From Nvidia to Ascend: DeepSeek's Founder Makes Training on Huawei Chips a Strategic Bet

On Sunday, September 21, 2026, The Information published an exclusive that crystallizes a shift the AI hardware world has watched accelerate for over a year: DeepSeek founder Liang Wenfeng has told investors that training on Huawei chips is one of the company’s biggest strategic bets, and that he expects Huawei Technologies to begin delivering AI training chips to DeepSeek as early as the fourth quarter of this year, according to two people familiar with the matter.

The report, by Juro Osawa and Qianer Liu, also includes a detail that places the bet in context: DeepSeek is currently training a 2-trillion-parameter model and has plans for an 8-trillion-parameter successor. If the Huawei timeline holds, the world’s most famous efficiency-focused AI lab will attempt its most ambitious training runs yet — on Chinese silicon, not Nvidia’s.

Why Training Is the Line That Matters

DeepSeek’s relationship with Huawei hardware is not new, but until now it has been primarily an inference story.

In April 2026, Reuters reported — citing The Information — that DeepSeek’s V4 model would run on Huawei’s latest chips, making it the first frontier Chinese model to be served natively on Ascend silicon. In July, Reuters reported that DeepSeek was developing its own AI chip. And on September 4, Bloomberg reported DeepSeek’s plan to deploy at least 160,000 Huawei Ascend 950DT accelerators at a gigawatt-class data center under construction in Inner Mongolia — a cluster that could become one of the largest known concentrations of Huawei AI chips anywhere.

But those reports, however large in scale, described serving models to users. Training is a categorically harder problem, and it is where DeepSeek’s Huawei story has historically stalled.

In August 2025, the Financial Times reported that DeepSeek had tried and failed to train its R2 model on Huawei’s Ascend chips, suffering unstable performance during extended multi-GPU runs and ultimately reverting to Nvidia hardware while delaying the launch. That failure became the canonical citation for skeptics of Chinese self-sufficiency in AI compute: Huawei could rack clusters, the argument went, but frontier training demanded the software maturity, interconnect reliability, and ecosystem depth that only Nvidia’s CUDA stack had refined over fifteen years.

This is precisely what makes Liang’s statement to investors significant. It is not a procurement announcement — it is the founder of the lab that failed once, telling the people funding him that the next attempt is not an experiment but a strategy. Per The Information’s meta description, “a major priority for his company is to use more domestic chips to train its models,” with Huawei expected “to start delivering training chips to DeepSeek as early as the fourth quarter this year.”

The 2-Trillion-Parameter Backdrop

The scale of the models involved raises the stakes considerably.

DeepSeek built its reputation by extracting frontier-adjacent performance from constrained hardware — V3 was trained on 2,788,000 GPU-hours of Nvidia H800, a fraction of what Western labs spent. A 2-trillion-parameter model, and a planned 8-trillion one, are a different class of endeavor. Training runs at that scale stress every layer of the stack: memory bandwidth, interconnect topology, fault tolerance across tens of thousands of accelerators, and the software’s ability to keep utilization high for weeks.

Doing it on Ascend means DeepSeek’s engineers must solve those problems atop Huawei’s CANN software ecosystem, which lacks CUDA’s decade-plus of tooling, documentation, and community muscle memory. The August 2025 R2 failure showed exactly where that gap bites: unstable performance during extended multi-GPU runs is a systems-software symptom, not a raw-FLOPS one.

The counterargument is that DeepSeek is uniquely positioned to close the gap. This is a lab whose core competency is software engineering under hardware constraint — the dual-stream attention and FP8 training optimizations of the V3 era were exactly this kind of work. If anyone can absorb the friction of a less mature stack, it is the team that turned export-restricted hardware into a global efficiency benchmark.

Three Timelines Converging

The Q4 delivery window does not exist in isolation. Three clocks are ticking simultaneously.

Huawei’s roadmap clock. At its Huawei Connect conference on September 17, 2026, the company pulled forward the launch of its next-generation Ascend 960DT chip to Q1 2027 — two quarters earlier than the previously announced Q3 2027 window — and disclosed that the Ascend 960PR’s FP4 throughput target has doubled. A chip vendor accelerating its flagship by six months while its largest customer publicly commits to training on domestic silicon is not a coincidence; it is a coordinated ecosystem push. Huawei’s current top part, the Ascend 950DT, has seen market prices pass 250,000 yuan per unit amid a broader 20–50% price rise in Chinese AI chips driven by high-bandwidth-memory shortages.

The geopolitical clock. The report lands three days before the September 24 Trump–Xi summit in Washington, and a week after Treasury Secretary Scott Bessent confirmed the US and China agreed to establish an AI dialogue working group with a proposed national-security incident notification mechanism. US export controls have progressively locked Nvidia’s most capable data-center GPUs out of the Chinese market — which is exactly the vacuum Huawei’s Ascend line was built to fill. Meanwhile, House China Committee Chair John Moolenaar is publicly urging the administration to tighten controls further and close the “cloud services loopholes” that let Chinese labs distill American models. Every tightening of the screw makes DeepSeek’s domestic bet less optional and more existential.

The industry clock. The same week as The Information’s report, US hyperscalers are executing a parallel pivot away from Nvidia dependence: Microsoft is negotiating over 300,000 units of its Maia 300 accelerator for delivery next year, Meta’s Iris chip entered production this month, OpenAI’s Broadcom-co-designed “Jalapeño” processor has moved into a broader production strategy, and Amazon signed a multigenerational custom-chip agreement with Qualcomm. The geography differs, but the logic is identical — the biggest AI buyers are deciding that single-supplier dependence is the bigger risk, even at a hardware performance discount.

What to Watch

If Huawei does begin training-chip deliveries in Q4 2026, the first observable signal will not be a press release but an absence: the absence of a reversion. The 2025 R2 cycle showed what failure looks like — delays, a quiet retreat to Nvidia, and a launch pushed back. A 2-trillion-parameter model trained through to completion on Ascend hardware, at utilizations that don’t bankrupt the efficiency story DeepSeek is built on, would be the strongest evidence yet that the CUDA moat is crossable at frontier scale.

The second signal is what happens to the 8-trillion-parameter plan. That model’s training run would land squarely in the era of Huawei’s 960-generation silicon — meaning DeepSeek’s roadmap and Huawei’s roadmap are now, functionally, one roadmap. For a company of DeepSeek’s size to fuse its frontier ambitions to a single domestic supplier’s execution is either the boldest infrastructure bet in Chinese AI or its single greatest concentration of risk.

The third signal is competitive. The Information’s own keyword trail for the story includes Moonshot AI — a reminder that Kimi’s maker and every other Chinese frontier lab faces the same export-control wall and is watching whether Ascend training actually works. If DeepSeek succeeds, Huawei gains not just one customer’s order but a proof case that reshapes procurement across the entire Chinese AI sector. If it fails again, the fallback options are thinner than they were in 2025.

The Bet, Stated Plainly

Strip away the parameters and the diplomacy, and Liang Wenfeng’s message to investors reduces to one claim: the constraint that built DeepSeek — too little compute, too tightly controlled — has become the reason to master the compute the controls left behind. The lab that proved you could do more with less Nvidia now intends to prove you can do frontier without Nvidia at all.

Q4 2026 is when that thesis gets its first real test.