13 SEP 2026 — DeepSeek has ordered 160,000 Huawei Ascend 950DT accelerators for a gigawatt-scale site at Ulanqab in Inner Mongolia. The order was reported on 4 September. The price of the chip was reported on 10 September. Put the two together and the order is worth about 40 billion yuan, or roughly 5.9 billion US dollars.
That is arithmetic on two separate pieces of reporting rather than a figure either company has published, and neither has confirmed the order at all.
What the numbers say together
Huawei is asking more than 250,000 yuan for the 950DT, up 20 to 50 per cent on quotes from two months earlier as the high-bandwidth memory shortage reached Chinese accelerator makers. At that price, 160,000 units comes to 40 billion yuan — about 5.9 billion dollars at current rates, which you can check against a currency converter if the rate has moved by the time you read this.
The more striking figure is the share of supply. Component shortages, advanced memory among them, are expected to hold Huawei's 2026 production of the 950DT to somewhere between 200,000 and 300,000 units.
One customer has therefore ordered between 53 and 80 per cent of a year's output of the part.
Inference, not training
The detail that separates this from a straightforward capacity announcement is what the chips are for. DeepSeek does not currently plan to use the 950DT for training, even though Huawei designed and marketed it for that harder job.
Buying a training-class accelerator for inference is not obviously irrational, for a few possible reasons.
The 950DT might just be what Huawei can supply in volume to a buyer locked out of Nvidia. The memory — 144 GB per card, on Huawei's own HBM design — is enough to hold a large model resident for cheap serving. Or DeepSeek's real training constraint might be somewhere else entirely, making inference capacity the better investment.
Nothing published distinguishes these, and the distinction matters for anyone reading the order as a signal about Chinese training capability. An inference build says a company expects demand for its existing models. It says considerably less about the next one.
The gigawatt is the harder half
The Ulanqab campus is described as gigawatt-scale, drawing enough at full utilisation to supply roughly 750,000 homes.
That number belongs beside the chip count because the chips are the cheaper half of the problem. We have followed the pattern in AI data-centre power for months, and it holds here: accelerators can be bought with a purchase order, and a gigawatt of firm supply cannot.
Inner Mongolia is a deliberate choice in that light. It has the grid capacity, the land, the cold winters that cut cooling load, and the coal and wind that make the power available at all. What it does not have is proximity to users, which is a real cost for inference specifically — latency matters more when serving a query than when running a training job that takes weeks.
What this says about the export controls
The order is the clearest measurement yet of what the export restrictions have and have not done.
They have not denied China inference capacity at scale. A domestic accelerator exists, it carries 144 GB of memory on a domestically designed HBM stack, and a major lab is committing to it in volume for a gigawatt site. That is the capability the controls were meant to withhold, being built anyway with local parts.
They have, however, made it scarce and expensive. The chip costs 20 to 50 per cent more than it did in July because the memory comes through grey channels, and Huawei can make only 200,000 to 300,000 of them this year. A single order now absorbs most of that. Every other Chinese buyer is competing for a remainder that one customer has already claimed.
As policy, that is a supply constraint, not a denial. A constraint of that shape tends to consolidate a market rather than shrink it: buyers who can commit billions get served, and the rest wait.
The timeline is the soft part
DeepSeek wants at least part of the capacity running by the end of 2027 or early 2028, subject to Huawei's production.
That qualifier is carrying most of the risk. An order for 53 to 80 per cent of annual output cannot be filled in one batch by a manufacturer already short of memory, so the schedule depends on a supply chain under the same export-control pressure that pushed the chip's price up in the first place.
This looks more like a commitment than a scheduled delivery. Both halves — the silicon and the gigawatt — are years out, and the announcement effect arrived immediately.
What to watch
Confirmation, first. Neither DeepSeek nor Huawei has acknowledged the order, and a reported purchase intent is not a signed contract with a delivery schedule.
Then the price. If Huawei is allocating most of a year's output to one buyer, the remaining customers are bidding for what is left, and the 950PR's move from 60,000 to over 80,000 yuan this year suggests that market is already tightening. A single order this size is itself a reason for the next quote to be higher.
