Nvidia's share of China's AI chip market fell to 34 percent in the first half of 2026, IDC data show

Nvidia's share of China's AI chip market fell to 34 percent in the first half of 2026, according to IDC figures that Reuters reported on October 13, 2026. The company shipped 678,000 AI data center chips into China in the six months to June, down 44 percent from a year earlier. For anyone asking how much of China's AI chip market Nvidia still holds, that is the newest number, and it comes with caveats about what it measures.

The caveats matter because this story is easy to over-read. Losing market share inside China is not the same as China catching up in compute.

What the numbers say

Reuters, as republished by Investing.com, cites IDC as the only research source. It reports that Nvidia's share of Chinese AI chips fell "amid U.S. export controls and advancements by Huawei." The main driver was American restrictions on Nvidia's most advanced parts. But Beijing played a part too: in the report's words, "Beijing also largely balked at the small pool of chips that Washington did approve."

Among Chinese developers, the report lists Huawei with 28 percent, the largest share. Behind it are T-Head, Alibaba's chip design unit, at 7 percent, Cambricon at 5 percent, Kunlunxin, Baidu's unit, at 4 percent and Hygon at 3 percent. The article says these rivals are also making more advanced chips but gives no Huawei shipment or revenue figures.

It does not say whether the shares are measured in units or in revenue, or give the report's date beyond the first half of 2026. Treat the percentages as indicative until IDC publishes its own note, which we could not find publicly.

The comparison with last year

IDC's earlier full year 2025 numbers, reported by Reuters in April 2026, put the China market at roughly 4 million AI accelerator cards. Nvidia shipped about 2.2 million for a 55 percent share. Chinese vendors together shipped about 1.65 million, or 41 percent, with Huawei at about 812,000 and Alibaba's T-Head at about 265,000. AMD shipped about 160,000 for 4 percent.

Set the two reports side by side and the direction is clear: from 55 percent of a full year to 34 percent of a half year. Be careful with the arithmetic, though. A half year against a full year, and possibly a different counting basis, means the 21 point gap is a rough guide, not a measured change. The April article also noted that local governments building AI computing centers reportedly often carried implicit "buy Chinese" directives, which would help explain domestic gains independent of chip quality.

Why share is not the same as compute

Epoch AI's report "Will Huawei catch up to Nvidia by 2030?", updated September 24, 2026, puts the gap in physical terms. Epoch estimates Huawei produced about 800,000 Ascend units in 2025, against about 4.5 million Nvidia units sold. For 2026 it projects about 1.5 million Huawei units against 5.9 million Nvidia units. In H100-equivalent compute, Huawei's 2026 output is about 0.88 million against roughly 23 million for Nvidia, under 4 percent of the total.

The reason is per-chip performance as well as volume. Epoch says the Ascend 950DT's FP8 performance is about half that of the H100, a chip that began shipping in 2022. And the report's headline conclusion is that Huawei "currently trails Nvidia in both per-chip performance and AI chip production volume."

Epoch models several memory scenarios. If Huawei relies on Chinese-made high bandwidth memory only, its compute output stays around 1 percent of Nvidia's through 2028. Even with ten times as much smuggled memory, it reaches roughly 16 percent in 2028, which Epoch says would not close the gap. Its central 2030 estimate of about 24 million H100-equivalents is roughly what Nvidia produces in 2026, which Epoch reads as about four years behind on volume.

How the pieces fit

Read together, the two sources describe a market in which Chinese buyers are being steered toward domestic chips by both governments, while the domestic chips are still far behind in aggregate capability. The tension has a visible side effect that we have followed elsewhere: the legal channel shrinks while enforcement cases over diverted Nvidia hardware keep appearing. Our tracker of chip smuggling cases logs the prosecutions that follow, and Epoch's separate finding on China and the US supply chain exposure shows how deep the dependence runs.

It also reframes the export control argument. If Chinese labs get fewer Nvidia chips and the domestic substitutes are slower and scarcer, the policy is doing something real to China's training capacity. If Beijing's own restrictions are suppressing Nvidia purchases that Washington allowed, then loosening US approvals would not automatically restore Nvidia's share. The Reuters report points to both forces and does not weigh them.

Our take

We would read 34 percent as a measure of policy on both sides of the Pacific, not a measure of technology. The figure that tells you who is ahead in the race for compute is Epoch's: roughly 25 times less AI compute from Huawei in 2026, with a gap that narrows only if memory supply improves faster than anyone has shown. The things to watch are IDC's own publication of the first half data, Huawei's high bandwidth memory output, and whether any further US approvals for Nvidia parts change Beijing's stance. Until one of them moves, share loss in China is a footnote to the compute gap, not a reversal of it.