DeepSeek reportedly puts Huawei Ascend into training — not just the inference lane

What happened

On September 23, 2026, TrendForce reported that DeepSeek reportedly plans to expand use of Chinese-made AI chips, including Huawei’s, for training as well as inference. The piece synthesizes Commercial Times citing The Information on closed-door investor remarks by founder Liang Wenfeng. Huawei could reportedly deliver a new batch of AI training chips as early as the fourth quarter of 2026. How many chips that batch contains is still uncertain. Liang reportedly said he believes Huawei’s chips could catch Nvidia within a few years. That is the founder’s line, carried second-hand. It is not a lab benchmark.

Chosun Biz, inside that same TrendForce account, describes the split Chinese labs have been living with. Domestic chips, Huawei included, have mostly handled inference, which needs less compute. Large-scale training has still leaned on Nvidia. Huawei said in April that its chips had been partly used to train DeepSeek’s lightweight V4-Flash model. DeepSeek is training a 2-trillion-parameter model and has talked about an eventual 8-trillion-parameter target. Liang’s rough math, per Chosun Biz via TrendForce, is that an OpenAI-scale training job would need around 50,000 Nvidia GB300 chips or roughly 200,000 Huawei Ascend 950 chips. That is a founder estimate, not a purchase order.

The same synthesis includes a supply check. Tom’s Hardware notes that Huawei’s next Ascend 900-series accelerators are aimed at China, not export, and that SMIC’s limited 7nm- and 6nm-class capacity is still restricting Huawei AI hardware supply. A Securities Times roadmap note, again via TrendForce, brings the Ascend 960DT forward to the first quarter of 2027, three quarters earlier than the prior plan, and aims the Ascend 960PR at the third quarter of 2027. Those dates are roadmap context. Earlier Bloomberg reporting cited by TrendForce put DeepSeek’s Inner Mongolia plan at least 160,000 Ascend 950DT chips, framed mostly as running models rather than training them. The training-intent report is the newer piece. The two uses stay separate.

On fundraising, the same Chosun Biz account via TrendForce says DeepSeek is investing about 30 billion yuan in compute expansion and finalizing a second round near 50 billion yuan, toward a 500 billion yuan valuation. Those figures come from trade press. They stay labeled that way.

Reuters added a same-day report from Beijing, a question-and-answer by Eduardo Baptista and Laurie Chen. Under U.S. export controls, Huawei has become the de facto primary supplier for advanced AI workloads in China. The strategy Reuters describes is to build the best chip the sanctions allow, then use advanced packaging and networking to link thousands of chips into super nodes, competing at the system level rather than on single-chip FLOPS. Reuters flags two open questions. One is whether that system architecture meaningfully narrows the Nvidia gap. The other is whether Huawei can scale fabrication and yields enough to meet demand. Until both answers are clearly yes, China’s AI compute stays constrained. That frame keeps a training plan separate from training capacity.

Why it matters

Training is the expensive chapter of the AI boom. Inference can limp along on last year’s silicon. Frontier training cannot. If DeepSeek truly shifts work toward Huawei Ascend for training, domestic chips start showing up as a spending plan. A closed-door quote does not move the binding constraint. SMIC’s advanced-node supply, and Huawei’s ability to ship training silicon a customer can actually use, still decide whether a fourth-quarter 2026 batch is a real cluster.

For readers outside China, set this next to the Ascend 960 schedule and the earlier Inner Mongolia figure of at least 160,000 chips. A plan to train on Ascend and parity with Nvidia are different claims. Reuters’ super-node account is the practical China strategy under sanctions. Korea still owns HBM. Taiwan still owns the densest advanced-packaging queue. Training on Ascend in China still depends on wafer starts and yields these reports do not measure.

What to watch is confirmation from another major outlet of fourth-quarter 2026 training-chip shipments, with a volume, a product, and a customer; whether DeepSeek’s next model notes name Ascend for training or for serving; any concrete talk of SMIC 7nm- and 6nm-class utilization; Ascend 960DT timing from samples into mass production in the first quarter of 2027; and whether the fundraising round actually closes near 50 billion yuan and the 500 billion yuan valuation trade press is floating. Keep every shipment figure labeled as reported until a filing or a major outlet locks it down.

The Inner Mongolia plan, framed mostly as inference, is DeepSeek’s 160,000-chip ask. The factory limit beside it is Huawei’s Ascend shortage. The next chip date is Ascend 960 landing early.