The Battle for China’s AI Chip Throne: Huawei Ascend, Domestic Substitution & What It Means for Foreign Partners

1. The Three-Year Transformation in Numbers

The speed of China’s AI chip substitution is without precedent in semiconductor history. According to IDC’s 2025 China Cloud AI Accelerator Market Report published in April 2026, China’s total AI accelerator card deliveries reached 4 million units in 2025. Of these, domestic Chinese vendors delivered 1.65 million units, capturing 41% market share — a leap from approximately 30% in 2024. NVIDIA delivered about 2.2 million units for a 55% share, down from a historical high of roughly 95%.

The forward-looking projections are even more dramatic. Bernstein Research estimates that by 2026, domestic vendors will collectively hold more than 50% of the market, with Huawei alone reaching approximately 50% share. Morgan Stanley’s May 2026 report is even more bullish on Huawei: it projects Huawei Ascend will capture 62% of China’s domestic AI accelerator market in 2026, with Cambricon at 14%, and NVIDIA collapsing to just 8%. TrendForce noted in August 2026 that domestic solutions could approach 90% share in the high-end segment.

2. Huawei Ascend: The Dominant Force

Huawei is not merely the largest domestic vendor — it is rapidly becoming the central pillar of China’s AI compute supply. In 2025, Huawei Ascend shipped 812,000 units, representing 49.2% of all domestic shipments and roughly 20% of the entire Chinese market. The company has articulated a “one generation per year, doubling compute” roadmap, with the Ascend 950 series launching in 2026, followed by 960 in 2027 and 970 in 2028.

The software ecosystem is equally important. Huawei’s CANN (Compute Architecture for Neural Networks) — the domestic answer to NVIDIA’s CUDA — has matured to the point where leading Chinese model companies are actively porting their training stacks. The landmark moment came when DeepSeek V4 — a trillion-parameter model — migrated its training pipeline from CUDA to CANN, proving that domestic chips can now handle front-line large-model training, not just inference.

3. Other Key Players: “One Superpower, Multiple Strong”

While Huawei dominates, the domestic landscape is evolving into a “one superpower, multiple strong” pattern, with vendors differentiating by scenario, architecture, and customer segment.

4. What This Means for Foreign Businesses

For overseas firms evaluating or currently partnering with Chinese AI companies, the chip landscape shift introduces four distinct risk dimensions:

4.1 Performance & Compatibility Risk

Domestic chips are no longer uniformly “inferior” — in specific scenarios (especially inference with Chinese-origin models), Huawei Ascend 950PR can outperform NVIDIA H20 by 2.87×. However, breadth of compatibility remains a challenge. Not every model, framework, or workload has been ported to CANN or other domestic software stacks. If your Chinese partner is running a specialized workload, ask specifically: has this workload been benchmarked on their chosen domestic chip? Generic claims of “Ascend-compatible” are insufficient.

4.2 Supply-Chain Continuity Risk

Export controls cut both ways. Your Chinese partner may face difficulty procuring NVIDIA chips for new capacity, but they also may face constraints in servicing overseas customers depending on the specific chip, end-use, and jurisdiction. A partner who has proactively built a hybrid fleet (NVIDIA for export-compliant workloads + domestic for China-domestic workloads) is far more resilient than one who bet entirely on a single vendor.

4.3 Policy & Localization Risk

Chinese government-related projects and state-owned enterprise procurement now explicitly favor domestic chips, with domestic share in government and telecom scenarios reaching 42% in 2025. If your Chinese partner serves government or SOE clients, their chip mix is effectively mandated. This creates a parallel architecture problem: the infrastructure serving Chinese domestic clients may be fundamentally different from the infrastructure serving your overseas workloads. Understanding this split is essential for compliance and performance planning.

4.4 Vendor Financial Risk

The domestic substitution wave is creating enormous financial volatility. Cambricon’s 453% revenue jump is spectacular, but its 88.66% customer concentration and 178.67% inventory surge signal both opportunity and fragility. Smaller vendors like Moore Threads (738M RMB revenue) are growing fast but remain thinly capitalized. Before committing to a long-term partnership, verify your vendor’s financial health with an independent professional enterprise credit report.

5. The Software Stack Issue: CANN vs. CUDA

The semiconductor industry has an old saying: “chips without a software stack are unusable.” The true battle for China’s AI chip throne is not being fought on the silicon — it is being fought in the software layers that sit between the chip and the model.

NVIDIA’s CUDA has a 15+ year head start, with an ecosystem of libraries, frameworks, and developer mindshare that took nearly two decades to build. China’s domestic response includes:

The migration cost from CUDA to CANN varies dramatically by workload. For standard inference serving of popular models (LLaMA-family, Qwen, DeepSeek), third-party platforms like SiliconFlow and Infinigence-AI now provide abstraction layers that make the underlying chip nearly invisible to the application developer. For training frontier models or running highly custom kernels, the migration cost remains substantial — often requiring model architecture adjustments and months of engineering effort.

6. Due Diligence in the Age of Domestic Substitution

Evaluating a Chinese AI company’s chip strategy requires a specialized due diligence framework. Here is a practical checklist tailored to the 2026 reality:

• Map the chip inventory: What is the exact breakdown of NVIDIA, Huawei Ascend, Cambricon, and other chips in your partner’s fleet? What is the ratio for your specific workloads vs. their overall capacity? An independent standard business credit report can verify procurement contracts and capital expenditures against declared capacity.
• Assess migration readiness: Which models and workloads have been ported to domestic chips? Request benchmark data showing performance deltas between NVIDIA and domestic silicon for your use case. Generic “we support Ascend” claims are insufficient.
• Verify software stack maturity: Is your partner using CANN, MindSpore, or a third-party abstraction layer? Have they demonstrated production-grade stability on domestic chips, or are they still in pilot phase?
• Evaluate export-control exposure: Does your partner’s chip mix create compliance obligations for your company under U.S., EU, or other jurisdictional export control regimes? Consult with your own compliance team, but verify the factual basis through an independent official enterprise credit report that confirms the partner’s actual technology stack.
• Check financial resilience: Domestic chip vendors themselves carry financial risk. If your partner is heavily dependent on a vendor like Cambricon (88.66% customer concentration) or Moore Threads (thin margins), their supply continuity is only as strong as their primary chip supplier’s balance sheet.
• Assess policy compliance: For partners serving government or SOE clients, verify their domestic chip ratio meets the implicit localization thresholds (42%+ in gov/telecom as of 2025). Non-compliance risks contract termination and operational disruption.