Technical Deep Dive
The core of this shift lies in the architectural and software differences between Nvidia's CUDA and Huawei's CANN (Compute Architecture for Neural Networks). CUDA has been the de facto standard for GPU computing for over a decade, with a mature toolchain (cuDNN, TensorRT, NCCL) and a vast library of optimized kernels. Huawei's CANN, while younger, is designed from the ground up for its Ascend series, which uses a Da Vinci architecture—a heterogeneous array of AI cores (Cube, Vector, Scalar) that differs fundamentally from Nvidia's CUDA cores.
Architecture Comparison:
| Feature | Nvidia H100 (Hopper) | Huawei Ascend 910B | Huawei Ascend 910C (rumored) |
|---|---|---|---|
| Process Node | TSMC 4N | SMIC N+2 (7nm-class) | SMIC N+2 (enhanced) |
| FP16 TFLOPS | 1979 | ~320 | ~400 (est.) |
| HBM Memory | 80GB HBM3 | 64GB HBM2e | 96GB HBM3 (est.) |
| Interconnect | NVLink 4.0 (900GB/s) | HCCS (200GB/s) | HCCS (400GB/s est.) |
| Software Stack | CUDA 12.x, TensorRT | CANN 7.x, MindSpore | CANN 8.x (est.) |
| Power (TDP) | 700W | 310W | 400W (est.) |
Data Takeaway: While Nvidia's H100 still dominates raw performance, Huawei's 910B achieves competitive inference throughput per watt—critical for China's energy-constrained data centers. The 910C, if it reaches 400 TFLOPS, will close the gap to Nvidia's A100 in training tasks, though H100 remains out of reach.
Software Lock-in Dynamics:
CANN's key innovation is its operator fusion and graph compilation engine, which automatically optimizes neural network graphs for Ascend hardware—similar to XLA for TensorFlow but hardware-specific. Huawei has also open-sourced MindSpore (a PyTorch-like framework) and provides migration tools (e.g., `msadvisor`) that automatically convert CUDA kernels to CANN-compatible operators. On GitHub, the `mindspore-ai/mindspore` repository has surpassed 4,000 stars, while the `Ascend/samples` repo provides over 500 code examples for model conversion. However, the conversion success rate for complex models (e.g., Mixture-of-Experts architectures) remains around 70-80%, meaning significant manual tuning is still required.
Takeaway: The technical gap is narrowing fast. Huawei's strategy is not to beat Nvidia on peak FLOPS but to match performance on the most common AI workloads (LLM inference, vision transformers) while offering better power efficiency and a fully domestic supply chain. The real battle is software ecosystem maturity, and here Huawei is investing billions to replicate CUDA's developer lock-in.
Key Players & Case Studies
Huawei's Ascend Strategy:
Huawei has deployed a three-pronged approach: (1) hardware iteration—from 910B to the upcoming 910C and the server-grade Atlas 900 cluster; (2) software ecosystem—CANN, MindSpore, and the MindX SDK for model deployment; (3) developer incentives—free cloud credits, training programs, and a dedicated partner network. Major Chinese cloud providers (Alibaba Cloud, Tencent Cloud, Baidu AI Cloud) now offer Ascend-based instances, and ByteDance has reportedly deployed tens of thousands of Ascend chips for internal recommendation systems.
Nvidia's Response:
Nvidia has not completely abandoned China. It continues to sell the H20 (a cut-down H100 with reduced interconnect bandwidth) and the L20 (inference-focused). But Huang's admission signals a strategic pivot: the company is doubling down on its Blackwell architecture (B200, GB200) for Western hyperscalers, targeting 1,000W+ TDP with liquid cooling, and expanding into automotive (Drive Thor) and robotics (Isaac). Nvidia is also pushing its 'AI foundry' model, offering custom model training services to governments and enterprises outside China.
Case Study: Baidu's ERNIE Bot Migration
Baidu, a long-time Nvidia customer, began migrating its ERNIE 3.5 and 4.0 model training from A100 clusters to Ascend 910B clusters in late 2024. The migration required rewriting 40% of the training pipeline due to differences in distributed communication libraries (NCCL vs HCCS). Training throughput dropped 25% initially, but after six months of optimization, it recovered to 90% of original performance. This case illustrates the high switching costs but also the feasibility of migration.
Competitive Landscape Table:
| Company | Chip | Target Market | Key Advantage | Key Weakness |
|---|---|---|---|---|
| Nvidia | H100, B200 | Global (ex-China) | CUDA ecosystem, NVLink | Export restrictions, high power |
| Huawei | Ascend 910B/C | China, Belt & Road | Domestic supply, power efficiency | Software immaturity, HBM access |
| AMD | MI300X | Global (ex-China) | Open-source ROCm, competitive price | Smaller developer base |
| Intel | Gaudi 3 | Global (ex-China) | Ethernet-based scaling | Late to market |
| Cambricon | MLU590 | China | Specialized for inference | Limited training performance |
Data Takeaway: Huawei is the only non-Western player with a credible full-stack AI solution (chip + software + cloud). Its main bottleneck is access to advanced HBM memory, which is controlled by Samsung and SK Hynix under US pressure. If China's domestic HBM production (e.g., CXMT) matures by 2026, Huawei's position becomes unassailable in China.
Industry Impact & Market Dynamics
The bifurcation of AI computing standards is already reshaping global supply chains and business models.
Market Share Shift:
| Year | Nvidia China Revenue Share | Huawei Ascend Share (China DC) |
|---|---|---|
| 2022 | 25% (est.) | <5% |
| 2023 | 17% | 15% |
| 2024 | 8% | 40% |
| 2025 (proj.) | 3% | 60%+ |
Data Takeaway: Nvidia's China revenue has collapsed from ~$10B annually to under $3B, while Huawei's AI chip revenue is projected to exceed $8B in 2025. The inflection point was the H800 ban in October 2023, which forced Chinese companies to accelerate domestic adoption.
Cost Implications for Multinationals:
Companies operating in both China and the West (e.g., Tesla, Apple suppliers, global banks) now face a painful choice: maintain two separate AI infrastructure stacks or choose one and accept performance penalties in the other. Migrating a large language model from CUDA to CANN costs an estimated $2-5M per model and takes 3-6 months, according to industry estimates. For a company with 20 production models, that's $40-100M in migration costs.
Sovereign AI Movement:
Countries from India to Saudi Arabia are watching closely. India's Centre for Development of Advanced Computing (C-DAC) is evaluating Huawei's Ascend for its national AI compute facility, while the EU's EuroHPC Joint Undertaking is funding research into RISC-V AI accelerators. The message is clear: no country wants to be locked into a single vendor for a technology as strategic as AI.
Takeaway: The AI chip market is transitioning from a winner-take-most dynamic (Nvidia) to a multi-standard, multi-vendor landscape. This will increase costs and complexity in the short term but may foster innovation and resilience in the long term.
Risks, Limitations & Open Questions
Huawei's Challenges:
1. HBM Supply Chain: Huawei relies on Samsung and SK Hynix for HBM2e and HBM3. US export controls could tighten further, cutting off this supply. China's CXMT (ChangXin Memory Technologies) is developing HBM but is at least 2-3 years behind.
2. Software Maturity: CANN still lacks the breadth of CUDA libraries for specialized domains (e.g., computational chemistry, climate modeling). The developer community is growing but remains a fraction of CUDA's.
3. Performance Ceiling: The 910C is expected to match the A100, but the H100/B200 gap remains large. For cutting-edge research (e.g., training a 1-trillion-parameter model), Nvidia hardware is still essential.
Nvidia's Risks:
1. Over-reliance on Hyperscalers: If Google, Microsoft, and Amazon develop more custom chips (TPU, Maia, Trainium), Nvidia's dominance could erode even outside China.
2. Geopolitical Uncertainty: Further export controls could cut off sales to other large markets (e.g., Middle East).
3. Architectural Bet: Blackwell's extreme power density (1,000W+) requires liquid cooling infrastructure that many data centers lack, potentially slowing adoption.
Open Questions:
- Will the US impose secondary sanctions on companies using Huawei chips? This could force a binary choice for multinationals.
- Can Huawei maintain its rapid iteration pace (new chip every 18 months) given semiconductor equipment restrictions?
- Will the two ecosystems converge via open standards (e.g., MLIR, ONNX) or remain permanently split?
AINews Verdict & Predictions
Our Editorial Judgment:
Nvidia's concession is not a defeat—it's a rational response to an irrational policy. The US government created a protected market for Huawei, and Huawei is executing flawlessly. But the long-term winner is neither Nvidia nor Huawei; it's the concept of 'AI sovereignty.' Every major economy will now demand domestic AI chip options, fragmenting the market into at least three blocs: the US/Nvidia ecosystem, the China/Huawei ecosystem, and a European/Indian/RISC-V alternative.
Specific Predictions (2025-2027):
1. By Q1 2026, Huawei's Ascend 910C will achieve parity with Nvidia's A100 in training throughput for models up to 100B parameters, making it the default choice for Chinese enterprises.
2. By Q3 2026, at least one major US cloud provider (likely Oracle or IBM) will offer Huawei Ascend instances for customers needing China-compliant AI workloads, creating a 'dual-stack' cloud.
3. By 2027, the cost of maintaining dual AI stacks will lead to a new middleware market—companies like Anyscale or Ray will offer abstraction layers that transparently route workloads to CUDA or CANN hardware.
4. The biggest loser will be AMD, which is caught in the middle: its ROCm software is too immature to challenge CUDA, and it cannot sell into China due to export controls.
What to Watch Next:
- Huawei's HBM strategy: Watch for announcements of domestic HBM production or strategic partnerships with Chinese memory makers.
- Nvidia's China-specific product: Will Nvidia launch a fully compliant, low-performance chip for China, or exit entirely?
- Regulatory escalation: The US Commerce Department's next rulemaking on chip exports (expected late 2025) could target Huawei's supply chain more aggressively.
The AI chip war is not ending—it's entering a new, more complex phase. The era of a single global AI standard is over.