"Winning with Quantity Despite Falling Short on Performance: Huawei AI Chip's Counterattack - The International Impact of Nvidia's Instability"

On the other hand, there is a range in these figures. U.S. government officials indicated at a hearing in June this year that Huawei's production of advanced AI chips in 2025 would not exceed 200,000 units, showing a divided assessment of supply capacity.Reuters

What's News: Turning Point in Numbers

• 910C to approximately 600,000 units in 2025—about double this year. In 2026, the entire Ascend family to up to 1.6 million dies. Observations suggest overcoming stagnation due to U.S. sanctions.Bloomberg

• NVIDIA's "Wavering Products for China"—H20 faces increased uncertainty due to new energy-saving standards in China, raising concerns about the company's declining composition ratio.Business Insider

• Improvement in Production Yield—Huawei's advanced AI chips have improved yield to about 40%, reaching a profitable level, according to the UK FT. This enhances the feasibility of mass production plans.Financial Times
  
  

Background: Filling the "Gap" in U.S.-China Tech Dominance

Due to strict export controls from 2024 onwards, NVIDIA's cutting-edge GPUs have become difficult to obtain in China. Combined with unique Chinese standards, the company's lineup for China (such as H20) has faced frequent revisions. This demand-supply mismatch has created a strong incentive for "domestic alternatives," accelerating the capabilities of Chinese players like Huawei and SMIC.Business Insider

NVIDIA CEO Jensen Huang recently stated, "China is only 'nanoseconds' behind the U.S.," calling for easing excessive export restrictions. Despite market uncertainties, the company aims to maintain its presence with "regulation-compliant" products.Tom's Hardware

Technical Discussion: Performance vs. Scale Trade-off

The standalone performance of the Ascend 910C is considered less efficient compared to the top-tier NVIDIA GPUs of the generation. Meanwhile, Huawei is adopting an approach to expand cluster size and optimize communication with technologies like optical interconnects, pushing with "quantity over quality." A representative example is the CloudMatrix 384 cluster, which theoretically surpasses NVIDIA's large-scale systems in BF16 performance, but is evaluated as less power-efficient. Due to low power costs and the necessity for domestic procurement, there is room for adoption in China based on "overall utility."Tom's Hardware

Supply and Demand Impact: Who Will Use This Chip?

China's big tech companies (Alibaba, Tencent, Baidu, ByteDance, etc.) are expanding the adoption of domestic chips to reduce dependency on NVIDIA and ensure procurement stability. They are redesigning necessary performance and costs for each model, employing a strategy to fill gaps with "best-effort scaling." This trend encourages the expansion of the domestic ecosystem (software frameworks, HBM, optical communication, cooling).Tom's Hardware

Outlook: "Battle of Numbers"

Reports via Bloomberg/Japanese economic media indicate 600,000 units in 2025 and 1.6 million dies in 2026, while U.S. government estimates suggest fewer than 200,000 units in 2025, showing a strong lower limit. Which is closer depends on the reality of yield, equipment procurement, and supply chain (HBM, advanced packaging, EUV/DUV alternatives). Additionally, an ambitious plan to triple AI chip production by 2026 in China as a whole is reported, and the depth of policy and capital cannot be ignored.Reuters

SNS Reaction: "Polarization" of Enthusiasm and Skepticism

• Optimists: Many cite reports of "910C expanding to 600,000 units in 2026, with a total of 1.6 million dies for the Ascend family," expressing expectations to "fill NVIDIA's gap domestically." The numbers have been widely circulated on X.X (formerly Twitter)

• Cautious Voices: On Reddit, comments like "far from next-gen NV's VR200 class" and "compensating for power efficiency lag with quantity" are prominent. Many realistic voices suggest "performance falls short, but wins in procurement possibility."Reddit

• Skeptics: Doubts about the credibility of production targets. Comments like "optimistic for 600,000 units next year based on this year's results and past yields" and rebuttals to the "H100 equivalent" theory were observed.Reddit
  
  

Editorial Perspective: Victory is Not Determined by "Absolute Performance"

The main battlefield of global AI infrastructure is a multi-dimensional equation of pure performance × availability × cost × energy. While NVIDIA leads in absolute performance, Huawei follows with supply certainty and scale design. The strategy of enhancing business continuity by "stacking" available computing resources horizontally is rational in the domestic market, and if software optimization (such as CANN and PyTorch compatibility) catches up, scenarios where it can "compete adequately" depending on the task will expand.

However, the key lies in power efficiency and HBM procurement. The difference in performance per power directly impacts TCO and cannot be ignored even in China, where there is a rush to establish new data centers. The medium- to long-term outcome depends on system optimization, including power, cooling, and optical wiring.

Risk Factors

1. Regulatory Changes: Additional U.S. regulations and strengthened energy-saving standards in China influence both product roadmaps.Business Insider

2. Yield and Equipment Constraints: Even with improvements to the 40% range, constraints on advanced packaging and exposure equipment remain.Financial Times

3. Speed of Generational Change: NVIDIA's next-generation introduction cycle is fast. The pattern of "catching up only to face the next generation" continues.Tom's Hardware
   
   

Summary

While the numbers are optimistic, the reality of power efficiency and supply chain constraints is also harsh. Nevertheless, in the Chinese market, where "availability" holds value, Huawei's plan to double the Ascend serves as a significant accelerator for sovereign AI. If the goal of 1.6 million dies by 2026 is realized, the tectonic shift from NVIDIA's dominance will become clear, at least in the Chinese domestic market.Bloomberg