China's AI Chip Market Share Breaks 50%, but Nvidia Still Dominates Frontier Training
0xBroomberg
Goldman Sachs projects domestic AI chips will exceed 50% of China's 2026 shipments, but Nvidia still leads frontier training by roughly one generation — market share gains do not equal technological parity, and that gap defines the real contest.
"Over 50%" — how is that number actually calculated?
Goldman's framework: Nvidia still holds 55% of China's 2026 AI chip shipments, Huawei Ascend takes 20%, Alibaba's T-Head 7%. This means → by Goldman's definition, "domestic majority" has not arrived yet; Nvidia remains the single largest supplier.
Bernstein and IDC use a different basket: domestic vendors will account for roughly 56% of AI accelerator server shipments, up from 41%–46% in 2025. In plain terms = one counts chips, the other counts whole servers — naturally the conclusions differ.
Bernstein separately cites *The Economist*, putting Huawei at 50% and Nvidia's export-compliant share at just 8%. This reflects a basic rule: same market, different definitions, and the numbers can diverge by multiples. Always check the methodology before reading the headline.
Who is selling, and how much? — vendor revenue breakdown
Bernstein estimates 2026 annual accelerator revenue: Huawei leads at $12.1 billion, followed by AMD at $3 billion, Cambricon $2.1 billion, and Hygon $2 billion.
Alibaba's T-Head brings in $1.2 billion, Baidu $700 million — materially smaller. This means → Huawei alone roughly equals the other five combined; even within the domestic camp, the structure is "one giant, many followers."
On procurement mix, internet companies contribute 63% of Huawei Ascend's 2026 shipments and 60% of Cambricon's; Hygon draws about 35% from the public sector. In plain terms = cloud operators are the biggest buyers of domestic chips, government procurement is the second engine.
What did Z.ai prove? — domestic chips at scale for the first time
Zhipu AI's Z.ai has partially activated a 1 GW data center, running its GLM platform on multiple clusters each using over 10,000 domestic chips. This means → domestic silicon has moved from "it works in the lab" to "it runs at scale" — a demonstrated fact, not a roadmap.
After the announcement, Z.ai's stock surged as much as 41% intraday, closing up 36.9% at HK$1,219; Hua Hong Semiconductor and Montage Technology each rose more than 17%, MiniMax gained over 15%.
China plans to invest roughly RMB 2 trillion (~$295 billion) in data centers over five years. Deutsche Bank projects chip shipments rising from 4 million units in 2025 to 5 million in 2026; JPMorgan expects domestic AI chip shipments to grow from 1 million in 2025 to 5 million by 2028.
How far has Huawei caught up? — inference gap narrowing, training still one generation behind
Goldman estimates daily token output: Ascend 910B produces roughly 400 million, 910C doubles that to 800 million, while Nvidia's H800 reaches 2.6 billion. In plain terms = running for the same 24 hours, Nvidia generates more than three times the "AI text volume" of Huawei's latest chip.
Against Nvidia's export-compliant H20, Huawei's 950PR is more competitive: 1.56 PFLOPS at FP4 vs. H20's 0.56 PFLOPS, with 112 GB high-bandwidth memory and 2 TB/s interconnect. This means → against the Nvidia chips actually allowed into China, Huawei already wins on paper.
But Nvidia's global flagships are far ahead: H100 hits 3.96 PFLOPS at FP8, B200 reaches 18 PFLOPS at FP4 with 8 TB/s memory bandwidth. This reflects the real benchmark: Huawei's true competitor is the B200, not the H20 — and the gap is still more than one full generation.
How much more expensive, how much more power? — the cost disadvantage
Goldman estimates infrastructure cost: Nvidia at roughly RMB 24 million per EFLOPS, Huawei at RMB 51 million, some Chinese super-node systems exceeding RMB 100 million. In plain terms = for the same compute, Huawei costs more than twice Nvidia's price, and certain domestic solutions cost four times as much.
Huawei's CM384 system is priced at approximately RMB 135 million and draws about 800 kW. The dense design cuts server count but raises cooling and infrastructure requirements.
This means → domestic chips currently compete on availability, not cost-efficiency — they can be procured, but running them is significantly more expensive.
Where is the last wall? — software ecosystem and frontier training
Huawei says its CANN Next software stack has achieved roughly 80% CUDA compatibility, but migration and optimization costs remain. This means → developers can "mostly run their code," but switching from the Nvidia ecosystem is not free.
Inference workloads are expected to account for over 50% of future AI data center demand, with training at 20%–30%. This trend favors Huawei, T-Head, and Cambricon, because inference cares less about peak single-chip performance and more about system-level efficiency.
Yet frontier training still points to Nvidia: some Chinese model developers reportedly source Nvidia chips through unofficial channels, with B300-class servers quoted at roughly $1 million. Put simply = paying $1 million through gray-market channels is itself proof that domestic solutions cannot yet replace Nvidia at the highest end of training — and whether they eventually can is the key test for domestic chips' long-term valuation thesis.
Content is for reference only, not financial advice.