DeepSeek has told Huawei it wants at least 160,000 Ascend 950DT accelerators for a new data center in Ulanqab, Inner Mongolia, a facility designed to eventually draw a full gigawatt of power 1. On the surface, that looks like the loudest vote yet for Chinese silicon over Nvidia from the country's most closely watched AI lab. Read past the headline number and the story is narrower, and more useful to investors: DeepSeek is moving only inference, the work of serving finished models to users, onto Huawei hardware. Training, the far more compute-intensive job of building the next model, stays on whatever Nvidia chips the company can still secure 2 3. That split marks the real boundary of what Chinese AI hardware can do today, not what Beijing wants it to do, and it is a more precise signal than the 160,000-chip figure by itself.

The Line DeepSeek Won't Cross Yet

The caution is not abstract. Last year DeepSeek spent months trying to train a model on older Ascend chips with Huawei engineers working on site, and the effort never produced a single successful full training run 3. That failure is why the new order is being described by people close to it as an inference-only deployment, with training workloads continuing to run on Nvidia hardware DeepSeek already holds or can still import under Beijing's narrower approved channels 2 3. Inference is the more forgiving workload: it tolerates lower interconnect bandwidth and less mature software tooling than training a frontier model from scratch. Betting a gigawatt facility on Ascend chips for that job is a materially smaller risk than betting the next model release on them.

Why Ulanqab

The site was not chosen for symbolism. Industrial electricity in the region runs around 0.35 yuan per kilowatt-hour, and the average annual temperature is about 4.3 degrees Celsius, cold enough to allow free-air cooling for roughly ten months of the year 2. For a facility meant to run inference around the clock at gigawatt scale, that combination of cheap power and near-free cooling does more for unit economics than any single chip specification. DeepSeek wants at least part of the capacity live by late 2027 or early 2028, a timeline explicitly tied to how fast Huawei can actually deliver chips rather than to DeepSeek's own construction schedule 1.

The Ceiling Is Memory, Not Logic

That delivery risk is the piece markets have been underpricing. SMIC can already fabricate enough Ascend-class logic die for well over a million chips a year. What China cannot yet do at scale is wrap enough high-bandwidth memory around that logic 5. Huawei is targeting roughly 600,000 to as many as 1.6 million Ascend dies produced in 2026, but HBM supply, dominated by CXMT output with a shrinking stockpile of foreign memory filling the gap, is the actual constraint on how many finished, sellable chips reach customers 5. Estimates of 2026 output for the 950DT specifically cluster around 200,000 to 300,000 units 1 4. If that holds, DeepSeek's order alone could absorb most of a full year of Huawei's supply of this chip, ahead of every other domestic buyer competing for the same wafers.

The Number Behind the Number

This order lands on top of a market share shift that was already underway. Bernstein now expects Nvidia's share of China's AI chip market to fall from roughly 40% to about 8% by the end of 2026, with Huawei's share climbing to close to half 6. That forecast predates DeepSeek's move, which puts China's highest-profile AI lab onto the Ascend side of the ledger at meaningful scale for the first time, a validation Huawei has not previously had from a developer of DeepSeek's stature. It does not, however, resolve the training gap: Beijing has separately steered its own approved Nvidia H200 allocations toward training use only, pushing routine inference work toward domestic chips by policy as much as by DeepSeek's own choice 4.

The Investor Read

The distinction worth pricing is between two different claims. One is that Huawei has cloned Nvidia's full stack; the other is that inference, the highest-volume, most predictable revenue line in AI infrastructure, is now provably fundable on Chinese silicon at gigawatt scale, while training remains a chokepoint neither Huawei nor SMIC has closed. Only the second claim is currently supported by evidence. That favors suppliers tied specifically to the inference buildout, domestic memory producers such as CXMT, and the packaging and testing houses serving Huawei's Ascend line, over a broader thesis that Nvidia has been engineered out of China's AI stack entirely. DeepSeek is also reportedly developing its own inference chip to reduce dependence on both Huawei and Nvidia, a project that would take three to five years to reach volume production even if it succeeds, underscoring that this order is best read as a multi-year hedge rather than a resolved technology bet 3.