Market share
Huawei Shipped 572,000 AI Accelerators in China in the First Half of 2026
That is 28% of China's data center accelerators by units, against NVIDIA's 34%, and the gap is closing from both ends.
| Vendor | Share of units | Units | Logic foundry |
|---|---|---|---|
| NVIDIA | 34% | 678,000, down 44% year on year | TSMC |
| Huawei | 28% | 572,000, up from more than 300,000 | SMIC |
| Alibaba T-Head | 7% | Not reported | Not disclosed |
| Cambricon | 5% | Not reported | SMIC (reported) |
| Kunlunxin (Baidu) | 4% | Not reported | Not disclosed |
| Hygon | 3% | Not reported | Not disclosed |
| All others | 19% | Not reported | Mixed |
IDC tracked 11 Chinese suppliers, and nine of them shipped more than a year earlier. The vendors IDC names add to 47% of units, against 41% for Chinese chips across 2025. NVIDIA's units fell because US licensing limits what it can sell in China, not because demand fell: Reuters reported in January 2026 that Chinese orders for the H200 passed two million.
Huawei sells systems as well as chips. The Atlas 900 A3 SuperPoD (CloudMatrix 384) links 384 Ascend 910C chips and is sold to Alibaba, Tencent, Baidu, ByteDance and DeepSeek, among others. Huawei makes up for a weaker chip with more of them per system, and pays for it in power: the 384-chip system is reported to draw about 560 kW.
Supply chain
Every Ascend Link Is in China or Comes From Stock Bought Before the Controls
- Made in US
- Some US volume
- US planned
- Made abroad
- Not disclosed
| Component | Supplier | Role | Location | Evidence | US status |
|---|---|---|---|---|---|
| Logic die | SMIC | Primary | Shanghai, China | Claimed | Made abroad |
| Logic die | TSMC | Second source | Taiwan | Confirmed | Made abroad |
| HBM | Samsung | Primary | South Korea | Confirmed | Made abroad |
| HBM | SK hynix | Second source | South Korea | Confirmed | Made abroad |
| HBM | CXMT | Second source | Hefei, China | Claimed | Made abroad |
| Advanced packaging | Not disclosed | Primary | China | Inferred | Made abroad |
| Final test | Not disclosed | Primary | China | Inferred | Made abroad |
None of the seven links is made in the US, and Huawei has no US route to add. The full record, with sources for each row, is on the Ascend 910C supplier ring. The same dataset shows the opposite pattern for the other nine chips tracked: TSMC appears in all nine, and the supplier rings hub compares all 10.
Foundry
SMIC N+3 and Hybrid Bonding Stand In for EUV
| Chip | Phone | Process | Note |
|---|---|---|---|
| Kirin 9000 | Mate 40 (2020) | TSMC 5nm | Last HiSilicon chip made at TSMC under normal terms |
| Kirin 9000S | Mate 60 Pro (2023) | SMIC N+2, 7nm class | First proof of 7nm-class volume without EUV |
| Kirin 9010, 9020 | Pura 70, Mate 70 (2024) | SMIC N+2 | Same node, redesigned cores |
| Kirin 9030 | Mate 80 (late 2025) | SMIC N+3 | TechInsights: denser than N+2 but well short of 5nm |
| Kirin 9050 Pro | Mate 90 Pro Max (2026) | SMIC N+3 compute die on N+2 die, hybrid-bonded | LogicFolding; first mass-produced Huawei stacked logic |
SMIC makes every current Huawei chip. Its N+2 process (7nm class) uses deep ultraviolet (DUV) scanners with multiple patterning, and N+3 pushes the same tools further; TechInsights confirmed N+3 in the Kirin 9030 in December 2025 and found it denser than N+2 but well short of commercial 5nm nodes. With no EUV, each step costs more exposures and yield. The Kirin 9050 Pro takes another route: two dies, one on each node, joined copper to copper so that each major block spans both. Huawei says LogicFolding uses more than 50 million hybrid bonds and raised transistor density by 53.5%. China still imports most hybrid bonding equipment; Piotech, the leading domestic maker, reported 136 million yuan of bonding tool revenue in 2025.
SMIC itself is running full. Second-quarter 2026 revenue was $3.006 billion, its first quarter above $3 billion, with 93.7% utilization and 90.2% of revenue from China. See Process Nodes and Hybrid Bonding.
Memory
HBM, Not Logic Wafers, Caps Ascend Output
| Source | What | Status |
|---|---|---|
| Stockpile | Samsung and SK hynix HBM bought before the December 2024 controls; about 11.7 million stacks (SemiAnalysis estimate) | Finite; found in 910C teardowns |
| CXMT, Hefei | HBM3, then HBM3E | HBM3 sampling to Huawei; HBM3E in small-scale production by late August 2026, tested by T-Head and Cambricon (reported) |
| Huawei-designed memory | HiBL 1.0 (128 GB, 1.6 TB/s) in the Ascend 950PR; HiZQ 2.0 (144 GB, 4 TB/s) in the 950DT | Announced September 2025; maker not disclosed |
HBM caps Ascend output more than logic wafers do. SMIC has wafer capacity for more than a million Ascend dies a year, but every package needs memory stacks, and China's own HBM is early: SemiAnalysis put CXMT's overall yield on eight-layer HBM3 at about 25% in June 2026, and lower for 12-layer parts. CXMT's scale is not the limit. It listed on Shanghai's STAR Market in July 2026 and reported first-half revenue of 150.3 billion yuan, up 874%, on commodity DRAM prices. See HBM.
Roadmap
Huawei Answers the Per-Chip Gap With Chip Count
| Chip | Timing | Logic | Memory | System |
|---|---|---|---|---|
| Ascend 910C | 2025, in volume | Two 910B dies, SMIC N+2 | Stockpiled HBM | Atlas 900 A3 SuperPoD: 384 chips |
| Ascend 950PR | 2026; mass production from April (reported) | SMIC | HiBL 1.0, 128 GB, 1.6 TB/s | Inference prefill and recommendation |
| Ascend 950DT | 2026 | SMIC | HiZQ 2.0, 144 GB, 4 TB/s | Atlas 950 SuperPoD: 8,192 chips in 160 cabinets, fourth quarter 2026 |
| Ascend 960 | 2027 (planned) | Not disclosed | Twice the 950's capacity | Atlas 960 SuperPoD: 15,488 chips in 220 cabinets |
| Ascend 970 | 2028 (planned) | Not disclosed | Not disclosed | Higher performance than the 960 |
Bloomberg reported in September 2025 that Huawei planned about 600,000 Ascend 910C chips for 2026, double 2025, and up to 1.6 million dies across the line, with yields still low. A March 2026 report put the 950PR plan at about 750,000 units for the year; neither figure is confirmed by Huawei. The roadmap answers the per-chip gap with scale: superpods of thousands of chips, Huawei's own memory and its CANN software stack, which Huawei said in 2025 it would open-source along with the MindSpore framework.
Controls
Cut Off From TSMC Since 2020
| Date | Measure | Effect on Huawei |
|---|---|---|
| May 15, 2020 | US Foreign Direct Product Rule extended to Huawei | TSMC stopped making HiSilicon chips; Kirin 9000 was the last |
| 2023 | Kirin 9000S ships on SMIC N+2 | Showed 7nm-class volume without EUV |
| 2024 | TSMC dies reach the 910C through Sophgo (about three million, SemiAnalysis) | TSMC says it halted the orders; the die bank is used up |
| December 2, 2024 | US controls on HBM sales to China and more chipmaking tools | New HBM imports cut; stockpile and CXMT remain |
| April 2025 | US license required for NVIDIA H20 sales to China | Last legal NVIDIA product for China removed for a time |
| September 2025 | Beijing tells state buyers to stop buying NVIDIA GPUs (reported) | Ascend favored in state-linked purchases |
| January 13, 2026 | US moves H200 licenses for China to case-by-case review | NVIDIA licensed for small quantities; supply limited |
Each round of US controls has pushed Huawei further into a domestic chain, and Beijing's buying rules have handed that chain demand. The broader picture across foundry, memory, tools and materials is on China Bifurcation.
Record
Page Record
- Covers
- Huawei and HiSilicon: China AI accelerator share, the Ascend supplier ring, SMIC N+2 and N+3, the Kirin 9050 LogicFolding stack, HBM sources, the Ascend roadmap and SuperPoDs, and the US controls since 2020
- Sources
- IDC via Reuters and TechNode (China AI accelerator shares and units, first half of 2026, October 14, 2026); TechInsights (Kirin 9030 and SMIC N+3, December 11, 2025); Notebookcheck citing a Geekerwan teardown (Kirin 9050 Pro, October 3, 2026); SemiAnalysis and Huawei via KuCoin News (LogicFolding bonds and density, Piotech, October 16, 2026); SMIC second-quarter 2026 results via AnySilicon; Korea JoongAng Daily citing The Information and SemiAnalysis (CXMT HBM3E, revenue and IPO, September 1, 2026); Reuters factbox (Ascend roadmap and SuperPoDs, September 18, 2025); Bloomberg via Gulf News (2026 output plans, September 2025); Newsquawk (950PR production plan, March 27, 2026, reported); Igor's Lab citing NVIDIA's 10-K and Reuters (H200 licensing, 2026); SemiconductorX chip supplier dataset (Ascend 910C ring)
- Last verified
- October 2026