DeepSeek’s next big move is not another model launch. It is infrastructure.
The Chinese AI company is reportedly preparing to deploy at least 160,000 Huawei Ascend 950DT accelerators at a large data centre under construction in Inner Mongolia. If the plan materialises, it would create one of the largest known clusters of Chinese-made AI chips and mark a significant step in Beijing’s long-running effort to reduce dependence on Nvidia hardware.
That sounds like a simple “Huawei replaces Nvidia” story.
It is not.
The more interesting detail is that DeepSeek is reportedly planning to use the Huawei chips mainly for inference, not for training its frontier models. Training, for now, is still expected to rely heavily on Nvidia accelerators.
So China may be building an alternative AI stack, but it has not yet fully escaped the hardware it wants to replace. Technological independence, like most break-ups, appears to be complicated.
Why 160,000 Chips Matter
The scale is substantial.
DeepSeek’s planned Inner Mongolia facility is reportedly being designed at roughly 1 gigawatt of power capacity. At full utilisation, that is enough electricity to rival the consumption of hundreds of thousands of homes. The Huawei deployment would account for only part of the eventual compute infrastructure planned for the site.
The exact capital expenditure for the data centre and chip deployment has not been publicly disclosed, so any neat multibillion-dollar price tag circulating online should be treated cautiously.
What is clear is that Huawei has spent heavily building the technology behind this push. In 2025, the company invested CNY192.3 billion in research and development, equal to 21.8% of annual revenue. Its cumulative R&D spending over the previous decade exceeded CNY1.38 trillion.
That is what semiconductor competition actually looks like: less dramatic keynote lighting, considerably more money.
Huawei’s 950DT Is Built For The AI Era
Huawei says the Ascend 950DT is designed for both model training and the decode stage of AI inference.
The processor is expected to offer 144GB of memory, memory bandwidth of 4TB per second and total interconnect bandwidth of 2TB per second. Huawei has scheduled the chip for availability in the fourth quarter of 2026.
The company is also building the Atlas 950 SuperPoD around the same chip family, with configurations scaling to thousands of accelerators working as a single logical computing system.
DeepSeek therefore gives Huawei something equally valuable: a demanding customer at enormous scale.
If the deployment works well, Huawei gets a powerful proof point for its hardware ecosystem. DeepSeek gets a domestic inference platform less exposed to US export controls.
On paper, everybody smiles.
Silicon tends to be less sentimental.
The Nvidia Problem Has Not Disappeared
DeepSeek’s reported decision exposes the central weakness in China’s semiconductor transition.
Running an already-trained model and training a frontier model are very different workloads. Training demands exceptional software maturity, high-speed networking, memory performance and the ability to keep enormous clusters working efficiently for long periods.
DeepSeek has experimented with Huawei hardware for model development, but reports indicate that it still prefers Nvidia for its most demanding training workloads.
And Nvidia is hardly looking vulnerable globally.
The company reported $96.2 billion in quarterly revenue for its latest fiscal quarter, including $89 billion from its Data Center business, up 117% from a year earlier. Nvidia also said its current outlook assumes no Data Center compute revenue from China.
China may become a smaller part of Nvidia’s sales story.
That is not remotely the same as saying Nvidia has become a smaller part of AI.
The Upside For China Is Bigger Than DeepSeek
The significance of the planned deployment is not whether 160,000 chips immediately outperform Nvidia.
It is whether Chinese companies can build an ecosystem where domestic accelerators become progressively “good enough” across more workloads.
Large-scale deployment produces operational data, exposes software weaknesses, gives developers experience and creates commercial demand that can finance another generation of hardware.
DeepSeek has already pushed efficiency aggressively. Its current DeepSeek-V4 family includes a 1.6-trillion-parameter Pro model and a smaller Flash variant, both supporting context windows of up to one million tokens. The company has also continued to position lower-cost inference as part of its competitive strategy.
Pair efficient models with domestic hardware, and China begins constructing something strategically more important than a single benchmark victory: an AI supply chain it can continue operating under geopolitical pressure.
The Catch Is Supply
There is another inconvenient detail.
Huawei itself has production constraints. High-end memory and other advanced components remain difficult to source at scale, and fulfilling an order as large as DeepSeek’s may take considerable time.
So the 160,000-chip plan is less a declaration of independence than a stress test of China’s entire semiconductor ecosystem.
Can Huawei manufacture enough?
Can the software mature quickly enough?
Can DeepSeek operate enormous domestic clusters efficiently?
And can China eventually move from inference substitution to genuinely competitive frontier-model training?
Those questions matter far more than whether Nvidia loses one order.
For now, DeepSeek’s move sends a clear message: China is no longer waiting for a perfect domestic alternative before deploying one.
Sometimes ecosystems are not built after the technology becomes mature.
They become mature because somebody is willing to deploy 160,000 chips and discover what breaks.
Read More: India’s 2026 AI Ambition
