Huawei’s Ascend 910C—China’s Answer to Nvidia’s AI Chip Dominance

Huawei is poised to ship its Ascend 910C AI chip in mass volumes as early as next month, offering Chinese AI firms a potent alternative to Nvidia’s restricted H20 chips. The 910C pairs two 910B chiplets to match the performance of Nvidia’s H100, leveraging SMIC’s 7 nm N+2 process despite yield challenges. Early deployments in Huawei’s CloudMatrix 384 cluster demonstrate 300 PFLOPs of BF16 compute, though at the cost of higher power draw compared to rivals. Mass shipments will reshape China’s AI infrastructure, reduce dependence on U.S. chips, and intensify the tech decoupling between Washington and Beijing.

Introduction

China’s leading telecoms giant, Huawei, has quietly prepared its newest AI accelerator, the Ascend 910C, for large‑scale deployment. This move follows U.S. export curbs on Nvidia’s H20 chips—measures designed to limit China’s access to cutting‑edge AI hardware. Industry insiders say Huawei will begin mass shipments next month, filling a strategic void for domestic AI developers.

Background on U.S. Export Controls

Since 2022, the U.S. government has required special licenses for exporting high‑end AI chips—including Nvidia’s H20 model—to China, citing national‑security risks. These restrictions have forced Chinese AI firms to seek local alternatives to power large‑scale model training and inference. As licensing delays mount, training budgets and project timelines face uncertainty—heightening demand for a homegrown solution.

Technical Overview of the Ascend 910C

Architecture and Performance

The Ascend 910C effectively doubles the power of its predecessor by integrating two Ascend 910B chiplets on a single package. Each chiplet, manufactured on SMIC’s 7 nm N+2 node, contributes to a combined throughput that rivals Nvidia’s H100 GPU across key benchmarks. According to HuaweiCentral, the 910C can deliver around 800 TFLOPs of FP16 performance with roughly 3.2 TB/s of memory bandwidth—about 80% of the H100’s specs.

Tom’s Hardware tests show the Ascend 910C achieving 780 TFLOPs of BF16 compute per chip, translating into 300 PFLOPs at the cluster level when deployed in CloudMatrix 384. While powerful, the system consumes 559 kW—over four times the power draw of Nvidia’s GB200‑based clusters—an inefficiency offset by China’s lower electricity costs and abundant grid capacity.

Manufacturing and Supply Chain

SMIC handles the primary chiplet fabrication, but some Ascend 910C units reportedly incorporate TSMC‑made silicon from pre‑2020 stockpiles, prompting U.S. scrutiny and an ongoing investigation. TSMC itself maintains it has not supplied Huawei since 2020 and continues to follow export regulations strictly. Bloomberg notes that Huawei’s ability to secure enough 7 nm wafers remains a bottleneck, though stockpiles and alternative foundry partnerships help mitigate shortfalls.

Market and Geopolitical Implications

Huawei’s mass deployment of the Ascend 910C will likely become the hardware backbone for China’s leading AI labs, including ByteDance, Baidu, and China Mobile. This shift reduces Chinese firms’ reliance on Nvidia and signals a broader push for semiconductor self‑sufficiency amid intensifying U.S.‑China tech tensions. Financial analysts forecast that diverting even a fraction of China’s $17 billion annual Nvidia spend to Huawei could reshape global chip-market dynamics.

TechRadar reports that the Ascend 910C‑powered CloudMatrix cluster already exceeds Nvidia’s GB200 in raw petaflop capacity, underscoring China’s system‑level ingenuity despite chip‑level efficiency gaps. SemiAnalysis adds that optical interconnects and high‑density packaging further bolster performance, though energy inefficiency and yield risks loom large.

Challenges and Risks

• Yield Constraints: SMIC’s N+2 process yields remain lower than those at leading-edge foundries, risking supply shortfalls.
• U.S. Policy Escalation: Washington could broaden controls to Chinese fabs, tightening access to critical equipment and design tools.
• Power Inefficiency: At 559 kW per cluster, operating costs and infrastructure requirements may deter some adopters.

In addition, deep‑dive tests from DeepSeek Research suggest inference performance on the 910C trails the H100 by about 40%, potentially affecting latency‑sensitive applications.

Future Outlook

Looking ahead, Huawei plans to refine the N+2 process yields and explore EUV upgrades in collaboration with domestic foundries. Reports indicate follow‑on chips (e.g., Ascend 910D) are already in design, promising 50% better energy efficiency and tighter integration. Meanwhile, U.S. policymakers are evaluating further export curbs—a move that could either entrench Huawei’s position or spur renewed multilateral dialogue on tech trade rules.

Conclusion

Huawei’s Ascend 910C emerges as a bold challenge to Nvidia’s AI GPU hegemony, leveraging chiplet design and China’s system‑integration strengths to fill a strategic void. While power efficiency and policy risks persist, mass shipments set for next month will mark a pivotal shift in global AI infrastructure sourcing. Stakeholders should monitor yield improvements, export‑control developments, and competitive responses from Nvidia and other chip leaders.