DeepSeek is buying 160,000 Huawei chips, and the detail that matters is what it will not use them for

DeepSeek is preparing to order at least 160,000 of Huawei's next-generation Ascend 950DT accelerators for a new gigawatt-scale data centre at Ulanqab in Inner Mongolia, according to a Bloomberg report.

If it completes, it would be one of the largest known clusters of domestically manufactured Chinese AI accelerators anywhere — a roughly $2.56 billion order placed by China's most closely watched AI lab, with Nvidia locked out of the market.

It reads as a decoupling milestone. The details complicate that reading considerably.

The Caveat

DeepSeek does not currently plan to use the 950DT chips for training.

That is the part worth sitting with. Huawei designed and marketed the 950DT for training — the computationally brutal step where a model is actually built. DeepSeek intends to use them to run its models, not to create them, and has so far relied on Nvidia accelerators for training even after attempting to train earlier models on Huawei silicon.

So the largest domestic-silicon order yet placed by a Chinese lab is, on current plans, an inference order. The dependency that actually constrains DeepSeek — access to Western chips for training — is not what this purchase resolves.

The Second Constraint

There is also a supply problem, and it is Huawei's.

Component shortages, particularly for high-end memory, are reported to cap total 950DT output at the low hundreds of thousands of units this year. Against that ceiling, a single 160,000-unit order is enormous, and fulfilling it could stretch beyond twelve months.

The data centre is scheduled to come online partially by late 2027 or early 2028 — a timeline that tells you this is a bet on capacity that does not exist yet, from a supplier that cannot currently build it at the required rate.

Why It Still Matters

None of which makes the order insignificant. Three things follow from it regardless.

It commits Huawei's constrained output to a domestic customer, at a scale that squeezes everyone else in China queuing for the same chips. It gives Huawei a reference deployment at gigawatt scale, which is the credential its accelerator business has lacked. And it demonstrates that a serious Chinese lab will now plan multi-year infrastructure around domestic silicon rather than waiting for export policy to loosen.

That last point is the durable one. Export controls were designed to make Chinese labs dependent on American hardware. A 160,000-chip order placed against a supplier that cannot yet deliver is not the behaviour of a company expecting that dependency to be restored.

The Energy Question

A gigawatt is the other number in this story, and it gets less attention than it deserves.

That is a full nuclear reactor's worth of generation dedicated to a single facility. Inner Mongolia is a plausible site for exactly that reason — cheap land, cold air, and proximity to coal and wind generation — but a build of this scale is a regional energy commitment as much as a computing one.

What To Watch

Whether the order is confirmed at the reported volume or quietly reduced as Huawei's memory constraints bite. Whether DeepSeek's stated inference-only plan holds, or whether it shifts training onto the 950DT once the cluster exists — which would be the genuinely significant development. And whether other Chinese labs follow with orders of their own, or wait to see whether this one is delivered.

For now the correct reading is narrower than the headline number suggests. This is a very large bet on Chinese silicon for the easier half of the problem.