DeepSeek plans 160,000 Huawei Ascend chips for inference while training stays on Nvidia
Huawei claims its Ascend chips have outsold Nvidia in China as of October 2026, but DeepSeek keeps its training work on Nvidia hardware. Unit sales and training workloads are separate measures, and sizing either company's China business needs both.
The Investor · Invest desk
Key figures: Nvidia's share of China's AI accelerator market may fall to around 8% in 2026, from about 40% in 2025. Domestic chipmakers could take approximately 50% by 2026. Moving training from CUDA to CANN is predicted to carry roughly 50% higher costs and time.
- around 8% Nvidia's 2026 China AI accelerator share, projected, claim 16 down from about 40% in 2025
- approx. 50% Domestic chipmakers' possible share by 2026, claim 15 projection; source not named
- roughly 50% Predicted extra cost and time, training CUDA to CANN, claim 14 predictor not named in the report
| Figure | Value | Note | Claim |
|---|---|---|---|
| Nvidia's 2026 China AI accelerator share, projected | 8% around | down from about 40% in 2025 | 16 |
| Domestic chipmakers' possible share by 2026 | 50% approx. | projection; source not named | 15 |
| Predicted extra cost and time, training CUDA to CANN | 50% roughly | predictor not named in the report | 14 |
What happened
- DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips for inference, according to the cryptobriefing.com report.
- Eric Xu, Huawei's rotating chairman, has framed the company's pitch around dependable domestic supply over raw peak performance.
- Moving a training workload from Nvidia's CUDA to Huawei's CANN is predicted to carry roughly 50% higher costs and time.
- The report projects Nvidia's share of China's AI accelerator market at around 8% in 2026, down from about 40% in 2025.
- Washington has granted limited approvals for certain Nvidia models despite its broader restrictions on advanced chip exports to China.
Why it matters
- exposure A falling accelerator share does not remove the training demand that still runs on Nvidia, so the China exposure to size is the base of labs training on it.
- constraint Until CANN narrows the training gap, Huawei's sales to the top labs are capped at the inference side of their budgets.
- decision A lab like DeepSeek has to run two stacks, Ascend for inference and Nvidia for training, and carry Washington's approval and enforcement risk on the Nvidia side.
The slide from 40% to 8% is 32 points (80% of Nvidia's share) [10], and domestic makers at roughly 50% leave 42% of the market to other suppliers [15][11]. The article does not name who produced either projection, and Huawei's claim comes without a unit count. If share is the measure, outselling a vendor projected at 8% needs Huawei to clear 8% and nothing more [13][12], so the claim fits any Huawei share from just above 8% to the full domestic 50%.
DeepSeek shows where that volume goes. Inference deployments are rolling out with more agility than training ones [5], and the reason is software. CUDA is deeply entrenched across the AI industry [3], Huawei's alternative is called CANN [4], and applied to a training run the predicted 50% penalty means about 1.5 times the cost and 1.5 times the time [14][17]. None of DeepSeek's 160,000-chip order moves a training run, on the article's account [1].
Smuggling is the second channel. US authorities have arrested people involved in smuggling Nvidia chips into China, with cases ongoing as of October 2026 [6]. Chips that arrive that way would not appear in Nvidia's reported China sales, so we'd expect the Nvidia hardware running in China to exceed what a market share of 8% implies.
In our view Huawei's domestic opportunity today is inference, and Nvidia's China position is better measured by who trains the frontier models than by accelerators sold. If training is a small pool of spend next to inference, holding it is a small prize, and the 32-point slide is the bigger fact [10]. If CANN closes the 50% gap, the holdout ends, and with it what the article calls the source of Nvidia's remaining leverage [8].
What to watch
- Whether any top Chinese lab announces a training run on domestic hardware, a sign CANN is closing the training gap.
- How Washington handles its limited Nvidia approvals: tighter rules would speed the shift to Huawei, looser ones could give labs a reason to stay put.
- Whether Huawei publishes a unit count behind its October 2026 claim that Ascend outsold Nvidia in China.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption40
- Hype gap+25
- Incentives
- Insufficient
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips for inference, while keeping Nvidia hardware for its training workloads.
- [2]
Eric Xu, Huawei's rotating chairman, has framed the company's pitch around dependable domestic supply rather than raw peak performance.
- [3]
Nvidia's CUDA platform, the toolkit developers use to write code that runs efficiently on Nvidia chips, has become deeply entrenched across the AI industry.
- [4]
Huawei offers its own alternative to CUDA, called CANN.
- [5]
Inference deployments appear to be rolling out with more agility than training deployments, which helps explain why Huawei's sales can climb while Nvidia keeps a foothold at the top end.
- [6]
US authorities have arrested individuals involved in smuggling Nvidia chips into China, with cases still ongoing as of October 2026.
- [7]
Washington has granted limited approvals for certain Nvidia models despite its broader restrictions on advanced chip exports to China.
- [8]
Winning training, where switching costs are highest and performance matters most, is the harder prize for Huawei and the source of Nvidia's remaining leverage.
- [9]
Many of China's leading AI developers still rely on Nvidia hardware to train advanced large language models.
- [10]
Nvidia's projected China share falls 32 percentage points, from about 40% to about 8%, which is an 80% fall in its share.
- [11]
If domestic makers take about 50% and Nvidia about 8%, about 42% of China's AI accelerator market is left to other suppliers.
- [12]
If share is the measure, Huawei outselling Nvidia at a projected 8% needs Huawei above 8% and, as a domestic maker, no more than the full domestic 50%.
- [13]
Huawei claims its Ascend AI chips have outsold Nvidia within China as of October 2026.
- [14]
A transition for training workloads from CUDA to CANN is predicted to carry roughly 50% higher costs and time.
- [15]
Domestic AI chipmakers could capture approximately 50% of China's AI accelerator market by 2026.
- [16]
Nvidia's share of China's AI accelerator market may fall to around 8% in 2026, down from about 40% in 2025.
- [17]
A roughly 50% higher cost and time for training transition equals about 1.5 times the cost and 1.5 times the time.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- cryptobriefing.comChina struggles to shift tech firms away from Nvidia chips
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.