Technical Deep Dive
Huawei's Ascend series processors are built on a Da Vinci architecture that emphasizes a unified, scalable compute unit called the Cube. Each Cube contains 16 AI cores, and the architecture scales by tiling multiple Cubes on a die. The Ascend 910B, currently the flagship, delivers 256 TFLOPS of FP16 compute and 512 TFLOPS of INT8 performance, with 32 GB of HBM2e memory providing 1.2 TB/s bandwidth. The upcoming Ascend 910C is expected to double memory bandwidth to 2.4 TB/s using HBM3 and increase compute density through a more aggressive 7nm-class process.
However, the manufacturing process is the critical bottleneck. SMIC's N+2 process, which is equivalent to TSMC's 7nm node, uses deep ultraviolet (DUV) lithography with multiple patterning due to the absence of extreme ultraviolet (EUV) tools. This increases manufacturing complexity, reduces yields, and limits the number of chips per wafer. A typical 300mm wafer yields approximately 80-100 usable dies for a chip the size of the Ascend 910B (roughly 600 mm²), compared to 120-150 for a comparable TSMC-manufactured chip. This yield gap directly translates into higher per-chip costs and constrained supply.
| Process Node | Equivalent TSMC Node | EUV Required? | Estimated Yield (600mm² die) | Cost per Wafer (est.) |
|---|---|---|---|---|
| SMIC N+1 | 10nm | No | 60-70% | $4,500 |
| SMIC N+2 | 7nm | No | 40-55% | $5,500 |
| SMIC N+3 (planned) | 5nm | Yes (limited) | 20-30% (est.) | $8,000+ |
Data Takeaway: The yield penalty for non-EUV manufacturing is severe, reducing usable dies by 30-50% compared to TSMC. This directly limits Huawei's ability to scale production even if demand is unlimited.
Huawei's software stack, CANN (Compute Architecture for Neural Networks), has been optimized to reduce memory bandwidth pressure and improve utilization. Recent benchmarks show that with CANN 7.0, the Ascend 910B achieves 85% of the training throughput of an Nvidia A100 on GPT-3 scale models, up from 65% a year ago. For inference, the gap is narrower, with the 910B matching the A100 on batch sizes of 32 or larger for Llama 2-70B. These improvements come from better operator fusion, memory pooling, and automatic mixed-precision scheduling.
A relevant open-source project is the MindSpore framework, which Huawei has been pushing as an alternative to PyTorch. MindSpore's GitHub repository has accumulated over 4,000 stars and supports automatic graph optimization that is tightly coupled with the Ascend hardware. However, the ecosystem remains fragmented, with many Chinese AI startups still preferring PyTorch with custom Ascend backends.
Key Players & Case Studies
The primary players in this ecosystem are Huawei (designer and integrator), SMIC (foundry), and a cluster of Chinese cloud providers and AI startups that are the primary customers.
Huawei has adopted a dual strategy: selling chips directly to cloud providers like Alibaba Cloud, Tencent Cloud, and Baidu AI Cloud, while also deploying them in its own Ascend cloud service. The internal consumption helps validate performance and absorb initial supply, but external sales are where the revenue growth lies. Huawei's chip division, HiSilicon, continues to design the architecture, but it cannot manufacture at scale without SMIC.
SMIC is the only foundry in China capable of 7nm-class production. Its capacity for N+2 is estimated at 15,000-20,000 wafers per month (WPM), of which Huawei reportedly takes 70-80%. The remaining capacity is allocated to a handful of other clients, including Bitmain for cryptocurrency mining chips and some automotive AI chips. SMIC's new fab in Shanghai, expected to add 10,000 WPM of N+2 capacity, will not begin volume production until late 2026 at the earliest.
Competing Solutions: Several Chinese AI chip startups are attempting to fill the gap, but none have achieved production scale.
| Company | Chip | Process Node | FP16 TFLOPS | Status |
|---|---|---|---|---|
| Huawei | Ascend 910B | SMIC N+2 | 256 | In production |
| Cambricon | MLU370 | TSMC 7nm (stockpiled) | 128 | Limited supply |
| Biren Technology | BR100 | TSMC 7nm (stockpiled) | 256 | Pre-production |
| Enflame | T20 | SMIC N+2 | 192 | Sampling |
Data Takeaway: Huawei has a commanding lead in both performance and production volume. Cambricon and Biren rely on stockpiled TSMC wafers from before the export restrictions, limiting their ability to scale. Enflame's T20 is promising but still in sampling, with no volume production timeline.
Case Study: ByteDance ByteDance, which runs massive AI workloads for Douyin and recommendation systems, has been one of Huawei's largest customers. In Q1 2025, ByteDance placed an order for 50,000 Ascend 910B chips for training its next-generation video generation model. Delivery has been delayed by 8-12 weeks due to wafer shortages, forcing ByteDance to extend its use of Nvidia A100 chips it had stockpiled before the export ban. This case illustrates how the capacity bottleneck directly impacts the pace of AI model development in China.
Industry Impact & Market Dynamics
The $12 billion revenue target represents approximately 20% of the global AI chip market (excluding Nvidia's dominance) and would make Huawei the second-largest AI chip supplier by revenue after Nvidia. However, achieving this requires shipping roughly 1.5 million Ascend 910B chips (at an average selling price of $8,000), which would consume nearly all of SMIC's N+2 capacity for the year.
| Year | Huawei AI Chip Revenue | SMIC N+2 Capacity (WPM) | Estimated Shipments (units) | Market Share in China |
|---|---|---|---|---|
| 2023 | $2.5B | 10,000 | 300,000 | 15% |
| 2024 | $6.0B | 15,000 | 750,000 | 35% |
| 2025 (est.) | $12.0B | 20,000 | 1,500,000 | 55% |
Data Takeaway: Huawei's revenue growth is directly tied to SMIC's capacity expansion. If SMIC cannot increase N+2 capacity beyond 20,000 WPM, Huawei will miss its revenue target by 20-30%.
The broader market dynamics are reshaping China's AI supply chain. Cloud providers are now competing for wafer allocation, not just chip availability. This has led to a new business model: long-term wafer reservation agreements, where cloud companies prepay for future capacity. Alibaba Cloud recently signed a $1.5 billion prepayment agreement with SMIC for N+2 wafers, effectively locking in supply for its own AI infrastructure.
Another emerging trend is the use of chiplet architectures to reduce die size and improve yield. Huawei is reportedly working on a multi-die version of the Ascend 910C that would use four smaller dies connected via a silicon interposer, similar to AMD's approach. This would allow each die to be manufactured on a less advanced node (N+1 or even 28nm) while achieving aggregate performance comparable to a monolithic 7nm chip. The trade-off is increased power consumption and packaging complexity, but it could bypass the capacity bottleneck entirely.
Risks, Limitations & Open Questions
Yield Risk: SMIC's N+2 yields are reportedly in the 40-55% range, meaning nearly half of all wafers produce defective chips. Improving yield is the single most impactful lever for increasing supply, but it requires process tweaks that take months to validate. If yields drop further due to equipment wear or process drift, Huawei's shipments could fall by 30% or more.
Equipment Sanctions: The US and its allies continue to tighten export controls on semiconductor manufacturing equipment. SMIC's ability to maintain and repair its DUV lithography tools depends on a supply chain that is increasingly restricted. Any disruption in spare parts or maintenance services could force fab shutdowns.
Alternative Foundries: Huawei has explored using other Chinese foundries like Hua Hong Semiconductor (which focuses on 28nm and above) and CXMT (DRAM specialist), but none have the process capability for high-performance AI chips. The only other option is Samsung, which could theoretically manufacture chips for Huawei under a foundry agreement, but geopolitical risks make this unlikely.
Software Ecosystem Fragmentation: While CANN has improved, it still lags behind CUDA in terms of library support, debugging tools, and community contributions. Many Chinese AI developers report spending 20-30% of their time on porting and optimization work when moving from Nvidia to Ascend. This friction slows adoption and reduces the effective performance of Ascend chips in real-world deployments.
Ethical and Strategic Concerns: The concentration of AI chip production in a single foundry (SMIC) creates a single point of failure. If SMIC faces a natural disaster, labor strike, or geopolitical incident, China's entire AI industry could grind to a halt. Diversifying foundry sources is a national security priority, but there are no viable alternatives in the near term.
AINews Verdict & Predictions
Huawei's $12 billion revenue target is achievable only if SMIC can push N+2 yields above 60% and add at least 5,000 WPM of new capacity by mid-2026. Based on current trends, we predict Huawei will reach $9-10 billion in AI chip revenue this year, falling short of the target by 15-25%. The shortfall will be partially offset by higher average selling prices as Huawei prioritizes high-margin chips for cloud providers over lower-margin edge AI chips.
Prediction 1: By Q3 2026, Huawei will announce a chiplet-based Ascend 920 that uses four 28nm dies to achieve performance comparable to the 910B. This will allow it to bypass the N+2 capacity bottleneck and scale production to 3 million units per year using more available 28nm capacity.
Prediction 2: The Chinese government will invest $5-10 billion in a second advanced foundry, likely a joint venture between SMIC and a state-owned enterprise, to be built in Chengdu or Wuhan. This facility will target 5nm-class production using imported DUV tools, with first wafers expected in 2028.
Prediction 3: The software ecosystem gap will narrow faster than expected. By end of 2026, CANN will support 90% of the PyTorch operators commonly used in LLM training, and MindSpore will gain significant traction among Chinese AI startups due to government procurement preferences.
What to watch next: The key indicator is SMIC's quarterly earnings calls, specifically any mention of N+2 yield improvements or capacity expansion timelines. Also watch for Huawei's chiplet patent filings and any announcements about new foundry partnerships. The next 12 months will determine whether China's AI chip independence is a realistic goal or a strategic mirage.