TAIPEI, 17 AUG 2026 — TrendForce expects domestically developed chips to take close to 90 per cent of China's high-end AI chip market this year, leaving Nvidia, AMD and every other overseas supplier to share the remainder. The research firm describes the shift as a dual-track move to domestic GPUs and proprietary ASICs rather than a single company displacing another.

This is a forecast from a market research firm, not a disclosure by anyone selling chips. What makes it worth reading is that the companies underneath it have just filed results, and those results are not forecasts.

The forecast

≈90%Domestic share of China's high-end AI chip market in 2026
≈10%Left to Nvidia, AMD and other overseas suppliers
+108%Cambricon first-half revenue growth, to RMB 5.996bn
+147%Moore Threads first-half revenue growth, to RMB 1.736bn

TrendForce published the assessment on 10 August. It attributes the shift to three things: sustained government promotion of domestically produced AI chips, policy support weighted toward the domestic firms with the most room to grow, and a supply chain that has matured in advanced process nodes, advanced packaging and thermal management.

The last point — a maturing supply chain for packaging and cooling — is the one that carries the weight of the argument.

Why packaging and cooling decide this

A high-end AI accelerator is more than a piece of silicon. It is a die, high-bandwidth memory, and an advanced package, all engineered to stay cool in a dense rack.

Designing a competitive chip has never been the hard part of Chinese AI silicon. Manufacturing it at volume has been, and packaging capacity in particular has been the industry's binding constraint everywhere, not only in China. We have written about that repeatedly this month: Intel raised US$20 billion aimed largely at advanced packaging, and SMIC is raising prices because it cannot add capacity fast enough.

TrendForce's argument is that this constraint has eased inside China enough to matter. If that is right, the market share number is a consequence rather than a cause, and it is the packaging and thermal line in the analysis that should be tracked rather than the headline percentage.

The companies underneath the forecast

Three sets of first-half results published this month give the forecast something firmer to stand on.

Cambricon reported first-half revenue of RMB 5.996 billion, up 108 per cent year on year, with net profit of RMB 2.311 billion, up 123 per cent. Almost all of that revenue — RMB 5.994 billion of it — came from cloud products. Research and development spending rose 30 per cent, which is a slower rate than revenue and worth watching in a company scaling this fast.

Moore Threads reported revenue of RMB 1.736 billion for the half, up 147 per cent, against RMB 1.506 billion for the whole of 2025. Its net loss narrowed to RMB 11.56 million, effectively to break-even, and it plans to issue H-shares and seek a Hong Kong listing.

Hygon has guided to first-half revenue of RMB 8.5 billion to 9.3 billion, growth of between 56 and 70 per cent, with net profit of RMB 1.7 billion to 1.83 billion.

Three companies growing at those rates simultaneously is not a single national champion being propped up. It is a segment.

Dual-track matters more than the percentage

The most significant detail for buyers is the description of the shift: a move to both domestic GPUs and proprietary ASICs.

A GPU is a general-purpose accelerator that runs whatever a developer writes for it. An ASIC is built for a narrower set of operations and is faster and cheaper within that range, and useless outside it. A market moving to both at once is a market where the largest buyers are confident enough about their own workloads to commit silicon to them.

That confidence is the real barrier to reversal. Export policy can change quickly. A cloud provider that has designed its inference fleet around a specific accelerator, and rewritten its software to match, does not switch back because a rule changed.

What this means for buyers in this region

For buyers in ASEAN, the consequence is a bifurcating software stack, driven by procurement rather than policy.

Chinese cloud providers and equipment vendors sell aggressively into Southeast Asia, and the infrastructure they offer will increasingly run on domestic silicon. That silicon does not run CUDA. Code, tooling and staff skills built for one stack do not transfer to the other without work, and the cost of that work lands on whoever chose the platform.

An enterprise in the region signing a multi-year AI infrastructure agreement with a Chinese provider is making a technical decision with a long tail. It is not necessarily the wrong decision — price and availability are real advantages, particularly while Western accelerators are supply-constrained — but it should be made with the portability question answered rather than deferred.

The related question for regional governments is what a bifurcated market does to sovereignty arguments. Buying compute from a Chinese provider reduces exposure to United States export policy and increases exposure to Chinese policy. Neither is neutrality, and the two are frequently discussed as though one of them were.

What we could not establish

The underlying TrendForce report is subscription-only, and we have read the publicly available summary rather than the full analysis. We could not establish how "high-end" is defined for the purposes of the 90 per cent figure, which is the definition the whole number rests on, nor the prior-year comparison on the same basis, nor the unit and revenue split behind it.

Two figures attributed to the same firm do not sit together neatly. The high-end AI chip figure leaves overseas suppliers about 10 per cent, while reporting of a separate TrendForce forecast puts Nvidia and AMD's combined share of China's AI server chip market at 21 per cent in 2026, down from 34 per cent. AI server chips and high-end AI chips are different populations, so these are not necessarily in conflict, but anyone quoting a single number should say which one they mean.

We also could not establish how much domestic volume is bought on merit versus procurement direction, the performance gap against current Western parts, whether packaging capacity is merely improved or sufficient for forecast volumes, and the extent of Cambricon's reliance on state-linked customers for its cloud revenue.

What to watch

Moore Threads' Hong Kong listing would be the most informative event here. A prospectus requires disclosure that a mainland listing does not always surface, including customer concentration and supply arrangements, and it would be the first properly documented look inside one of these companies.

Next, watch the packaging figures rather than the chip figures. Domestic advanced packaging capacity is the constraint that determines whether these growth rates continue, and it is reported far less often than market share.

Finally, watch what Chinese vendors bring to this region. When domestic accelerators appear in AI infrastructure proposals in Jakarta, Kuala Lumpur and Hanoi, the bifurcation is no longer a domestic Chinese story but a regional procurement question.