At a provincial intelligent-computing centre in western China, the queue is not for electricity and it is not for floor space. It is for cards. An integrator arrives with a purchase order and a delivery window that has slipped twice, and a facilities manager walks him past racks that are installed, cooled and empty.
The meeting is short. He is offered less than he asked for, at a price above his quotation, with delivery next quarter. He takes it. Nobody treats this as a crisis. It is how the market works now.
That is the right frame for the number that travelled around the world this year: Huawei's Ascend chips took roughly half of China's accelerator market.
Key takeaways
- Bernstein Research, cited by AP, puts Ascend at about 50% of China's AI accelerator market in 2026, up from 40%, with Nvidia down to 8% from 40%.
- Huawei targets 1.6 million Ascend dies in 2026, nearly triple the prior year, while SMIC is estimated to give 15–20% of advanced-node output to Ascend.
- The Council on Foreign Relations estimates the best US chips at about 5× Huawei's best today, widening to roughly 17× by the second half of 2027.
- On aggregate compute, Huawei may reach about 5% of Nvidia's 2026 output in the most favourable scenario; the median estimate is near 1%.
- Ascend 950DT was suggested at about RMB 250,000 (≈ US$35,200 / HK$275,000) in September, up 60% in three months.
A share made by scarcity
Read as a verdict on silicon, the share is misleading. Read as a supply statistic, it becomes ordinary. Nvidia's access to the Chinese market was narrowed by export controls; Huawei could sell whatever it could make. When the alternative is unavailable, half the market is arithmetic.
The upstream numbers tell the same story. SMIC is estimated to devote 15–20% of its advanced-node output to Ascend, with utilisation above 93% (SemiAnalysis). That is a foundry running hot on a narrow product mix, not a fab with spare capacity.
So the question is not why buyers switched, but what they are buying and how much of it exists.
The comparison that actually matters
Fifty per cent measures units sold inside one protected market. The number that matters to anyone modelling compute is aggregate: how much usable arithmetic China's champion produces against Nvidia's global output.
The Council on Foreign Relations puts the best US accelerators at about 5× Huawei's best on total processing performance, widening to roughly 17× by the second half of 2027. In the scenario most favourable to Huawei, the company produces about 5% of Nvidia's aggregate AI compute in 2026. The median estimate is near 1%.
Both figures can be true at once. A company can hold half of a protected market and one per cent of the world's compute. The first number describes procurement. The second describes capability — and the training budget available to a Chinese lab.
The silicon is shaped for a different job
Huawei's parts are memory-heavy and interconnect-heavy relative to raw throughput:
- Ascend 950PR: 2 PFLOPS FP4, 128 GB of proprietary HiBL 1.0 memory, 2.0 TB/s interconnect.
- Ascend 950DT: 144 GB of HiZQ 2.0, roughly 4.0 TB/s.
- Ascend 910C: about 53 billion transistors, two 910B dies in a chiplet package, 128 GB of HBM.
- Ascend 960 is scheduled for 2027.
That profile suits serving more than training the next frontier model. It is a defensible bet that leaves the training gap open.
The price is the tell
Ascend 950DT carried a suggested price of about RMB 250,000 (≈ US$35,200 / HK$275,000) per card in September, up 60% in three months. The 950PR moved from about RMB 60,000 to more than RMB 80,000. The 910C went from about RMB 90,000 to more than RMB 110,000.
Semiconductor prices normally fall as volume rises. These are climbing while Huawei pushes toward 1.6 million dies. What the curve shows is a binding constraint meeting a buyer with no substitute.
The constraint is memory and packaging. CXMT's 2026 output of roughly 2.2 million HBM stacks supports only 250,000–400,000 Ascend-class packages, according to a SemiAnalysis estimate, while SMIC can make die for more than a million chips. Die capacity sits well ahead of what can be assembled and sold. That gap, not the lithography headline, decides whether 50% becomes 60% or stalls.
What would move the line
Three variables, in order of leverage:
- Memory supply: CXMT's stack output, and the yield of Huawei's own memory variants.
- Foundry yield: SMIC's N+2 7nm process runs at around 40%, which sets cost per good die.
- System design: whether interconnect and cluster architecture can keep compensating for per-chip deficits.
None of these is a policy announcement; all are engineering schedules, which is why the share will move slowly in either direction.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the headline share is best understood as a measure of what Chinese buyers were permitted to buy, not of what Chinese silicon can do.
Honest limitations
Nearly every figure here is an estimate rather than a disclosure. The 50% share comes from one research house via a wire service, the aggregate ratios from a single think-tank model with wide scenario bands, and the SMIC allocation and yield numbers are third-party inferences. Huawei publishes performance for its own parts but not shipment volumes or customer mix, and CXMT does not publish stack output at all. The 1.6 million die figure is a target, not an outcome.
What readers can do now
- Operators: specify memory capacity, interconnect and delivery quarter in procurement, and stop negotiating against peak floating-point numbers that describe a workload you do not run.
- Investors: treat memory and packaging capacity as the swing variable in any China compute model; the logic die is no longer the scarce input.
- Researchers: benchmark Ascend on memory-bound serving workloads, where the design is aimed, and report training results separately rather than blending them into one score.