China AI Export Controls 2026: New GPU Restrictions, Domestic Chip Progress, and the Global Semiconductor Realignment
The US tightened AI semiconductor export controls again in 2026. New restrictions target memory bandwidth and advanced packaging. China's domestic chip industry shows real progress with SMIC 5nm and Huawei Ascend. The global semiconductor map is redrawing.
The US-China technology competition entered a new phase in 2026. The Biden-era export controls have been tightened, expanded, and in some cases restructured under the Trump administration's second term. The result is a semiconductor landscape that looks fundamentally different from 2023 — new restrictions targeting specific technical parameters, a Chinese domestic chip industry that is further along than most Western analysts predicted, and a global supply chain that is rapidly bifurcating.
The New Restriction Layer: Memory Bandwidth and Advanced Packaging
The Biden administration's October 2022 and October 2023 export controls primarily targeted compute performance — limiting the total processing power of chips that could be exported to China. These rules were effective for their time but had loopholes. Companies like NVIDIA could tune chips (the A800, H800, and H20 series) to stay just under the compute limits.
The Trump administration's January 2026 Executive Order on AI Semiconductor Security closed many of those loopholes by introducing two new restriction parameters.
<<<BOLD>>>Memory bandwidth<<<BOLDEND>>> became a controlled parameter. Any chip with memory bandwidth exceeding 1.5 TB/s is subject to export licensing requirements when destined for China or other restricted entities. This targeted NVIDIA's H100/B100-class chips directly — their HBM3 memory interfaces exceed 2 TB/s bandwidth. The H20 chip, which NVIDIA specifically designed for the China market, has a memory bandwidth around 1.2 TB/s — just under the limit — making it technically compliant but strategically inadequate for frontier AI training.
<<<BOLD>>>Advanced packaging<<<BOLDEND>>> was the second new target. The January 2026 rule restricts the export of chiplet-based designs where multiple dies are interconnected via advanced packaging technologies (CoWoS, InFO, EMIB) into a single logical chip with total compute capacity exceeding the threshold. This closes the loophole where a company could export multiple sub-threshold chips for assembly in China. Now the entire aggregated system is evaluated, not just the individual die.
The practical effect: China cannot legally import any chip capable of training frontier AI models. The H20, which was borderline adequate for fine-tuning and inference of 2024-era models, is not sufficient for training GPT-5-scale models or their equivalents. Chinese AI companies must either use stacked domestic chips (less efficient), access cloud instances through proxies (high risk), or use the chips they already have in inventory.
SMIC's 5nm Reality
The most significant development on the Chinese side is SMIC's progress on 5nm-class manufacturing. In April 2026, SMIC confirmed that its N+3 process (equivalent to TSMC's N5/node 5nm-class) has entered risk production at its Beijing fab. The yield data has not been independently verified, but third-party analysis of chips pulled from commercial products suggests yields of approximately 35% to 45% — compared to TSMC's 5nm yields of above 90%.
The yield gap matters. At 35% to 45% yield, SMIC's 5nm is feasible for Huawei's Ascend AI accelerators and a limited run of Kirin mobile processors, but it's not economically viable for most commercial applications. The cost per good die is roughly 2.5 times higher than TSMC's 5nm, even accounting for China's lower labor costs and government subsidies.
The equipment question remains critical. SMIC's 5nm capability depends on deep ultraviolet lithography (DUV) from ASML — specifically the TWINSCAN NXT:2050i immersion system. China has stockpiled DUV systems before export controls took effect and is believed to have between 60 and 80 DUV units in inventory. But the company cannot access extreme ultraviolet lithography (EUV), which is under a Dutch export ban. Leading-edge logic requires EUV for the most critical layers at 5nm and below. SMIC is using multi-patterning on DUV to achieve 5nm-class features, a technique that is technically viable but dramatically increases process complexity and defect rates.
Huawei Ascend: The Domestic AI Chip Standard
Huawei's Ascend series has become the de facto standard for domestic AI compute in China. The Ascend 910C, released in late 2025, is the current flagship. Benchmarks show it delivers approximately 65% of the FP16 tensor performance of an NVIDIA H100 for inference workloads, and about 40% for training. The gap is narrowing — the 910C is a meaningful improvement over the 910B's 50% inference performance relative to H100.
The software stack remains the weak point. Huawei's CANN (Compute Architecture for Neural Networks) and MindSpore framework have improved significantly. In a June 2026 evaluation by China's Beijing AI Research Institute, standard models (Llama 4 70B, Qwen 3 72B) deployed on Ascend 910C clusters achieved 78% to 91% of the training throughput of equivalent H100 clusters, after optimization. The gap is acceptable for domestic deployment but not competitive for frontier research.
The installed base of Ascend chips in China is estimated at roughly 600,000 units as of June 2026 — a number that has doubled since June 2025. Huawei's cloud division, Huawei Cloud, offers Ascend-backed GPU instances at roughly 70% of the cost of NVIDIA H100 instances through Chinese cloud providers, though the per-chip performance is lower.
The Global Realignment
The export controls have created a bifurcated global semiconductor market that is becoming semi-permanent.
<<<BOLD>>>Tier 1 (US and allies):<<<BOLDEND>>> Full access to TSMC 3nm and 2nm, NVIDIA B200 and next-generation Blackwell Ultra GPUs, EUV lithography, advanced packaging. US companies and allied nations (Japan, South Korea, Netherlands, Taiwan) operate in an unrestricted technology environment.
<<<BOLD>>>Tier 2 (neutral and unaligned):<<<BOLDEND>>> Middle Eastern countries, India, Brazil, Southeast Asia. They can access most AI chips with some licensing requirements. UAE-based G42, for example, has access to NVIDIA H100s under a technology assurance agreement with the US government. The access is not unrestricted but is far broader than Tier 3.
<<<BOLD>>>Tier 3 (restricted):<<<BOLDEND>>> China, Russia, and entities on the Entity List. They are cut off from frontier AI chips, EUV lithography, advanced packaging services, and semiconductor design tools for leading-edge nodes. Domestic alternatives exist but at lower performance, higher cost, and with significant software ecosystem gaps.
The biggest unknown is whether Tier 2 countries will remain in their lane. India has been making aggressive moves toward domestic chip manufacturing, announcing a $15 billion semiconductor incentive package in March 2026. Saudi Arabia is investing heavily in AI compute infrastructure. If the Tier 2 countries start building significant independent AI compute capacity, the US-China competition framework may not hold.
