Huawei is narrowing Nvidia’s lead in China’s artificial intelligence chip market, gaining ground as restrictions on American technology push Chinese companies towards domestic suppliers. But rising shipments tell only part of a contest that also depends on software, manufacturing capacity and the cost of running increasingly complex AI systems.
The competition carries consequences beyond the two companies. For Beijing, a competitive domestic supplier would reduce dependence on foreign technology. For Washington, Huawei’s progress raises questions about whether export controls are slowing China’s AI development or helping establish a rival computing industry.
The available evidence points to growing commercial traction for Huawei, alongside continuing Chinese demand for Nvidia’s products.
Huawei gains ground in a restricted market
Nvidia accounted for 34% of China’s data-centre AI accelerator shipments in the first half of 2026, compared with Huawei’s 28%, according to IDC figures reported by Reuters on October 13. Nvidia shipped 678,000 accelerators, down 44% from a year earlier, while Huawei shipped 572,000, up from more than 300,000. US export restrictions and Beijing’s promotion of domestic alternatives have helped reshape the market. Those shipment figures measure units, however, rather than revenue or the computing power delivered by each processor.
That distinction matters. A company can gain market share by supplying processors that customers can obtain, even when competing products offer advantages for particular tasks. For buyers, availability becomes part of the calculation alongside performance.
There is also evidence that Nvidia retains considerable appeal. Reuters reported on October 20 that ByteDance, Alibaba and Tencent had ordered more than 3 million Nvidia RTX PRO 5500 workstation graphics cards, with a combined sticker price exceeding $35 billion. The report cited people familiar with the orders. Nvidia could not guarantee full deliveries, and the three buyers did not respond to Reuters’ requests for comment. The reported orders indicate potential demand, rather than completed sales.
Together, the reports suggest a market in which Chinese companies are expanding domestic purchases while continuing to seek American hardware. Huawei’s opportunity is substantial, but customer preference cannot be separated from restrictions on what buyers can obtain.
A challenge built around entire computing systems
Huawei’s strategy extends beyond developing a faster individual chip. It is building systems that connect processors, memory and networking equipment, seeking to make large groups of chips operate efficiently together.
In its September 2025 computing roadmap, Huawei described an Atlas 950 SuperPoD containing up to 8,192 Ascend processors, with availability planned for the fourth quarter of 2026. The company also outlined proprietary high-bandwidth memory for its Ascend 950 chips, targeting different requirements in model training and inference — the process of running a trained model to produce answers.
The approach makes communication between processors central to performance. When an AI workload is distributed across thousands of chips, moving information efficiently becomes essential. Adding processors alone does not establish that a system will complete a particular task faster or more cheaply.
Huawei has since brought forward parts of its timetable. In a September 2026 presentation, it said the Ascend 960DT would become available in the first quarter of 2027, three quarters ahead of its original schedule. The Ascend 960PR was scheduled for the third quarter. These remain company commitments, rather than independent confirmation of delivery or customer performance.
For customers evaluating such systems, the decisive comparisons include electricity consumption, reliability, installation costs and how effectively existing applications run. Headline computing specifications provide only part of that assessment.
Software will help determine who keeps customers
Nvidia’s competitive position rests partly on CUDA, its platform for programming graphics processors. Its development toolkit includes computing libraries, compilers, debugging tools and software for improving application performance.
That creates a practical challenge for competitors: customers need applications that work reliably, alongside capable hardware. Moving to another platform can require engineering time, testing and changes to software.
Huawei is investing in that layer. In September, it said its Compute Architecture for Neural Networks, or CANN, had moved to community-driven open-source development. The company reported more than 5,200 monthly active developers and said more than 40 models had been natively pretrained using Ascend and CANN.
Independent evidence also illustrates potential difficulties. A July research preprint examining two inference workloads on an Ascend 910 system reported that achieving reliable operation required software patches and disabling some performance features. The study concerns a specific setup and cannot establish how all Ascend products perform.
Huawei’s next test is whether growing purchases translate into dependable systems that customers continue using as their AI workloads expand. Winning those decisions would turn an opening created by supply restrictions into a more durable competitive position.
