Six Points of Room

Nvidia’s share of China’s data-center AI accelerator market fell to 34% in the first half of 2026, according to IDC data reported by Reuters, with shipments down 44% year over year to 678,000 units. Huawei shipped 572,000 — up from a little over 300,000 a year earlier — and took 28% of the market. Six percentage points now separate the company the world’s data centers were built around from the company Washington spent three years trying to slow down.

Read the shape of the market, not just the two leaders. China’s total AI accelerator shipments reached 2.008 million units in the period, and non-GPU accelerators rose nearly 80% to 1.016 million — now more than half of everything shipped. Huawei, T-Head, Kunlunxin and Cambricon account for 95% of that non-GPU segment. The restriction did not stop the orders from being filled; it changed which shelf they came from, and the shelf that grew was the domestic one.

What the Controls Actually Moved

Three years of export licensing were meant to hold back the top of the stack, and by volume they did move it: Nvidia’s unit shipments into China are down almost half. What the licensing could not do was hold the market share, because 34% of a market you are allowed to serve is a very different asset from the near-total position the same chips started from. Huawei’s gain is the mirror image of that decline, and at six points of separation the two lines are close enough that a single procurement quarter could cross them.

The export control was written to slow an industry down. What it produced was a second one.

The enforcement story runs alongside the market story and reads less cleanly. A contractor linked to Super Micro has pleaded guilty to helping divert roughly $2.5 billion in AI technology, Nvidia-equipped servers included, toward China; a co-founder remains charged and has pleaded not guilty. Senate Democrats are pushing tougher legislation on advanced chip and equipment sales ahead of a prospective Trump–Xi meeting. And in the other direction, Washington has eased access for Emirati entities including G42 and Core42, which can now take advanced chips from Nvidia, AMD and Cerebras without individual licences.

The Same Chip, Two Doors

That last part is the actual policy, and it is narrower than either party’s rhetoric. Supply into one country is being restricted while supply into another is being streamlined, both under the same stated concern about where compute ends up. A 25% tariff now applies to selected processors including Nvidia’s H200 and AMD’s MI325X, while chips imported for US data centres are exempt. The controls have become a routing system rather than a wall — and the only number in this report that describes outcomes rather than intent is the 34%, still falling, in a market Nvidia used to own.