A data center in Zhongguancun, Beijing, tests a newly arrived cluster of NVIDIA H20 chips. Performance reports indicate: single card FP32 computing power of 9.2 TFLOPS, which is only 32% of the A100 chip from three years ago. This seemingly compromised product, however, writes an intriguing chapter in the history of global semiconductors.
1. The Geopolitical Code in Technical Parameters
According to the MLCommons 2025Q2 benchmark, a clear technological gap has formed between the H20 and its competitors. This “customized” solution exhibits significant asymmetrical characteristics: in terms of FP16 precision supporting large model training, the H20’s performance only reaches 53% of the Ascend 910B; however, in the FP32 general computing domain preferred by data centers, its computing power slightly exceeds domestic alternatives. This precise performance balance down to the instruction set aligns with the U.S. Department of Commerce’s regulatory logic of ‘positioning without decoupling.’ 2. Corporate Choices in the Cost Maze
A procurement list from an autonomous driving company in Beijing reveals the rational game at the market end. Although the absolute cost-performance ratio of the H20 is not advantageous, for companies urgently needing to maintain their computing power lifeline, its “plug-and-play” ecosystem compatibility remains tempting. A CTO from an e-commerce platform admitted: “The cost of rewriting 2 million lines of CUDA code could be more fatal than the hardware price difference.” 3. Strategic Depth in Offensive and Defensive Shifts
Data from the Semiconductor Industry Association of America shows that in the first half of 2025, the share of NVIDIA’s data center revenue from China plummeted from 22% in 2024 to 9%. Behind this cliff-like decline lie two parallel evolving storylines:
1. The Leapfrog Effect of Alternative Technologies
The latest Cambrian Shiyuan 590 chip has achieved an energy efficiency ratio in recommendation algorithm scenarios that is 217% of the H20, while Baidu’s self-developed Kunlun chip generation three exceeds the H20’s throughput in natural language processing tasks by 40%. These breakthroughs in vertical fields are dismantling the traditional GPU moat.
2. The Butterfly Effect of Capacity Reconstruction
The expansion of TSMC’s 16nm production line in Nanjing has allowed the monthly production capacity of the Ascend 910B to exceed 200,000 units. Meanwhile, the hardware “fuse” design mandated by the U.S. Department of Commerce is forcing Chinese customers to accelerate their transition to distributed computing architectures. 4. The Chaotic Game Before Dawn
In this war of attrition with no winners, three paradoxes are emerging:
The Diminishing Marginal Returns of Technological Blockades: When the regulatory threshold is set at 15 TFLOPS, Chinese companies are building new heterogeneous computing paradigms below this critical point.
The Ecological Backlash of Market Fragmentation: The fragmentation of the CUDA ecosystem is giving rise to the emergence of OpenCL and Tsinghua’s computing framework.
The Innovative Distortion of Regulatory Arbitrage: NVIDIA has established a “Compliance Architecture Department” that spends $320 million annually on chip neutering designs.
As MIT Technology Review states: “As the chip war enters its 2.0 phase, the key to victory has shifted from nanometer line width to the ability to reconstruct architectural philosophy.” In this century-long game, the H20 acts like a prism: the U.S. sees it as a buffer to delay the opponent’s development, China sees it as a timer for independent innovation, while NVIDIA calculates the minimum survival cost to weather the cyclical storm. As all three parties search for suboptimal solutions in the computing power maze, the true disruptor may be born in the blind spots of existing rules.
Follow me↓ I will share my thoughts and insights from time to time