China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk

TrendForce now projects that domestically designed accelerators will take close to 90% of China's high-end AI chip sales in 2026, up from 45% last year, leaving Nvidia and AMD roughly 10% ^[3]. For anyone shipping AI product into China, that shifts the engineering problem from export licences to a target stack that is not CUDA, and the schedule risk from policy to somebody else's fab, packaging and thermal capacity ^[13][17].

Start with the base. Guancha.cn puts China's 2025 total available market above 4 million accelerator units, with Nvidia at 2.2 million units and a 55% share, and Huawei second at 812,000 units and 20.3% ^[6][7]. Alibaba's T-Head shipped 265,000, AMD 160,000, and Cambricon and Kunlunxin about 116,000 each, with everyone else under 100,000 ^[8]. AMD and Nvidia together supplied roughly 2.36 million units, 59% of unit volume ^[9], which leaves about 1.64 million units for domestic suppliers ^[19].

Now the arithmetic that matters. Holding the market at 4 million units, 90% domestic means 3.6 million domestic units, a 2.2X increase in a year, which is how TrendForce frames it: replacing 1.96 million foreign high-end parts inside twelve months ^[10][20]. But the growth figure the firm publishes for Chinese high-end AI processor shipments in 2026 is "more than 83%" ^[5]. Applied to that 1.64 million base, 83% gets you to about 3.0 million units, or 75% of a flat 4 million market, not 90% ^[21]. The two numbers reconcile only if the served market contracts to around 3.3 million units, or if the 90% is measured against a narrower high-end slice than the 4 million unit count ^[22]. Both readings say the same thing: the binding constraint is supply.

That constraint is concentrated. Cambricon, named alongside Huawei by DigiTimes as the biggest beneficiary of the shift ^[2], shipped about 116,000 units in 2025, roughly 2.9% of the market ^[8][24]. Huawei and Cambricon combined shipped about 928,000 units ^[25]. Being the designated winner is not the same as having the output to absorb a two million unit hole. SMIC, China's most advanced foundry, reported Q2 2026 revenue of $3.005 billion against $2.505 billion in Q1 2026 and $2.209 billion a year earlier ^[18], which is about 20% sequential and 36% annual growth ^[23]. Real, and not 2.2X.

The second-order decision is which stack you port to. TrendForce describes a dual-track model of domestic merchant GPUs plus proprietary ASICs, with Huawei, Cambricon and Biren on one side and in-house parts from Alibaba, Baidu, ByteDance and Tencent on the other ^[14][16]. The hyperscalers prefer their own silicon because it is cheaper than merchant accelerators and tuned to their workloads and data formats ^[15]. For a vendor, that means the Chinese target is not one non-CUDA runtime but several, with the custom ASICs reachable only through the cloud that owns them.

Nvidia's side of the ledger is already thin. Bernstein put its China share at 66% in 2024, 40% in 2025 and on track for 8% in 2026 ^[11], and Jensen Huang described the company's PRC share as "zero" in May after shipping no new accelerators officially in the first half ^[12]. Grey-channel parts still arrive precisely because local buyers remain dependent on CUDA and high-end hardware ^[13], which is the dependency this transition is meant to end.

Watch the forecast revisions as closely as the shipments: TrendForce's December 2025 outlook had domestic parts at about 50% of the 2026 high-end market, and the new number is nearly double that ^[4]. A forecast that moves that far in a few months is tracking policy signals more than order books. Watch SMIC's next two quarters and any disclosure on advanced packaging output, and watch whether hyperscaler ASIC programmes crowd out merchant supply from Cambricon and Biren rather than adding to it ^[14][16].