China's domestic AI chips in 2026: Huawei, Cambricon, SMIC and limits

China's home-grown AI chips are real products with growing sales, led by Huawei's Ascend line, Cambricon and the foundry SMIC. They are also held back by manufacturing limits that Huawei itself acknowledges. Most performance figures below are company claims or press estimates, and each is labelled as such.

This post explains how an AI chip is actually made, who does each step in China, what the main firms have published, and where the evidence is thin. It draws on Huawei's own announcements, SMIC's results releases for 2025 and the second quarter of 2026, and press reporting on Cambricon, with a date on every figure. It is not investment advice and does not forecast.

Key numbers at a glance

• 18 September 2025 (Huawei keynote): Huawei published a roadmap with Ascend 950 in 2026, Ascend 960 in Q4 2027 and Ascend 970 planned for Q4 2028. These are company plans.
• 384 (Huawei, March 2025 launch): maximum Ascend 910C chips in the Atlas 900 A3 SuperPoD, claimed at up to 300 PFLOPS.
• 8,192 (Huawei claim, planned Q4 2026): Ascend 950DT chips in an Atlas 950 SuperPoD.
• 15,488 (Huawei claim, planned Q4 2027): Ascend 960 chips in an Atlas 960 SuperPoD.
• US$9.33 billion (SMIC, results of 26 March 2026): 2025 revenue, up 16.2 percent from US$8.03 billion in 2024.
• 93.5 percent (SMIC, 2025): annual capacity utilization, up 8 percentage points.
• 21.0 percent (SMIC, 2025): gross margin, up from 18.0 percent in 2024.
• US$3,005.6 million (SMIC, results of 13 August 2026): second quarter 2026 revenue, up 20.0 percent on the first quarter, with a gross margin of 25.3 percent.
• 90.2 percent (SMIC, second quarter 2026): share of revenue from customers in China, up from 84.1 percent a year earlier.
• 6.5 billion yuan (press report, 30 April 2026): Cambricon's 2025 revenue, with net profit of 2.06 billion yuan, its first annual profit.
• 2.89 billion yuan (same report): Cambricon's Q1 2026 revenue, up 160 percent year on year.
• 1.65 million (IDC via Reuters via a press article, 30 April 2026): AI chips reportedly delivered by Chinese makers in 2025, a 41 percent share of China's AI server market.
• About 812,000 (same report): units attributed to Huawei in 2025.

How an AI chip gets made, in five layers

It helps to separate the layers, because China's position differs in each.

Design. A company writes the blueprint for the chip. Huawei, Cambricon and others do this in China.

Manufacturing. A foundry turns the blueprint into silicon. SMIC is the largest of the Chinese foundries discussed here. The "node" is shorthand for the generation of manufacturing process. Smaller node numbers generally mean more transistors in the same area and better efficiency, and they are harder to make.

Memory. AI chips depend on high-bandwidth memory, or HBM, stacked next to the processor. Without fast memory the processor waits for data.

Interconnect. One chip is never enough for a large model. Systems link thousands of chips so they behave like one computer. Huawei calls its building block a SuperPoD, meaning a large group of chips joined by a fast private network, and a SuperCluster is a group of SuperPoDs.

Software. The libraries and tools that let AI developers run their models on the chip. A chip with weak software is hard to adopt, however good the silicon.

Huawei: the roadmap and the claims

At Huawei Connect in Shanghai on 18 September 2025, Huawei's rotating chairman Eric Xu set out the Ascend plan. Every figure in this section is a Huawei statement or projection. None of the sources cited here contains independent measurements.

Huawei's plan, as published on its website:

• Ascend 950 series, built on a shared die in two variants. The 950PR targets inference "prefill" (reading a prompt) and recommendation workloads, with a low-cost in-house memory Huawei calls HiBL 1.0, planned for Q1 2026. The 950DT targets "decode" (writing the answer) and training, with a higher-bandwidth in-house memory called HiZQ 2.0 offering 144 GB and 4 TB/s of memory bandwidth, planned for Q4 2026. Huawei claims 1 PFLOPS in FP8 and 2 PFLOPS in FP4 for the 950, and an interconnect of 2 TB/s, which it says is 2.5 times that of the Ascend 910C.
• Ascend 960, planned for Q4 2027, roughly doubling the 950's compute, memory bandwidth, memory capacity and interconnect.
• Ascend 970, still being specified, with a target of doubling FP4 and FP8 compute again over the 960, planned for Q4 2028.
• A stated cadence of a new chip about once a year, with compute doubling each time.

A quick glossary. A PFLOPS is a quadrillion floating point operations per second. FP8 and FP4 are number formats that use 8 and 4 bits, which let chips do more arithmetic per second at lower precision. An EFLOPS is a thousand PFLOPS.

On systems, Huawei said its Atlas 900 A3 SuperPoD, launched in March 2025, holds up to 384 Ascend 910C chips and delivers up to 300 PFLOPS. Huawei said more than 300 had been deployed for more than 20 customers. For the future, it announced the Atlas 950 SuperPoD (up to 8,192 chips, 160 cabinets, about 1,000 square metres, 8 EFLOPS FP8, planned Q4 2026) and the Atlas 960 SuperPoD (up to 15,488 chips, 220 cabinets, about 2,200 square metres, planned Q4 2027). It also described Atlas 950 SuperClusters of 64 SuperPoDs, over 520,000 chips and a claimed 524 EFLOPS in FP8.

Huawei compared these systems with a named Nvidia product planned for the second half of 2026 and claimed large advantages in chip count, compute, memory and interconnect. Those are projections about unreleased systems on both sides, so the multiples are not repeated here. A system built from many more, individually weaker chips can win on total size while losing on efficiency, space and power. The comparison depends on what is measured.

On 2 March 2026 at MWC Barcelona, Huawei described the Atlas 950 SuperPoD again, saying it integrates 64 chips per cabinet and scales to 8,192, and showed a general-purpose TaiShan 950 SuperPoD. The release does not name customers or give deployment dates.

What Huawei admits

The most useful sentence in the keynote is a candid one. Xu said that the Chinese mainland will trail in advanced semiconductor process nodes for a long time, and that Huawei lacks access to advanced nodes. The company's stated answer is to combine large numbers of chips, which is what the SuperPoD strategy is. He also acknowledged gaps with Nvidia in individual chips and software ecosystems. Both statements come from Huawei's own materials.

The official sources cited here do not confirm that the 950PR actually shipped in Q1 2026, or that the 950DT and Atlas 950 SuperPoD are on schedule for Q4 2026. Treat those dates as plans.

SMIC: the factory side

SMIC, the Semiconductor Manufacturing International Corporation, published its 2025 results on 26 March 2026. According to its release:

• Revenue of US$9,327 million, up 16.2 percent from US$8,030 million.
• Gross margin of 21.0 percent, up from 18.0 percent.
• Profit attributable to owners of US$685 million, up 39.0 percent.
• Capacity utilization of 93.5 percent, up 8 points.
• Monthly capacity above 1 million 8-inch equivalent wafers.
• R&D of US$774 million, 8.3 percent of revenue.
• The company calls itself the second-largest pure-play foundry worldwide, which is its own statement and depends on the definition.

These are company-wide figures. SMIC makes many kinds of chips, mostly for phones, cars, consumer devices and industrial products. The release does not break out AI chips or name customers, so you cannot read SMIC's total as a measure of AI output. The 8-inch equivalent is a unit used to compare wafers of different sizes.

SMIC in 2026: the quarterly picture

SMIC's release for the second quarter of 2026, dated 13 August 2026, shows a sharp step up in the first half of the year:

• Revenue of US$3,005.6 million, up 20.0 percent from US$2,505.5 million in the first quarter and from US$2,209.1 million a year earlier. SMIC attributes the growth to higher wafer shipments, a higher average selling price and a change in product mix.
• Gross margin of 25.3 percent, compared with 20.1 percent in the first quarter and 20.4 percent a year earlier.
• Monthly capacity of 1,096,500 standard logic 8-inch equivalent wafers, up from 1,078,250, and a utilization rate of 93.7 percent.
• Capital expenditure of US$1,835.7 million in the quarter, up from US$1,562.8 million.
• Guidance for the third quarter of 2026 of revenue up 2 to 4 percent on the quarter and a gross margin of 26 to 28 percent. Guidance is a forecast, not a result.

Two details matter for the AI chip story. First, customers in China accounted for 90.2 percent of revenue in the quarter, up from 84.1 percent a year earlier, while the share from America fell from 12.9 percent to 8.2 percent. That is consistent with Chinese chip designers relying more on a domestic foundry, though the release does not say which customers or products drove the change. Second, management wrote that the industrial momentum and spillover effects generated by AI would persist into the second half and drive broad demand for chip manufacturing. That is a statement about demand, not a disclosure of AI chip volumes.

The application split still points away from AI accelerators as the main business. Consumer electronics made up 44.2 percent of wafer revenue, smartphones 16.9 percent, industrial and automotive 16.5 percent, and computers and tablets 15.6 percent. SMIC has no separate category for data centre or AI chips, so any estimate of its AI output comes from outside the company.

Cambricon and the wider market

Cambricon designs AI accelerators. According to a Yahoo Finance article of 30 April 2026, which relays company results and Reuters reporting:

• 2025 revenue was 6.5 billion yuan and net profit 2.06 billion yuan, its first annual profit since its founding in 2016.
• Q1 2026 revenue was 2.89 billion yuan, about US$423 million, up 160 percent, with net profit up 185 percent to 1 billion yuan.
• Customers are concentrated. The article says the top five clients made up 94 percent of revenue in the first half of 2025, and that the largest, reportedly ByteDance, accounted for about 80 percent. The word "reportedly" is the article's.
• A second listed designer, MetaX, reported Q1 2026 revenue of 561.9 million yuan, up 75 percent, with a net loss of 98.8 million yuan.
• The article cites IDC figures, via Reuters, that Chinese makers delivered 1.65 million AI chips in 2025, with Huawei at about 812,000, and that Cambricon is one of two firms on a government approved AI hardware procurement list alongside Huawei.
• It says Cambricon targets 500,000 accelerator shipments in 2026, up from an estimated 116,000 in 2025, and that this depends on SMIC's 7 nanometer class yields, which have been reported at around 20 percent. "Yield" means the share of chips on a wafer that work. That yield figure is a press report, not a company disclosure, and it may not be current.
• It reports an unnamed source saying Nvidia's share of China fell from a pre-control peak of 95 percent to under 60 percent, and a bank analyst projecting a fall to about 8 percent in coming years. Those are an unnamed estimate and an analyst projection, and no official source confirms either.

Two cautions apply. Press figures for Chinese chip makers sometimes contain unit errors, so this post uses only figures a source states plainly and flags the rest. And the article itself notes that domestic processors remain multiple generations behind Nvidia's current top architecture in raw performance.

Why the export rules matter here

The US controls described in the companion explainer on export controls as the official texts state them restrict both the chips that can go to China and the equipment used to make advanced chips there. Huawei's statement about lacking advanced nodes is the other side of the same story. If a country cannot buy the newest chipmaking tools, it builds bigger systems from older chips, pays more in power and floor space, and leans harder on memory and networking.

That is the logic of Huawei's published strategy. It also explains why Huawei is spending effort on its own HBM and its own interconnect standard, UnifiedBus, which it says it released openly to invite partners. Whether this closes the gap is an open question that none of the sources cited here settles.

What is firm and what is uncertain

Firm, because a company or regulator says so in its own materials:

• Huawei's published roadmap, dates and claimed specifications, as plans.
• SMIC's 2025 and second quarter 2026 results, including capacity, utilization and the regional revenue split.
• Huawei's admission that it lacks access to advanced nodes.

Uncertain:

• Whether Huawei's 2026 and later products ship on time and perform as claimed.
• Independent benchmarks of any Ascend chip. None appear in the sources cited here.
• SMIC's yield on advanced nodes and how much of its capacity serves AI chips.
• Market share numbers, which come from IDC via Reuters via a secondary article.
• Cambricon's full-year audited figures, which are taken here from a press report rather than the filing itself.

What to watch next

• Huawei Connect, held in the autumn, where roadmap updates usually appear.
• Evidence of Atlas 950 SuperPoD deployments with named customers.
• SMIC's third quarter 2026 results against its guidance of 2 to 4 percent revenue growth and a 26 to 28 percent gross margin, and any statement on advanced capacity.
• US rule changes. The timeline of the official texts sets out what each rule says.
• How domestic chips are used by Chinese model makers. See Chinese AI labs and their releases.
• The global picture in AI chip market share.

If your own team ships software that depends on open source libraries, you can check what is inside it with the free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• Huawei, Connect 2025 keynote by Eric Xu, 18 September 2025
• RCR Wireless, 22 September 2025: Huawei outlines roadmap for Ascend AI chips
• Huawei, MWC Barcelona 2026 announcement, 2 March 2026
• SMIC, 2025 annual results release, 26 March 2026
• SMIC, unaudited results for the three months ended 30 June 2026, HKEX filing, 13 August 2026
• Yahoo Finance, 30 April 2026: Chinese GPU maker Cambricon's Q1 revenue

Last reviewed: 19 October 2026.