The gap between Nvidia and Huawei in China’s AI chip market is now about 106,000 chips. That is the difference between the roughly 678,000 accelerators Nvidia shipped into the country in the first half of 2026 and the 572,000 Huawei shipped, according to research from IDC reported on Tuesday. A year ago, no one would have described that race as close.
The numbers are real, and they matter. But a shipment count tells you who sold hardware last quarter. It does not tell you who controls the software that developers build on, and that is where the more important contest is playing out.
What the IDC data says
IDC’s figures, relayed by Reuters and Benzinga, cover January through June 2026. Nvidia’s shipments of AI data center chips to China fell 44% from the same period a year earlier. Its share of China’s AI accelerator market dropped to 34%.
Huawei took second place with 28%, and it was the biggest gainer among Chinese chip designers. Its shipments rose from more than 300,000 a year earlier to 572,000. Behind it came Alibaba’s chip unit T-Head at 7%, Cambricon at 5%, Baidu’s Kunlunxin at 4% and Hygon at 3%, according to the Reuters summary.
The total market in China reached about 2.008 million accelerators in the half, per the Benzinga report. Inside that total, the mix is shifting. GPUs, the category Nvidia dominates, fell 26% to 991,000 units. Other accelerators, a bucket that includes Huawei’s Ascend line and custom chips from Chinese tech giants, rose nearly 80% to 1.016 million. By IDC’s count, non-GPU accelerators now outnumber GPUs in China.
A quick check of the math supports the headline numbers. Nvidia’s 678,000 chips is about 34% of 2.008 million, and Huawei’s 572,000 is about 28%.
Why Nvidia is losing ground
Both reports point to the same two forces, and they pull in the same direction.
The first is U.S. export controls. Washington continues to bar Nvidia from selling its most advanced AI chips in China, which limits what the company can offer there. The second is Beijing’s own policy. Per Reuters, China has restricted purchases of the smaller pool of chips the U.S. does allow, permitting only limited local sales. Benzinga adds that Beijing requires approval for imports of Nvidia products that comply with U.S. rules.
Meanwhile, Chinese rivals have been closing the performance gap. In September, Huawei introduced more than 10 chipsets spanning AI accelerators, processors and connectivity parts, according to Benzinga. Later that month, the AI lab DeepSeek partnered with Huawei on open-source software for its Ascend chips.
Nvidia’s own chief executive has been blunt about the situation. In May, Jensen Huang called Huawei “very, very strong,” and said Nvidia had “largely conceded” the China market as export controls pushed the domestic ecosystem forward, per Benzinga. In other words, the share loss is not a surprise to the company. It is a trend its leadership has described in public.
What the numbers leave out
IDC’s data is useful, but the coverage published so far has gaps that readers should keep in mind.
First, it is unclear which Nvidia products make up the 34% share. Benzinga notes that IDC did not say whether the figure includes the restricted B200 and B300 chips that have reportedly been smuggled into China. If smuggled hardware is not counted, Nvidia’s real footprint could be larger than the report suggests. That is speculation, and the available reporting does not settle it.
Second, the named shares add up to 81% (Nvidia plus the five Chinese vendors listed in the Reuters summary). The summaries do not account for the remaining 19%, and they do not explain IDC’s methodology. It is also not clear whether the share figures cover only data center chips or all AI chips, even though the shipment figure refers specifically to data center chips.
Third, this is a syndicated summary of a research firm’s report, not the report itself. Benzinga says Nvidia did not immediately respond to its request for comment. Treat the percentages as a solid directional signal and not as audited sales figures.
The real contest is software
Here is the part a chip count cannot capture. Nvidia’s strongest advantage has never been raw silicon alone. It is CUDA, the software platform that most AI developers learned and built on. Switching chips means rewriting code, retraining engineers and re-testing models, which is why customers have stayed with Nvidia even when alternatives looked cheaper.
That is why the DeepSeek and Huawei announcement deserves more attention than its size suggests. Benzinga presents the open-source software work on Ascend chips as a challenge to CUDA. If Chinese developers can build and run models on Ascend as smoothly as they do on Nvidia, the switching cost that protected Nvidia shrinks. Hardware share can be rebuilt over a few quarters. A software ecosystem, once it moves, tends to stay moved.
There is also a practical point for anyone outside China. A large, self-sufficient market with its own chips and its own developer tools becomes a second AI stack. Over time, models, benchmarks and tooling could diverge between the two. That would matter to global companies that need to deploy AI products in both places.
What to watch next
Three signals will tell you more than the next quarter’s share figure.
- Policy changes. Any loosening or tightening of U.S. export rules, or of Beijing’s approval process for Nvidia imports, would shift the picture quickly.
- Software adoption. Watch whether major Chinese AI labs announce training runs on Ascend or other domestic hardware, and whether the open-source tools improve.
- Nvidia’s disclosures. The company’s next earnings report should show how much China revenue it is still counting on, and how it describes the market.
The 106,000-chip gap is small enough that the lead could flip within a year. Whether it does will depend less on who ships more boxes and more on where developers choose to write their code.
Editor’s take
The share number is a scoreboard, and scoreboards lag. I would watch the software stack, not the shipment count. If Chinese developers keep moving real workloads onto Ascend, Nvidia’s position there will not come back even if export rules ease. If they do not, the chip gap says less than it seems to. This is the editor’s judgment call, not a prediction, and the data above cannot settle it either way.
