Nvidia now holds 34% of China's AI accelerator market, down from earlier levels, as Huawei expands its share on the mainland, according to market data from IDC reported this week.

The figure marks another contraction for the U.S. chipmaker in a geography that once generated a meaningful slice of its data-center revenue. IDC attributes the shift to a combination of U.S. export controls restricting the sale of Nvidia's most advanced accelerators to Chinese customers and a deliberate push by domestic suppliers, led by Huawei, to fill the resulting gap.

What does the 34% figure actually measure?

IDC's tracking covers the discrete AI accelerator market in mainland China, the segment served by GPUs, custom training chips, and inference processors sold to cloud operators, telecom carriers, and large enterprises. The 34% share is a snapshot of that installed-and-shipping market, not a forecast and not a measure of computing throughput.

The number reflects units and revenue booked into China. Nvidia's overall global data-center revenue continues to rise, but the geography of those sales is tilting away from the mainland as regulators on both sides redraw the rules of engagement.

Why is Huawei gaining?

Three mechanisms are working in Huawei's favor. First, U.S. Commerce Department restrictions block Nvidia from shipping its top-end H100 and H200 parts to China without a license, and a tailored H20 variant has faced its own licensing uncertainty. Second, Huawei's Ascend series of training and inference accelerators has reached volume production at SMIC, the country's leading foundry, using mature process nodes that fall outside the tightest controls. Third, large Chinese cloud customers, including Alibaba, Tencent, and Baidu, have publicly diversified their accelerator sourcing to reduce single-vendor exposure.

The result is a market that still wants Nvidia silicon for the most demanding training workloads, but increasingly accepts domestic alternatives for inference and for capacity expansion.

How much ground has Huawei actually taken?

IDC has not published a direct head-to-head share number for Huawei in the latest reading, but the firm characterizes the Chinese vendor as the principal share-gainer in the period. Analysts covering the China AI supply chain have pointed to Ascend 910B and 910C shipments as the bulk of that volume, with deployments at state-linked data-center operators driving the bulk of demand.

Nvidia's 34% reading is best understood as a directional data point rather than a precise market cap. The company retains a technological lead in raw floating-point throughput and software maturity, factors that have so far preserved its position in frontier-model training inside China. Where it is losing ground is in the long tail of inference deployments, where customers are price-sensitive and where domestic supply is now adequate.

What changes for Nvidia's revenue mix?

Nvidia reports revenue by geography but does not break out mainland China separately. The company has guided that data-center revenue from China-based customers will remain a low single-digit percentage of total data-center sales in coming quarters, reflecting both licensing outcomes and the substitution underway.

The commercial picture is therefore one of a shrinking, not vanishing, China footprint. Each percentage point of share lost to a domestic rival is a data-center dollar that does not flow back to Santa Clara, and the cumulative effect across several quarters is material.

What to watch next

Two near-term data points will determine whether the slide stabilizes or continues. First, the licensing path for Nvidia's next China-specific accelerator, expected to sit below the H20 in performance, will set the floor under its addressable market. Second, the pace of Ascend volume at SMIC through the second half of the year will indicate whether Huawei can convert share gains into sustained revenue.

For now, Nvidia retains the most capable silicon sold legally into China. The question is how long that advantage translates into bookings before IDC's next reading.