DeepSeek orders 160,000 Huawei chips for a site that cannot be filled yet

160,000 AI accelerators, ordered from a supplier that cannot currently build them. Bloomberg reported on 4 September that DeepSeek plans to deploy at least that many Huawei chips at a new data centre in Inner Mongolia, which would make it one of the largest known clusters of Chinese silicon anywhere.

The parts are Huawei's next generation Ascend 950DT. The site is expected to draw enough electricity to power roughly 750,000 homes, which puts it in the same power class as the American builds going up this year. The report was picked up by Mobile World Live, TFTC and Seeking Alpha, which all describe the order as at least 160,000 accelerators.

What is the DeepSeek Huawei Ascend order?

It is a reported plan to buy at least 160,000 Ascend 950DT accelerators for a single new data centre in Inner Mongolia. An accelerator is a chip built to do the heavy matrix arithmetic behind AI models far faster than a general purpose processor. Nvidia's GPUs are the best known examples, and Huawei's Ascend line is its in-house alternative.

DeepSeek is the Chinese AI lab best known for releasing capable models at low prices. Huawei is its most prominent domestic hardware partner, and the two have been working together on software as well as silicon, as covered in DeepSeek and Huawei going after CUDA, the programming layer that keeps most developers tied to Nvidia.

Nothing in the report says the chips have shipped. This is a plan, and the headline number describes what DeepSeek wants to install rather than what is running.

That distinction matters for readers who see 160,000 and assume a finished cluster. A data centre needs buildings, cooling, networking and a steady power supply as well as chips, and the reports describe the site as planned. The chip count tells you about ambition, while the delivery schedule tells you about timing.

What will DeepSeek use the chips for?

Inference, not training. DeepSeek does not intend to train models on these chips. It still relies on Nvidia hardware for that, despite several attempts to move off it.

Training is the long, expensive process of teaching a model from data. Inference is what happens afterwards, every time someone sends a question and gets an answer. The two stress hardware differently, and our explainer on inference chips versus training chips covers why a chip can be good at one and awkward at the other.

The order buys serving capacity for models trained somewhere else. That is a much narrower claim than the headline number suggests. It is also where Chinese demand actually sits now that assistants are being used at the scale of a billion monthly users, because every one of those users generates inference work.

The split also fits a wider pattern. Nvidia won the training race, and inference is a different problem with different winners, a point made in Nvidia won training, inference is a different problem entirely. A domestic chip that is merely adequate can still be useful if it is cheap to run and available.

Why can Huawei not fill the order?

Memory is the constraint. High-end memory shortages are expected to cap 950DT production in the low hundreds of thousands of units this year, and delivery on an order this size could take more than twelve months. So this is a statement of intent priced in gigawatts, with no date attached to it.

Modern accelerators pair the processor with stacks of fast memory sitting right beside it. A chip without that memory cannot be sold, however good the processor is. That makes memory supply a hard ceiling on output, separate from how many processors can be made.

The arithmetic is simple. If output is capped in the low hundreds of thousands, a single customer asking for at least 160,000 would take a large share of everything Huawei can make this year, leaving little for its other buyers. That is why the delivery estimate runs past twelve months.

Do export controls explain the shortage?

Only partly. Export controls were written to slow Chinese AI compute by restricting access to chips, and the thing holding this build back is memory supply. It is the same shortage pushing up the price of the RAM in your own machine, and another case of power and parts rather than policy setting the ceiling on what gets built.

China has been trying to close the memory gap at home. Our report on CXMT G5 DRAM entering mass production shows one domestic effort, although the 950DT report points to high-end memory as the part still in short supply.

The point for readers is that restricting chips and restricting what can be built are different things. A buyer can have the money, the site and the power, and still wait on a component.

Is DeepSeek building its own chip too?

DeepSeek has also been linked with designing its own silicon, which we covered in DeepSeek building its own AI chip to break free of Nvidia. The Huawei order sits alongside that effort rather than replacing it. Buying in Ascend hardware gives DeepSeek capacity now-ish, while anything it designs itself is further off.

Taken together, the picture is a company trying to loosen its dependence on one supplier from several directions at once, while still using that supplier for training.

At a glance: what is known and what is not

What to watch

Two numbers will tell you whether the cluster is real: Huawei's 950DT shipment figures over the next two quarters, and whether DeepSeek moves any training work onto Ascend instead of inference alone. Until one of those moves, 160,000 is a purchase order rather than a data centre.

A third signal is softer. If memory supply improves, the cap on 950DT output could rise and the delivery estimate shorten. If it does not, the order will be filled in slices, and DeepSeek will keep leaning on Nvidia and on whatever capacity it can rent.

For anyone following China's AI build-out, the useful lesson is to read orders against what factories can deliver. The broader spending picture, a 2026 to 2030 infrastructure plan from China's industry ministry that the South China Morning Post put at about 532 billion dollars, is in China budgeting 532 billion dollars to stop needing Nvidia, and this order shows how much of that money still depends on parts that are in short supply.

Key takeaways

• DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a new Inner Mongolia data centre, according to a Bloomberg report of 4 September.
• The site is expected to draw enough electricity to power roughly 750,000 homes.
• DeepSeek intends to run inference on the chips rather than training, and it still uses Nvidia hardware to train.
• Huawei cannot fill an order this size, because high-end memory shortages are expected to cap 950DT output in the low hundreds of thousands this year.
• Delivery could take more than twelve months, so the order is a statement of intent until 950DT shipment figures say otherwise.