The AI Chip War: How Export Controls Reshaped the Global Semiconductor Industry in 2026
The global semiconductor industry is undergoing its most significant restructuring since the 1980s US-Japan trade conflicts. New US export controls on advanced AI chips — first imposed in October 2022, tightened repeatedly through 2025, and codified in the CHIPS and Science Act follow-on legislation of early 2026 — have created a bifurcated market: one for the West and its allies, and another for China and the Global South.
Here is what the AI chip war looks like in 2026, how companies are adapting, and what the long-term consequences are for the industry.
The Current Regulatory Landscape
The US export control regime as of May 2026 includes three main layers:
Layer 1: Performance Density Thresholds. Any chip with a total processing power above 200 TOPS (trillion operations per second) and a performance density above 10 TOPS/mm² is restricted from export to China without a license. This covers NVIDIA's H100, H200, B100, and B200 GPUs, plus AMD's MI300X and Intel's Gaudi 3.
Layer 2: Entity List Expansion. Over 350 Chinese companies are now on the Entity List, including Huawei's entire semiconductor design chain, SMIC's advanced fabrication facilities, and all companies affiliated with the Chinese Academy of Sciences in AI-related fields.
Layer 3: Foreign Direct Product Rule (FDPR). Any chip made anywhere in the world that uses US-origin technology — including chips fabricated in Taiwan, South Korea, or the Netherlands — falls under US export controls. This extraterritorial reach is the most contested element of the regime.
The FDPR is what makes the US controls so powerful — and so controversial. ASML's extreme ultraviolet (EUV) lithography machines, which contain US-origin components, cannot be exported to China even though they are made in the Netherlands. TSMC, which uses US-origin design tools and equipment, cannot manufacture advanced chips for Chinese customers.
China's Response: Indigenous AI Chips
China has responded with unprecedented investment in domestic AI chip development. The "Semiconductor III" national program, announced in March 2026, allocates ¥1.2 trillion ($166 billion) over five years to build a self-sufficient AI chip ecosystem.
| Chinese AI Chip | Company | Node | Performance | Status |
|---|---|---|---|---|
| Ascend 910C | Huawei | 7nm (SMIC N+2) | ~300 TOPS | Mass production |
| Sophon SG2044 | Bitmain | 12nm | ~128 TOPS | Sampling |
| T-Head TH-600 | Alibaba | 7nm (SMIC N+2) | ~256 TOPS | Mass production |
| Cambricon MLU370 | Cambricon | 7nm (TSMC - stockpiled) | ~196 TOPS | Limited availability |
The problem for China is that SMIC's 7nm process (N+2) has yield rates estimated at 40-50%, compared to TSMC's 90%+ on equivalent nodes. This means Chinese AI chips cost roughly twice as much to produce as their Western equivalents — even before accounting for the massive government subsidies that support the industry.
Huawei's Ascend 910C, launched in late 2025, is the most credible Chinese AI chip. In inference workloads (running already-trained models), it achieves approximately 75% of the performance of an NVIDIA H100. In training workloads, the gap widens to roughly 50% due to immature software ecosystems.
The Gray Market Problem
Despite the controls, NVIDIA's high-end chips are reaching China through a network of intermediaries. Industry estimates suggest that 20,000 to 40,000 H100-equivalent GPUs per quarter are smuggled through third countries — primarily Singapore, Malaysia, and Vietnam.
The typical route: a shell company in Singapore places an order through a Malaysia-based distributor. The chips are shipped to a "testing facility" in Vietnam, where paperwork shows they are used for domestic AI workloads. In practice, many are trucked across the land border into China's Yunnan province.
US Customs and Border Protection has stepped up enforcement, but the volume of global semiconductor trade — hundreds of millions of chips per year — makes comprehensive inspection impossible. The US is now pushing for a "chip fingerprinting" system that would allow each restricted GPU to be tracked through its supply chain. NVIDIA has resisted, citing customer privacy concerns.
Impact on NVIDIA
NVIDIA is the company most affected by the controls. China represented approximately 20% of NVIDIA's data center revenue before the first controls were imposed. That figure has dropped to 5% in fiscal 2026 — a loss of roughly $4 billion in annual revenue.
But NVIDIA has more than compensated by selling to the rest of the world. Data center revenue reached a record $85 billion in fiscal 2025 and is projected to exceed $110 billion in fiscal 2026. The company's market capitalization has grown from $1 trillion in 2023 to over $3.5 trillion in May 2026.
NVIDIA's strategy has been to design China-specific chips that fall below the export control thresholds. The H20, introduced in 2024, is an H100 with reduced performance that just barely stays under the 200 TOPS limit. It sells at a 40% discount to the H100 and has become NVIDIA's best-selling product in China — a legal workaround that generates billions in otherwise lost revenue.
The Long-Term Consequences
The AI chip war is creating two separate semiconductor ecosystems. In the short term (2025-2027), the US maintains a clear advantage. In the medium term (2028-2031), China's domestic chips will close the gap to perhaps 80-90% of Western performance.
Whether that gap matters depends on the application. For military AI — where absolute performance matters less than reliability and supply chain security — China's indigenous chips may be sufficient. For cutting-edge AI research, including foundation model training, the US advantage remains decisive.
The biggest unknown is whether Taiwan's semiconductor industry remains intact. If China were to blockade or invade Taiwan, TSMC's advanced fabs — which produce the world's most advanced chips — would be disrupted or destroyed. That scenario, however remote, is the nightmare that keeps semiconductor executives awake at night.
FAQ
Q: Can I buy an NVIDIA H100 in China as an individual?
A: Technically no. In practice, gray market availability exists but prices are 3-5x the US retail price of $30,000.
Q: Is Huawei's Ascend chip competitive with NVIDIA?
A: In inference workloads, it is roughly 75% as capable. In training, roughly 50%. The software ecosystem (CUDA vs. Huawei's CANN) is the bigger gap.
Q: Will the export controls be relaxed under the new trade deal?
A: Partially. Mid-range chips (14-28nm) will be decontrolled. Advanced AI chips (7nm and below) remain restricted.
Q: How does this affect cloud computing prices?
A: Globally, GPU cloud prices are falling as supply increases. In China, GPU cloud prices are 2-3x global averages due to scarcity.
Q: What happens if SMIC achieves 5nm production?
A: SMIC is believed to be 3-5 years away from 5nm capability without EUV lithography. The company has developed a multi-patterning workaround that extends 7nm, but 5nm requires EUV.
Key Takeaways
- US export controls have created a bifurcated semiconductor market: advanced chips for the West and its allies, indigenous alternatives for China.
- China's domestic AI chips — led by Huawei's Ascend 910C — have reached 50-75% of NVIDIA's performance but suffer from poor yields and immature software ecosystems.
- A thriving gray market moves 20,000-40,000 restricted GPUs to China quarterly through Singapore, Malaysia, and Vietnam.
- NVIDIA has adapted by designing China-specific chips that legally comply with export thresholds, recovering billions in otherwise lost revenue.
- Two semiconductor ecosystems are emerging — a US-led system and a Chinese system — with different performance levels, supply chains, and strategic implications.








