Introduction: A “Shadow War” Changing the Global Tech Landscape
In May 2025, in a laboratory at a tech park in Shenzhen, engineer Xiao Li is testing an AI server equipped with the Ascend 910B chip—this device achieves nearly the same inference performance as NVIDIA’s H20 chip at one-third of the cost. At the same time, across the ocean at NVIDIA’s headquarters, Jensen Huang frowns at the report showing that revenue from China accounts for 13% of total income.
This is not merely a technical competition, but a life-and-death game concerning global AI hegemony. As U.S. policies clash with market dynamics, who will win the AI dominance of the next decade?
1. The Rise of Domestic Chips: From “Survival in the Cracks” to “Direct Confrontation”
1. Technological Breakthroughs: Performance Gaps Narrowed to “Millimeters”
– Dual Breakthroughs in Training and Inference: Huawei’s Ascend 910B achieves 148 TFLOPS of computing power at FP16 precision, close to NVIDIA’s H20 at 150 TFLOPS, and supports full-stack AI frameworks (such as MindSpore), reducing developer migration costs by 40%.
– Impressive Cluster Capabilities: The Moores Threads Kilo Intelligent Computing Cluster has a training accuracy error of only 1%, with computing power utilization comparable to the A100 cluster, already applied in multimodal large model training.
2. Policy Dividends: The “Accelerating Engine” for Domestic Substitution
– The Pressure of the Xinchuang Project: The goal of 70% localization of IT systems in state-owned enterprises by 2025 has propelled Huawei’s Ascend to increase its market share in government servers from 15% to 45%.
– Autonomous and Controllable Supply Chains: Yangtze Memory’s 3D NAND flash has entered Apple’s supply chain, and SMIC’s 28nm process yield has surpassed 85%, completely breaking the “bottleneck” dilemma.
3. Ecological Breakthroughs: From “Single Point Breakthrough” to “Full Chain Closed Loop”
– CUDA Alternatives: Huawei’s CANN framework has added over 80 integrated operators and more than 100 API interfaces, with the Llama separated deployment scheme improving large model inference efficiency by 30%.
– Developer Competition: The Ascend ecosystem has attracted over 300 enterprises, with training costs 35% lower than NVIDIA’s solutions and code migration cycles shortened by 50%.
Key Turning Point: In Q1 2024, the market share of domestic AI chips surged from 10% to 25%, while NVIDIA’s revenue in China plummeted by 42% year-on-year.
2. NVIDIA’s “Prisoner’s Dilemma”: Wanting Profit While Fearing Exclusion
1. The “Special Supply Chip Trap” Under Policy Red Lines
– The Embarrassment of the Castrated H20: Compared to the H100, the H20 has 41% fewer GPU cores and a 28% drop in FP16 performance, yet its price is still 20% higher than domestic chips.
– Capacity Allocation Conflicts: The Arizona factory in the U.S. allocates 70% of its capacity to North America, leaving only 10% for special supply chips to China, raising concerns among Wall Street analysts about “strategic imbalance”.
2. The “Cake” of the Chinese Market That Cannot Be Lost
– The Value of Data and Scenarios: China generates over 2EB of AI training data daily, three times that of the U.S.; losing this data would weaken NVIDIA’s technological iteration capabilities.
– The Backlash of Commercial Interests: In 2024, NVIDIA’s revenue in China reached $17.1 billion (accounting for 13% of total revenue); losing this market could result in a short-term gap of up to $18 billion.
3. The “Trust Crisis” in the Global Supply Chain
– TSMC’s Wavering: To avoid U.S. sanctions, TSMC’s Nanjing plant is accelerating the expansion of CoWoS packaging capacity, but the yield has dropped from 95% to 82%.
– The “De-Americanization” of European Allies: STMicroelectronics and GlobalFoundries are collaborating to build factories, focusing on automotive chips to avoid U.S. technology, leading to a 12% shrinkage in NVIDIA’s European market share.
🔥 The Fatal Paradox: The more NVIDIA tries to balance the Chinese and American markets, the more it risks being seen as an “unreliable partner” by both sides—this tearing between policy red lines and commercial interests is shaking the foundation of its global hegemony.
3. The Ultimate Showdown: The Struggle for Technical Standards and Ecological Discourse Power
1. The Division of Technical Routes
– Breakthroughs in Photonic Chips: The theoretical computing power of silicon-based photonic integrated chips from the Chinese Academy of Sciences is 100 times that of GPUs; if mass production is achieved by 2030, it will reshape the AI computing power landscape.
– The Independence of EDA Tools: BGI’s 28nm process covers the entire flow, and AI-assisted design tools will shorten the R&D cycle of domestic chips by 30%.
2. The Multipolarization of Geoeconomics
– The “Neutral Dividend” of Southeast Asia: Penang, Malaysia, accounts for 13% of global packaging capacity, becoming a buffer zone in the U.S.-China game.
– The “Middle Path” of Europe: Infineon in Germany has signed a cooperation agreement with Huawei to jointly develop industrial AI chips.
3. The “Offensive and Defensive Battle” of Developer Ecosystems
– Cracks in CUDA: After the DeepSeek open-source model was adapted for the Ascend 910B, training efficiency surpassed that of the A100, prompting NVIDIA to urgently lobby the U.S. government.
– The Defection of the Open Source Community: The PyTorch community has now supported Ascend chips, with the volume of code submissions from developers increasing by 210% year-on-year.
Conclusion: A War Without Winners, But China Must Win
As NVIDIA’s H20 chip runs on servers in China, and Huawei’s Ascend supports a smart city project in a Southeast Asian country, this game has long transcended commercial competition—it concerns the reconstruction of the global tech order.
For domestic chips, the real test lies not in the laboratory, but in mass production yield, ecological stickiness, and continuous innovation; for NVIDIA, finding a third path between “profits from the Chinese market” and “compliance with U.S. policies” will determine whether it can maintain its position as the AI computing power leader.
Historical experience tells us: technological autonomy is always the ultimate trump card in great power games.
Interactive Topic: “Do you think domestic chips can end NVIDIA’s hegemony by 2030?” Encourage comments and discussions.
Data Sources: Comprehensive reports from China Securities Journal, IDC, Reuters, etc.