Six points is not a lead. It is a rounding distance.
A year ago the arithmetic implied roughly 1.2 million Nvidia accelerators shipped into China in a half (that figure is back-calculated from IDC's 44% decline, not an IDC headline). This half it was 678,000. Huawei, in the same period, went from more than 300,000 units to 572,000. The distance between the American incumbent and its nearest Chinese rival is now 106,000 units in a two-million-unit market.
That is the context for the argument Nvidia's chief executive has made repeatedly in Washington: that export controls hand the Chinese market to Huawei. The IDC figures cannot isolate the effect of US export controls from Beijing's domestic-substitution policies, import approvals or changing customer demand. They do show the resulting market shift: Nvidia's shipments fell while domestic alternatives gained volume.
The category changed, not just the vendor
Read past the top two and a second shift appears. GPUs, the chips Nvidia dominates and the ones most coverage still uses as a synonym for AI hardware, were the minority in China in this half: 991,000 units, down 26%. Non-GPU accelerators, a category that includes chips optimized for specific AI tasks such as inference, rose nearly 80% to 1.016 million units.
That segment is almost entirely domestic. Four Chinese firms account for 95% of it, according to IDC, and its buyers are government agencies, internet companies and telecoms providers. In other words, the part of the Chinese market that is growing fastest is the part where the contest is no longer Nvidia against a rival GPU. It is a different class of chip, bought by institutions with reasons beyond raw performance to buy Chinese.
IDC's own framing points the same way: competition, it says, is moving from single-chip performance toward how well chips work together in a cluster. That is a less comfortable contest for an imported part, however fast it is, because clusters are bought as systems, serviced locally and planned over years.
What the dataset cannot tell you
Two limits belong in the piece, because both were flagged in the reporting itself. IDC did not specify which Nvidia products sit inside the 34%, and it did not say whether top-end chips that are barred from export but known to reach China through illicit channels are counted. If IDC excludes illicitly imported Nvidia accelerators, its shipment figures may understate Nvidia's presence in China. But Reuters could not establish whether those chips were included, and shipments alone cannot measure the installed base.
Unit share is also not revenue share, and it is not performance share. A market counted in boxes treats a high-end training GPU and a narrower inference accelerator as equals. Nvidia could hold a far larger share of the money, and of the serious training capacity, than 34% suggests. The honest reading is narrower: in units shipped through visible channels, in the half to June, Nvidia's position in China went from dominant to contested.
Watch the second half for one number in particular. IDC forecasts total 2026 accelerator shipments rising 19% even as GPU shipments fall 9%. If that divergence persists, non-GPU accelerators could extend their newly established lead in unit volume, although the forecast does not establish their future revenue or computing-capacity share.
Sources
- [1] Reuters (Eduardo Baptista and Che Pan, Beijing, October 13, 2026), reporting IDC figures for H1 2026 China data-centre AI accelerators, read in full via syndicationRead source
- [2] Same Reuters wire, corroborating syndicationRead source