GPU Export Controls Explained: What They Are, How They Work, and Who's Affected

What are GPU export controls and why do they exist?

GPU export controls are government restrictions on the transfer of high-performance graphics processing units and AI accelerators across national borders. They exist because GPUs are dual-use technology — originally designed for gaming graphics, they turned out to be the most efficient hardware for training large AI models. A cluster of GPUs powerful enough to train a frontier language model is also, in principle, capable of simulating nuclear weapons behavior, optimizing military drone swarm coordination, and running autonomous targeting algorithms.

The US Bureau of Industry and Security (BIS) administers the Export Administration Regulations (EAR), which classify products by their technical characteristics and restrict which destinations they can reach without a license. GPUs capable of training large AI models fall under ECCN 3A090, with controls applied based on Total Processing Performance (TPP) and memory interconnect bandwidth thresholds.

The strategic logic is explicit: if the US and its allies can prevent adversaries from accessing the compute infrastructure needed to train frontier AI at scale, those adversaries' AI development is constrained relative to the frontier — buying time for allied nations to build a durable technological lead. Whether the strategy succeeds is actively debated. What is not debated is that the rules are the primary instrument of US technology policy toward China in 2026 — and that China is pursuing a mirror-image strategy of its own.

The US export control timeline: from Huawei (2019) to the Pax Silica era (2026)

2019: Huawei is added to the BIS Entity List — its 220+ affiliates, chip design subsidiary HiSilicon, and any foreign product containing more than a de minimis share of US-origin technology are restricted. This established the template for supply chain weaponization: controlling not just the final product but the entire technology stack.

2020: SMIC (China's leading foundry) is listed, restricting equipment shipments that would advance its process nodes beyond 10nm.

2022 (October): BIS imposes the most sweeping controls since the Cold War — restricting advanced computing chips (A100, H100 class), chip-making equipment, and the software and technical data to develop them. ASML's EUV export restrictions to China are formalized under Dutch export policy.

2023 (October): BIS tightens compute thresholds, closing the A800/H800 workaround NVIDIA had created. Japan aligns its semiconductor equipment controls — 23 categories including equipment from Tokyo Electron and Kokusai Electric — restricting Chinese access to tools that would advance SMIC's process capabilities. The Netherlands aligns ASML's most advanced DUV controls.

2026 (current): The US-Taiwan Pax Silica Declaration (January 2026) formalizes deep supply chain integration, with $250 billion in Taiwanese semiconductor investments directed toward allied nations. South Korea joins the allied semiconductor framework. Controls continue to tighten quarterly.

China's counter-controls: gallium, germanium, graphite, and antimony

China is not merely the target of export controls — it is an active player in the same game, wielding its own chokepoints as a counter-strategy. Understanding the full bilateral control regime requires examining both sides.

Gallium and germanium (July/August 2023): China produces approximately 80% of the world's gallium and 60% of its germanium. Gallium is used in compound semiconductors (GaAs, GaN) essential for RF chips, power amplifiers, and advanced radar systems. Germanium is used in fiber optic systems, infrared optics, and certain transistors. China's export control — requiring a license for gallium and germanium exports — was explicitly framed as a response to US restrictions on ASML and chip equipment. The controls sent immediate supply shocks to Japanese, Korean, and European manufacturers who depend on Chinese gallium for compound semiconductor production.

Graphite (October 2023): China controls approximately 65% of global natural graphite production and nearly all commercial production of the synthetic graphite used in lithium-ion battery anodes — which underlie every AI server's power supply infrastructure. Controls on graphite exports affect battery supply chains that feed data center UPS systems and electric vehicle manufacturing.

Antimony (September 2023, tightened 2024): Antimony is used in flame retardants for electronics, semiconductor manufacturing chemicals, and certain military applications. China controls approximately 50% of global antimony production. Korean and Japanese electronics manufacturers are directly exposed.

The strategic logic mirrors the US approach precisely: China controls materials that the US-allied manufacturing base cannot easily replace, applying them as leverage in the same technology war. For South Korea, which relies on Chinese gallium for its compound semiconductor industry, and for Japan, whose specialty chemical cluster sources multiple controlled materials from China, these controls are not abstract geopolitical events — they are direct cost and availability pressures on existing production lines.

For investors monitoring Samsung, SK Hynix, Tokyo Electron, and Shin-Etsu: Chinese material export controls are a standing risk factor that affects input costs, qualification timelines for alternative sources, and supply chain resilience — independent of the Iran war helium and bromine disruptions that are simultaneously in play.

Which chips are controlled and what are the key thresholds?

BIS controls apply to chips exceeding specific Total Processing Performance (TPP, measured in tera-operations per second) and memory bandwidth thresholds. NVIDIA's H100, H200, and B200 (Blackwell) clearly exceed the thresholds and are restricted from export to China, Russia, and other controlled destinations without a license. NVIDIA's A800 and H800 derivative products — created to stay below the original thresholds — were closed by the October 2023 rule revision.

AMD's MI300X and subsequent Instinct AI accelerators exceed the thresholds. Intel's Gaudi series is controlled. Any chip manufactured at TSMC's N3 or below for AI compute is presumptively controlled under the manufacturing equipment restrictions that apply to TSMC's advanced nodes.

For Korean and Japanese companies: chips designed in Korea (Samsung's Exynos AI accelerators) or Japan (Rapidus's planned 2nm outputs) fall under ECCN 3A090 if they exceed the performance thresholds. The multi-jurisdictional compliance burden is the same for Samsung and SK Hynix as for NVIDIA — and for Japanese equipment makers who service the fabs that produce these chips, the same equipment controls apply to their maintenance and upgrade activities.

How to use AIChipMap to navigate the full bilateral control landscape

AIChipMap's export controls section covers all major active regimes from both sides — including the US BIS advanced computing rules, ASML/Netherlands EUV controls, Japan's 2023 equipment alignment, entity list actions against Huawei, SMIC, and YMTC, and critically, China's own counter-controls on gallium, germanium, graphite, and antimony.

Each regime page shows the current status, which supply chain nodes are targeted, which companies are affected, and a direct link to the source regulation. The bilateral coverage means you can see, in one view, both the controls that restrict access to AI chips from the US-allied side and the controls that restrict access to critical materials from the China side — mapping the full geopolitical exposure of your supply chain.

For Korean investors and supply chain teams monitoring Samsung (₩351,500) and SK Hynix (₩2,160,000): cross-reference their upstream material suppliers against China's gallium, germanium, and graphite controls. For Japanese users tracking Tokyo Electron (¥53,060) and Shin-Etsu Chemical (¥7,641): see which of their downstream customers are subject to US entity list restrictions and which of their material inputs are subject to Chinese export controls.

The goal is not just compliance — it is a complete map of the geopolitical chess board on which the AI chip supply chain operates.