Before the Storm: A Global Shock Triggered by a Chip Ban
On the night of April 17, 2025, at the VIP passage of Beijing Capital International Airport, NVIDIA CEO Jensen Huang appeared hurriedly in his signature leather jacket. This was his second secret visit to China in three months, with a low-key itinerary but hidden intentions—just 48 hours earlier, the U.S. Department of Commerce suddenly announced an “indefinite export control” on NVIDIA’s H20 chip designed specifically for China, causing a shockwave in the global semiconductor market: NVIDIA’s market value evaporated by over $50 billion in a single day, while stocks of companies like AMD and Intel plummeted, and the Chinese AI industry held an emergency meeting overnight.
The trigger for this storm was the H20 chip. As a “special version” product created by NVIDIA to circumvent U.S. export restrictions, the H20’s performance has been cut down to 15% of the flagship H100 chip, yet it remains the most powerful AI inference chip that Chinese companies can legally obtain. Data shows that in the first three months of 2025, the procurement amount for H20 ordered by Chinese cloud computing giants reached a staggering $16 billion, with ByteDance alone accounting for 30% of the orders. However, the U.S. government’s sudden policy shift caught all participants off guard: on April 9, the White House required that H20 exports be approved on a case-by-case basis, with a default assumption of non-approval. This policy translated into a $5.5 billion loss provision in NVIDIA’s financial report, covering inventory backlog, procurement defaults, and other chain reactions.
Computing Power Hunger: The “Arms Race” of Chinese AI Large Models
The explosive growth of domestic AI large models in China is pushing the demand for computing power to unprecedented heights. In 2024, the domestic large model DeepSeek-V3 achieved performance comparable to GPT-4o using only 2048 GPUs through a mixed expert architecture and algorithm optimization, with training costs reduced to $5 million, only 5%-10% of that of international peers. However, this efficiency has not alleviated the anxiety over computing power—the real-time computing resources required for large model inference still heavily rely on the stable supply of NVIDIA GPUs.
“Our computing power pool is like a well in the desert; every drop is a matter of life and death,” admitted the head of an AI lab at a leading internet company. According to internal data, in the first quarter of 2025, the computing power gap for leading AI companies in China generally exceeded 40%, with some star applications forced to delay their launch due to insufficient GPUs. This anxiety has directly translated into a frenzy of stockpiling H20 chips: the price of an 8-card H20 server skyrocketed to 1.4 million yuan, 30% more expensive than the domestic Ascend 910B solution, yet its inference efficiency still maintains a 2-4 times advantage.
Jensen Huang’s Gamble in China: A Moat of 4,000 Employees and 1.5 Million Developers
Huang’s choice to visit China at this time is no coincidence. NVIDIA has been established in China for 25 years, with 4,000 local employees building a complete system from R&D to sales, and its turnover rate of 0.9% is even lower than that of its Silicon Valley headquarters. More critically, there is an ecological barrier: China has the world’s largest CUDA developer community (1.5 million programmers) and 3,000 AI startups, which constitute NVIDIA’s indispensable moat.
However, policy risks are eroding this advantage. The U.S. Department of Commerce’s erratic chip controls on China have put NVIDIA in a “tightrope” dilemma: it must meet the U.S. government’s technology blockade requirements while preserving its $17.1 billion annual revenue from the Chinese market (which accounts for 13% of global revenue). Huang repeatedly emphasized in closed-door meetings that “the Chinese market is irreplaceable” and revealed that he is secretly developing an improved version of the H20 that complies with the new regulations, attempting to find a balance between performance and compliance.
Domestic Breakthrough: A Qualitative Change from “Usable” to “User-Friendly”
The U.S. blockade unexpectedly activated China’s semiconductor “backup plan.” Huawei’s Ascend 910B chip has achieved 80% of H20’s inference performance, and its unique dual-chip packaging technology has improved the yield rate to 92%; Cambricon’s Siyuan 590 has even surpassed NVIDIA’s comparable products in unit power consumption performance during autonomous driving scenario tests. Even more surprising is the breakthrough in the software ecosystem: Huawei’s CANN framework is now compatible with 98% of CUDA interfaces, reducing the migration cost for developers by 70%.
This breakthrough has formed a multi-faceted offensive:
Hardware Level: Kunlun chips have captured 15% of the domestic cloud computing market, and Alibaba’s Tianshu Lingguang 800 has established an advantage in image processing;
Architectural Innovation: Tianxu Zhixin has achieved a 30% cost reduction in specific scenario inference tasks through heterogeneous computing optimization;
Policy Support: Shanghai plans to build a cluster of 625,000 Ascend 910B chips by 2027, with a market scale exceeding 40 billion yuan.
The Future Battle: Innovation Catalyst Under the Technological Iron Curtain
The endgame of this chip game may be hidden in two sets of data: China’s semiconductor imports are expected to decrease by 12% year-on-year in 2025, but R&D investment is expected to grow by 28% year-on-year; the proportion of overseas revenue for U.S. semiconductor companies has dropped from 58% to 49%, while inventory turnover days have increased from 85 days to 112 days. When the domestic 14nm production line achieves a 70% self-sufficiency rate, and quantum chip prototypes break through 1,000 qubits, the technological iron curtain may instead become a catalyst for innovation.
Jensen Huang’s 48-hour trip to Beijing serves as a metaphor for this era—on the track where AI reshapes the world, no one can monopolize the throne. As Silicon Valley venture capitalist Sarah Guo said, “Today we impose limits on chips, and tomorrow we may lose the right to define AI ethics.” In this competition concerning the future, the victor may not be the one who runs the fastest, but the one who adapts best to change.
The Awakening Moment of Computing Power Sovereignty
The $16 billion H20 stockpiling frenzy is both a microcosm of the “weaning pains” of China’s AI industry and a catalyst for the rise of domestic computing power. As NVIDIA’s leather-clad leader rushes through Beijing at night, and Huawei’s Ascend lights up the starry sky of domestic computing power in data centers, this game has long transcended commercial competition, becoming a necessary question of national technology strategy. Perhaps as Huang said, “The greatest danger is not the strength of the opponent, but the arrogance of self-limitation.” In the vast sea of computing power, Chinese AI is writing its own answers.
(Data sources: Tencent News, China.com, Sina Finance, Eastmoney.com, and official disclosures from NVIDIA and Huawei)