Export Controls and the Geopolitics of GPU Supply

Artificial intelligence capabilities at the frontier;training and running the largest models;require massive computational resources. Those resources are currently concentrated in a handful of chip designs from NVIDIA, manufactured in Taiwan.

The United States government concluded, around 2022, that allowing adversary states to access frontier AI compute was equivalent to allowing them to acquire advanced weapons systems. The argument is straightforward: AI improves military targeting, signals intelligence, cyber capabilities, and autonomous systems.

Nations with access to frontier AI compute have military and economic advantages over those without. The export control architecture that followed is designed to prevent China;and other designated countries;from building or accessing the compute needed to train frontier AI models. This is not marginal policy: it is a deliberate attempt to create and maintain a permanent AI capability gap between the United States and its principal strategic competitor.

The US Department of Commerce's Bureau of Industry and Security (BIS) issued sweeping AI chip export controls in October 2022. The initial rules targeted chips above specific performance thresholds.

The H100, A100, and all comparable advanced AI accelerators were placed under licence requirements for export to China, Hong Kong, and Macau. NVIDIA responded by designing export-controlled variants: the A800, H800, and later L20 and H20;chips with reduced interconnect bandwidth or performance to fall below the regulated thresholds.

The Biden administration updated the rules in October 2023 to close these workarounds. A further update in 2024 created a three-tier system: close allies (Tier 1) receive unlimited access; most countries (Tier 2) receive controlled access with verification requirements; arms-embargoed countries and adversaries (Tier 3) receive no access. As of early 2026, approximately 40 countries face restrictions on advanced GPU imports. The Trump administration modified the framework again in 2025, and further updates are expected.

The export controls have had concrete effects on both sides of the restriction. Chinese AI companies;Baidu, Alibaba, ByteDance, and hundreds of smaller firms;faced an acute supply crisis in 2023.

Those with existing H100 and A100 inventory saw their assets appreciate dramatically as no replacement supply was available. Chinese AI labs have responded by accelerating domestic chip development: Huawei's Ascend 910B has been widely reported as an H100 equivalent, though performance claims remain disputed.

SMIC;China's most advanced domestic foundry;operates at 7nm process nodes using older DUV lithography rather than EUV, limiting its ability to produce chips at the scale and yield that TSMC achieves for NVIDIA. The capability gap is real but narrowing;domestic Chinese AI development has accelerated precisely because of the supply restriction. BIS enforcement actions have targeted several companies for re-exporting controlled chips via third countries. Singapore and the UAE have been identified as transshipment risks.

Export controls have accelerated investment in sovereign AI compute;national infrastructure built or funded by governments to ensure independent access to AI capability. France has committed €109M to a national AI compute infrastructure.

Saudi Arabia's Public Investment Fund invested $40B in AI infrastructure through the HUMAIN initiative in 2025. The UAE's G42 has built significant GPU clusters.

India is building a national AI compute programme targeting 10,000 GPUs. The UK's AI Compute Taskforce identified a significant shortfall in domestic compute capacity relative to national research ambitions. The distinction between Tier 1 countries (with unrestricted GPU access) and Tier 2 countries (with verification requirements) matters for sovereign programmes. A Tier 2 government procuring advanced GPUs must demonstrate to BIS that the hardware will not be re-exported and will be used for approved purposes;a compliance overhead that smaller nations find burdensome.

GPU infrastructure operators must navigate an export control compliance landscape that changes annually and carries significant penalty risk. Selling compute services to restricted-country customers;even inadvertently;can result in export violations.

Cloud providers must implement customer due diligence (end-user certifications, know-your-customer checks) for GPU access. For investors, export controls create structural opportunities.

Sovereign AI programmes are committed capital;governments buying GPU clusters regardless of commercial return expectations. These are long-term, credit-backed contracts that underpin infrastructure investment theses. Operators positioned in Tier 1 jurisdictions (UK, EU, Japan, Australia, Canada) can bid into sovereign AI procurement without the compliance overhead that Tier 2 status would require. For sovereign compute strategy, regulatory navigation, and government engagement support, get in touch at disintermediate.global/contact.

BIS export controls (2022-2026) restrict advanced AI chip sales to China and ~40 other countries; rules have been updated multiple times and continue to evolve

NVIDIA designed export-controlled variants (H800, L20, H20) but subsequent rule updates narrowed these workarounds

Chinese AI companies face real capability constraints; Huawei Ascend 910B is a domestic alternative with disputed performance claims

Sovereign AI programmes (Saudi Arabia $40B, France €109M, India 10,000 GPUs) are accelerating in response to supply risk

Cloud providers must implement export control compliance (end-user certifications, KYC) for GPU services;a meaningful operational requirement