IDC data reviewed by Reuters, reported on October 13, 2026, says Nvidia's share of China's data-center AI chip market fell to 34% in the first half of 2026, while Huawei climbed to 28%. That is a six-point gap, down from a lead that was far larger a year ago. The figures are widely repeated, but the original report is not public, so this post separates what is reported from what we could check.
Short version: the direction (Nvidia down, Chinese vendors up) matches everything else we have tracked this year, including China's push to keep Nvidia purchases under approval. The exact percentages are one analyst firm's estimate, relayed second-hand.
TL;DR: the claim in one table
| Question | Answer |
|---|---|
| Who published the numbers? | IDC, per a Reuters report dated October 13, 2026 |
| Nvidia, H1 2026 | 678,000 accelerators, 34% share, down 44% year over year |
| Huawei, H1 2026 | 572,000 accelerators, 28% share, up from just over 300,000 |
| Others | Alibaba T-Head 7%, Cambricon 5%, Baidu Kunlunxin 4%, Hygon 3% |
| Total market | About 2.008 million accelerators in six months |
| Did we read the IDC report? | No. We relied on outlets citing it |
| Is it consistent with earlier data? | Directionally yes; the exact baselines differ by source |
What exactly is being claimed
According to the summary carried by CXO Digitalpulse, which cites an IDC report reviewed by Reuters, the January to June 2026 picture looks like this:
- Nvidia shipped 678,000 AI accelerators into China, 44% fewer than a year earlier.
- Huawei shipped 572,000, up from a little over 300,000.
- GPU shipments fell 26% to 991,000, while non-GPU accelerators (custom ASICs and NPUs such as Huawei's Ascend line) rose nearly 80% to 1.016 million. For the first time in this series, non-GPU units outnumber GPU units.
- IDC forecasts total accelerator shipments in China to grow 19% in 2026 even as GPU shipments decline 9%.
That last point matters most. The market is not shrinking; the demand is being served by someone else.
What is confirmed and what is not
We could not open the Reuters piece or the IDC report. The same set of numbers appears across several secondary write-ups, which tells us the source is consistent, but it does not tell us the numbers are right. Here is where we landed.
| Item | Status |
|---|---|
| Nvidia 678,000 units and 34% | Reported consistently by multiple outlets citing IDC/Reuters; primary report not seen |
| Huawei 572,000 units and 28% | Same |
| Nvidia down 44% year over year | Same; internally consistent with the unit count |
| Which Nvidia products are counted | Not stated in the summaries we read |
| Cause: export controls plus Beijing's domestic push | Attributed to the report and to background reporting; IDC's own wording not seen |
| Full-year 2025 baseline | Conflicting (see below) |
The 2025 baseline is not settled
When IDC data for full-year 2025 was reported in April 2026, coverage of that earlier report said Chinese chipmakers took about 41% of the market and that Nvidia kept roughly 55% on about 2.2 million cards, with Huawei near 812,000 units. A separate Bernstein projection reported by AP in June 2026 described Nvidia and Huawei as roughly level at around 40% in 2025, with Nvidia falling toward single digits this year and Huawei near half (AP via WTOP).
Those three views do not line up neatly. Different firms define "China AI chip market" differently (servers versus cards, data center versus all accelerators, units versus revenue). So the safe reading is: every credible estimate shows Nvidia losing ground; none of them agree to the point.
Why the numbers are moving
Three forces show up in every account.
- US export rules. Washington restricted shipments of high-end Nvidia parts to China, then later eased some of them. Nvidia has already taken large charges tied to China-specific parts in earlier quarters, as our readers will remember from the H20 saga.
- Beijing's preference for domestic chips. Even when US sales were allowed again, Chinese authorities steered buyers toward local silicon and, per the report summaries, required approvals to import compliant Nvidia parts. We covered one such approval in the ByteDance and Alibaba RTX Pro 5500 story.
- A faster Huawei roadmap. At Huawei Connect on September 17, 2026, deputy chairman David Wang said the company is on a one-generation-a-year cadence for Ascend. The summary we read says the Ascend 960DT is now due in Q1 2027, three quarters ahead of the earlier plan, with the 960PR in Q3 2027, and that Huawei reports more than 1,000 Atlas 900 A3 SuperPoDs deployed. Those deployment figures are the company's own and are not independently verified.
Export controls gate letting a stream of AI chips through only partly, illustrating Nvidia China AI chip limits
How to read a shipment-share number
Shipment share is the easiest statistic to quote and the easiest to misread. A few cautions before anyone declares a winner.
- Units are not value. A million small inference cards and a hundred thousand flagship training parts are very different markets. Nvidia's revenue share in China may be higher or lower than its unit share.
- Units are not capability. Reporting relayed by AP says Huawei's Ascend 950 series is roughly comparable to Nvidia's H200 by some measures, according to analysts. That is a contested comparison, and software maturity (CUDA versus Huawei's stack) is a separate question.
- Smuggled and gray-market chips are invisible. Cases such as the Super Micro contractor guilty plea over AI server smuggling are a reminder that officially counted shipments are not the whole supply.
- Capacity, not just chips. Raw compute also depends on power. Our write-up of the SemiAnalysis China versus US gigawatt comparison shows how much of the picture sits outside the chip count.
What this means for builders and buyers
If you build or buy AI in China, or sell into it, the practical consequences are concrete.
- Model teams in China are increasingly training and serving on Ascend and other domestic parts. That is one reason Chinese labs keep shipping efficient open-weight models; see our China AI playbook on free models and cheap compute.
- Portability matters. Kernels written only for CUDA do not move to Ascend for free. Open-source inference stacks that target several backends are the hedge.
- Chip supply stays political. Executives in the sector face their own pressures; see China's travel restrictions on AI chip executives' families.
- Outside China, the story is demand. Nvidia's global business is limited by supply, not by China. Our AI chip supply chain roundup for October 2026 tracks that side.
Bridge between two landmasses with chips crossing, illustrating Nvidia and Huawei China AI chip supply routes
How you can check this yourself
- Look for the IDC release. IDC usually publishes a press summary or sells the tracker; search IDC's China accelerated server pages for "H1 2026".
- Find the Reuters report. Compare its wording on definitions (data-center AI accelerators versus servers) with the summaries.
- Cross-check Nvidia's own disclosures. Nvidia's next 10-Q discloses China revenue by region; if unit shipments fell 44%, China revenue should reflect it, with the caveat that China-specific products carry lower prices.
- Watch Huawei's claims separately. Deployment counts and roadmap dates come from Huawei; wait for third-party benchmarks.
- Compare methodology across estimates. Bernstein, IDC and others use different scopes; do not mix their percentages in one chart.
Why the non-GPU surge deserves its own paragraph
The most interesting number in the report may not be either percentage. Non-GPU accelerators, the custom chips and NPUs that do not follow the GPU programming model, grew nearly 80% while GPUs shrank 26%. That tells us Chinese buyers are not merely swapping one GPU for another. They are rebuilding serving stacks around different silicon, compilers and interconnects. Once a cloud provider has done that work for inference, switching back is costly even if Nvidia parts become easy to buy again. A policy reversal in Washington would therefore not simply restore the old split; it would fight an installed base that is already adapting. That is also why we treat the six-point gap as a trend line rather than a snapshot: the sign of the change is more informative than any single percentage.
What we expect next
IDC's own forecast, as relayed, is 19% growth in total accelerator shipments in China this year with GPUs down 9%. If that holds, the second half should widen the non-GPU lead unless US export policy changes again. The signals to watch are Nvidia's next earnings commentary on China, any new US licensing decisions, Huawei's Ascend 960 timing, and whether Chinese cloud providers disclose the mix of chips behind their model-serving prices.
We will update this post when IDC or Reuters text becomes available, or if the figures are revised.
FAQ recap
The FAQ above covers the headline questions. The one-line summary: the 34% versus 28% split is the best public estimate we have for the first half of 2026, it is directionally supported by earlier reports, and it is not yet verifiable against the primary document.
Related reading
- China approves Nvidia RTX Pro 5500 for ByteDance and Alibaba
- SemiAnalysis: China 24 GW compute versus US 56 GW
- China AI chip executives' family travel restrictions
- Super Micro contractor pleads guilty in AI server smuggling case
- Tencent and Oracle: 100,000 AI chips in Southeast Asia
- AI chip supply chain, October 2026
- China's AI playbook: free models, cheap compute
Figures are as reported on October 13, 2026 and may be revised when the primary IDC report is available.