketed it for that purpose.
• DeepSeek tried training earlier models on Huawei parts and still uses Nvidia accelerators for training.
• Shortages of components, including top-end memory, will limit 950DT output to the low hundreds of thousands this year, and filling the order could take more than a year.
• DeepSeek was reportedly in talks to raise billions of dollars to fund the buildout.

A secondary write-up from tech-ish places the site in Ulanqab, about 350 km north-west of Beijing, with partial operation by late 2027 or early 2028 and says neither company commented. Treat the location and timing as reported by that outlet, not confirmed.

Chip moving along supply-chain trays, illustrating DeepSeek Huawei Ascend 950DT supply

Reconciling the two reports

Three readings are possible, and the public record does not let us pick one:

1. Plans changed. Between early September and mid-October DeepSeek may have moved pre-training onto the 950DT's schedule, possibly because of the co-design work.
2. Different scope. The 160,000-chip cluster may be for serving, while a smaller allocation is for pre-training experiments that begin at the end of 2026.
3. Different sources, different emphasis. Anonymous sources describe plans that are still fluid, and "as soon as the end of 2026" is a possibility rather than a commitment.

We would not treat either account as settled. "As soon as" is the operative phrase in The Information's summary, and Bloomberg's "does not currently plan" was a statement about September.

Why supply decides the timeline

Design collaboration can shorten a roadmap on paper, but the chips still have to be built. The reporting points to two bottlenecks:

• Memory. The 950DT relies on Huawei's own high-bandwidth memory, and shortages of top-end memory are cited as the reason output stays in the low hundreds of thousands. For how memory partnerships shape the Western side of this market, see our post on Samsung, TSMC and NVHBM base dies.
• Launch timing. Coverage says Huawei plans to launch the 950DT in the fourth quarter of 2026. A slip there delays everything downstream, including any pre-training start at the end of the year.

Grid of chip squares with one lit row, representing a GPU cluster for AI training

Scale context helps: our post on China's 24 GW of compute versus the US's 56 GW covers the Inner Mongolia buildout, and Nvidia's reported China chip share shows how contested the market already is.

Why this matters beyond DeepSeek

For Nvidia. The more Chinese labs move even a slice of training to domestic silicon, the more China revenue that Nvidia has already begun to treat as small shrinks further. The reports do not say Nvidia is displaced today: Bloomberg's account has training still on Nvidia hardware.

For policy. US export controls assume Chinese labs need Nvidia's top chips for frontier training. A credible domestic training path would weaken that lever. It would also complicate the federal posture toward DeepSeek models; see our coverage of DHS ordering DeepSeek removed from federal civilian agencies.

For developers. DeepSeek's models stay open-weight and cheap regardless. If the co-design does produce hardware tuned to its architectures, the likely effect is lower serving cost and faster release cadence, which feeds the price pressure we described for DeepSeek V4.1 Flash and in the Jefferies analysis of cheap Chinese open models.

How we got here: a short timeline

Coverage compiled by tech-ish traces how DeepSeek's hardware position evolved. Treat the dates as that outlet's summary of earlier reporting:

The arc is consistent: a lab that was once squeezed into efficiency tricks by hardware limits is now trying to shape the hardware itself. The open question is whether the software stack and supply chain catch up to the ambition.

How to read anonymous-source chip reports

Chip-roadmap stories are among the hardest to verify because the facts live inside two private companies. A few habits help:

1. Separate the three layers: a design collaboration (cheap to start), a purchase order (needs money and supply) and a production training run (needs mature software and reliable interconnect). A report about the first does not imply the third.
2. Watch the verbs. "Plans to," "as soon as" and "could" describe intent. "Has trained" or "released a model trained on" would be evidence.
3. Look for confirmation in artifacts, such as a model card or technical report that names hardware, a Huawei product launch with specifications, or a customs and capacity datapoint.
4. Expect revisions. Between September and October the training story already moved from "not planned" to "possible by year-end." Another change before the end of 2026 would not be surprising.

What we could not verify

• The full text of The Information's report, including who the sources are and whether "pre-training" refers to a flagship run or a smaller experiment.
• Any statement from DeepSeek or Huawei.
• The dollar value of the order. One outlet cited roughly $2.56 billion, but Bloomberg-based coverage named no figure, so we are not using it.
• Whether the 950DT's software stack, including its collective-communication library, can support large pre-training runs reliably. Earlier coverage reported DeepSeek had trouble training on Huawei parts in 2025.

What to watch next

1. Huawei's fourth-quarter 950DT launch and its published specs.
2. A DeepSeek technical report that names the hardware used for training or serving.
3. Any official comment from either company.
4. Memory supply news that changes the "low hundreds of thousands" estimate.
5. Whether Washington adjusts export rules in response.

Related reading

• China 24 GW vs US 56 GW compute, SemiAnalysis
• Nvidia vs Huawei China AI chip share, claimed vs verified
• DHS orders DeepSeek removed from federal agencies
• DeepSeek V4.1 Flash open weights and pricing
• Jefferies on cheap Chinese open models
• Samsung, TSMC and NVHBM base die

Reports are anonymous-source journalism and details may change; accurate as of October 19, 2026.