Six Percentage Points Between NVIDIA and Huawei: The Competitive Landscape of China's AI Chip Market Is Undergoing Changes

Latest data from IDC shows that in the first half of 2026, NVIDIA's share in China's data center AI chip market dropped to 34%, followed by Huawei with 28%, narrowing the gap between the two to 6 percentage points.

A year ago, this gap stood at 31 percentage points.

More specific figures show: from January to June, NVIDIA's shipments of AI accelerators in the Chinese market reached 678,000 units, a year-on-year decline of 44%. Huawei shipped 572,000 units in the same period, almost doubling from more than 300,000 units in the same period last year. Jensen Huang, CEO of NVIDIA, has repeatedly criticized Washington's export restrictions, saying the policy forces the company to cede China's $50 billion AI chip market to Huawei.

This judgment is becoming a reality. But what is really worth exploring is: what did Huawei rely on to catch up? What exactly has NVIDIA lost?

The Cliff-like Drop on the Revenue Side

Behind the 44% decline in shipments is a more drastic contraction in NVIDIA's China data center revenue.

Colette Kress, CFO of NVIDIA, disclosed on the Q1 FY2026 earnings call that no data center Hopper architecture products were shipped to the Chinese mainland in the quarter, while this part of the business generated revenue of 4.6 billion U.S. dollars before the ban took effect. The company accrued $4.5 billion in inventory impairment and order cancellation losses, and expects $8 billion in revenue loss in the second quarter due to sales restrictions on H20. In the outlook for the second fiscal quarter, NVIDIA also assumed that any data center computing revenue from China would be zero.

Jensen Huang revealed at the 2026 Annual General Meeting of Shareholders that the proportion of revenue from China has dropped to about 9%. Three years ago, China once contributed at least one-fifth of NVIDIA's data center revenue.

This means that NVIDIA's revenue contribution in the Chinese market, which used to be at the level of billions of U.S. dollars annually, has dropped to nearly zero. The drop in share from 55% to 34% is largely the result of the overall shrinkage in total shipments — NVIDIA's shipments are declining, while Huawei's shipments are growing. What NVIDIA has lost is not just market share, but the actual space to participate in the Chinese market.

Ascend's Supply Bottleneck

Huawei's shipment volume of 572,000 units was achieved under supply constraints.

Around 812,000 units of Ascend 910C were shipped in the whole year of 2025, capturing nearly half of the domestic market share. Another 150,000 units were delivered in the first quarter of 2026, serving 37 leading large model enterprises. However, ByteDance and Alibaba have already cleared all existing inventory, while Tencent, Baidu and iFlytek are queuing up for supplies. The spot market is extremely short of available cards, and the secondary market once saw a premium of more than 30%.

The root of the production capacity bottleneck lies in two links: the yield rate of SMIC's second-generation 7-nanometer process, and the supply of high-bandwidth memory. The yield rate of 910C only climbed to 40% in early 2025, and it is still the ceiling of production capacity so far.

Bloomberg reported that Huawei plans to increase the production of its flagship chip 910C to about 600,000 units in 2026, almost doubling from 2025, and the total wafer output target of the Ascend series is raised to 1.6 million pieces. However, whether this target can be fulfilled depends on whether the process yield rate of SMIC can climb synchronously.

In other words, Huawei's 28% market share was obtained under the condition of limited production capacity. If supply was not a bottleneck, Huawei's shipment volume would be far more than 572,000 units.

Structural Changes in Non-GPU Accelerators

Another key clue to NVIDIA's declining share is hidden in the changes of product structure.

In the first half of 2026, China received 2,008,000 AI accelerators. Among them, shipments of GPUs widely used in AI model training fell by 26% to 991,000 units; while shipments of non-GPU accelerators dedicated to specific AI workloads increased by nearly 80% to 1,016,000 units, surpassing GPU shipments for the first time.

This crossover point is worth a closer look. GPUs are programmable general-purpose accelerators, suitable for mixed workloads of training and inference, and NVIDIA's dominant position is built on this architecture. Non-GPU accelerators are special-purpose chips designed for specific tasks — ASIC, FPGA or customized NPU, which can provide higher energy efficiency ratio in inference scenarios.

Huawei, T-Head, Kunlunxin and Cambricon together account for 95% of the non-GPU market share, and government agencies, Internet companies and telecom operators are adopting such products on a large scale. IDC predicts that China's total AI accelerator shipments will grow by 19% in 2026, while GPU shipments will decline by 9%.

This means that the growth engine of the market is shifting from the GPU track to the non-GPU track where domestic manufacturers have more advantages.

What Huawei Sells Is Not Just a Single Chip

If the rise of non-GPU accelerators is a change in product structure, then Huawei's real differentiation lies in that what it sells is not just a single chip.

In April 2025, Huawei Cloud released the CloudMatrix 384 super node, which connects 384 Ascend cards into a super cloud server through a new high-speed interconnection bus, providing a maximum computing power of 300 Pflops.

The significance of this solution is: the performance of a single Ascend chip may not be as good as NVIDIA's H100 or B200, but through system-level architecture innovation, Huawei can provide competitive computing power at the cluster level. Under the trend that MoE large models become the mainstream, system-level interconnection efficiency and memory bandwidth are often more critical than the peak computing power of a single chip.

This exactly echoes the trend pointed out in the IDC report: the focus of competition is shifting from individual chip performance to system-level capabilities, including the efficiency of processors working collaboratively in large computing clusters.

NVIDIA's product strategy is also moving in this direction — from selling GPUs to selling DGX SuperPOD, from chips to systems. However, in the Chinese market, the availability of NVIDIA's system-level solutions is limited by export control, while Huawei's CloudMatrix can be deployed in any data center in China.

Extension from Chips to Ecosystem

What is more noteworthy is that Huawei is extending the idea of system-level competition from the computing power layer to the entire AI infrastructure layer.

Xu Zhijun, Rotating Chairman of Huawei, previously announced the three-year roadmap for Ascend chips at the Huawei Connect conference, clarifying the rhythm of launching a new generation of chips every year. According to this roadmap, Ascend 950PR will be launched in the first quarter of 2026, Ascend 950DT in the fourth quarter of 2026, and Ascend 960 and Ascend 970 are scheduled for the fourth quarter of 2027 and 2028 respectively. Xu Zhijun also revealed that the Ascend 950 series chips will achieve fundamental improvements in self-developed HBM and other aspects.

According to the technical details announced by Huawei, Ascend 950PR will be equipped with 112GB of self-developed HBM high-bandwidth memory with a bandwidth of 1.4TB/s, which has been greatly improved in memory capacity and interconnection bandwidth compared with the previous generation. This means that Huawei is extending the independent supply chain from the chip manufacturing link to the storage link — and HBM supply is exactly one of the root causes of the 910C production capacity bottleneck.

This set of strategies is similar to NVIDIA's CUDA ecosystem strategy: instead of selling a single product, it builds a complete full technology stack covering interconnection, memory, software and systems around its own chips. NVIDIA locks in developers with CUDA, while Huawei locks in domestic customers with CloudMatrix and the CANN software stack.

The Meaning of 6 Percentage Points

6 percentage points is not a large figure in the scale of global semiconductor competition. But the trend it points to is clear: NVIDIA is transforming from a dominant player to a participant in the Chinese market, while Huawei is transforming from a substitute to a competitor.

This transformation is not brought about by a single technological breakthrough, but the result of the superposition of multiple factors. Export control has continuously compressed NVIDIA's product supply in the Chinese market — from the ban on A100 and H100, to the successive downgrade of special versions such as A800, H800 and H20, to the requirement of applying for a license for H20's export to China in April 2025. Domestic policies have opened up a demand window for local manufacturers, requiring Chinese enterprises to obtain approval before importing NVIDIA chips that meet U.S. restrictions, and this approval threshold itself is changing enterprises' procurement decisions. Huawei's own system-level innovation has turned Ascend from a usable alternative into a competitive choice.

What NVIDIA has lost is not just the figure of 34%. It has lost the pricing space in China's AI chip market, ecosystem influence, and participation in technical standards in one of the world's largest AI application markets. When Jensen Huang said "ceding it to Huawei", there is both helplessness and sobriety in those words.

What is worth observing next is whether Huawei's production capacity ramp-up can keep up with the growth rate of demand, whether NVIDIA's next-generation products can find a compliant path to enter the Chinese market, and whether the substitution of non-GPU accelerators for GPUs is a short-term phenomenon or a long-term trend. Among the 11 Chinese chip manufacturers tracked by IDC, 9 have achieved year-on-year growth. The speed of this market pattern change may be faster than everyone expects.

This article is from the WeChat official account "Tech Business", written by Tech Business, and authorized for release by 36Kr.