Data Panorama
The latest domestic chip research data outlines the new landscape of China’s AI computing power in 2026:
| Vendor/Series | 2025 Shipment | 2026 Estimated | Notes |
|---|---|---|---|
| Ascend 910 Series | ~520K units | 750K units | Q1 confirmed 150K cards shipped |
| Cambricon | - | Part of tier-1 | Q1 revenue 2.9B yuan, +53% QoQ |
| Hygon | - | Part of tier-1 | Tier-1 combined 1.6-1.7M units |
| Moore Threads/Biren etc. | - | ~1M units | Rapidly catching up |
| Kunlunxin | - | 300K+ units | Baidu self-developed + external sales |
| Total | - | ~3M units | Excluding internet giants’ self-developed |
What does 3 million units mean? If estimated at average inference performance per card, these 3 million domestic chips can support the inference needs of hundreds of millions of daily active AI applications — precisely covering the user base of current mainstream Chinese AI applications.
Ascend: Establishment and Concerns of the Leading Position
Huawei Ascend is the absolute leader in domestic AI chips. Several key signals:
Major customers vote with orders: ByteDance alone ordered 250,000 Ascend 950PR cards, Alibaba 120,000, Tencent 120,000. These three internet giants’ total procurement has reached 490,000 cards, approaching 65% of Huawei’s full-year 750,000 target.
Price advantage: Ascend 950PR is priced at approximately 70,000 yuan, less than one-third of NVIDIA’s H200. For Chinese internet enterprises primarily focused on inference scenarios, this is an extremely attractive cost-performance proposition.
Ecosystem binding: The DeepSeek V4 technical report for the first time lists Huawei Ascend 950PRO in the official hardware verification list. This means top AI models are proactively adapting to the Ascend ecosystem, not being forced.
But Ascend also faces concerns: Q4 2025 shipments were weak, with the full-year 520,000 units concentrated in Q2-Q3. Supply chain capacity and yield rate remain constraining factors.
Cambricon: Signals Behind the Better-Than-Expected Q1 Report
Cambricon’s 2026 Q1 report is an impressive scorecard:
- Revenue of 2.9 billion yuan, up 53% quarter-over-quarter, 60% above Goldman Sachs expectations
- EBITDA margin jumped from 26% last quarter to 42%
- Contract liabilities surged from 600,000 yuan to 396 million yuan — customers are pre-paying, indicating sufficient orders on hand
- Inventory of 4.5 billion yuan, slightly down from 4.9 billion last quarter — shipments are accelerating
Goldman Sachs simultaneously updated its coverage report, suggesting institutional confidence in Cambricon is strengthening. However, note that Cambricon’s product line focuses on training chips, and its market share in inference still needs expansion.
Landscape Assessment: Three Tiers Taking Shape
The domestic AI chip market has formed a clear three-tier structure:
First tier (Ascend/Cambricon/Hygon): Combined 1.6-1.7 million units. Complete software stacks (Ascend’s CANN, Cambricon’s Neuware, Hygon’s ROCm-compatible layer), covering both training and inference full scenarios.
Second tier (Moore Threads/Biren etc.): Approximately 1 million units. Differentiated advantages in specific scenarios (such as inference optimization, edge computing), but software ecosystems still need improvement.
Third tier (Kunlunxin and others): 300,000+ units. More serving parent company internal needs, limited external sales proportion.
Action Recommendations
- AI infrastructure procurement: If primarily inference-focused with limited budget, Ascend 950PR is the cost-performance first choice; if training capability is needed, Cambricon and Hygon are worth evaluating
- Developers: Recommended to adapt early to CANN (Ascend) and Neuware (Cambricon) ecosystems, domestic chip developer tools are rapidly iterating
- Investment perspective: Q1 reports from 10 Ascend supply chain companies showed net profit surging 32 times, supply chain opportunities have extended from chip design to packaging, testing, thermal management and other upstream/downstream segments