The U.S. government's restrictions on exporting advanced AI chips to China and other countries have fundamentally reshaped how nations build artificial intelligence infrastructure. What started as a targeted ban on NVIDIA's A100 and H100 chips in October 2022 has evolved into a comprehensive export control regime that now covers over 120 countries, with different access tiers based on geopolitical alignment. This escalating series of restrictions is forcing technology leaders and governments worldwide to rethink their AI strategies and invest billions in building their own domestic chip capabilities.

How Has the U.S. Export Control Regime Evolved?

The U.S. government has tightened AI chip export controls through five distinct regulatory steps, each one closing loopholes that companies and foreign governments exploited in the previous version. Understanding this timeline reveals how quickly the geopolitical landscape around AI infrastructure has shifted.

• October 2022: NVIDIA's A100 and H100 chips were banned for export to China and Russia, targeting the dominant platforms for AI model training at the time.
• October 2023: Export rules expanded to close the A800/H800 loophole, after NVIDIA created China-specific variants with reduced interconnect bandwidth to stay below export thresholds.
• April 2024: The H20, a further reduced China-spec chip, was initially permitted as NVIDIA attempted to design a commercially useful chip that could be exported legally.
• January 2025: The Biden administration established a three-tier framework covering 120+ countries, with Tier 1 allies receiving unrestricted access, Tier 2 nations requiring licenses with compute caps, and Tier 3 countries including China and Russia facing outright bans.
• April 2026: The H20 was banned entirely, and Huawei's Ascend 910C was added to the restricted list, prompting NVIDIA to recognize a $5.5 billion inventory write-down.

This progression demonstrates a fundamental shift in how governments view AI infrastructure. Compute has become a geopolitical asset, and the question of which chips a country can purchase, in what quantities, and for which purposes now involves export license applications, national security reviews, and diplomatic negotiations.

Why Are Nations Building Their Own AI Infrastructure?

The export control escalation has triggered an unintended but predictable consequence: every country that can afford to do so is now investing heavily in sovereign AI infrastructure programs. Nations that view AI capability as a strategic interest cannot accept dependence on U.S. chip export policy as a constraint on their AI development.

Japan has committed to sovereign AI infrastructure development, with Fujitsu's MONAKA custom CPU designed for Japanese sovereign AI workloads. Notably, Japan's approach is cooperative rather than confrontational; the country is securing domestic chip design capability while maintaining access to NVIDIA's ecosystem through NVLink Fusion partnerships.

The Gulf states are pursuing a different strategy. The UAE's G42 and Saudi Arabia's HUMAIN are building hyperscale data centers with direct NVIDIA partnerships, and both countries have negotiated government-to-government arrangements that include U.S. oversight provisions in exchange for access to the highest-performance chips. Combined, these nations have announced AI infrastructure investments exceeding $100 billion.

India's IndiaAI Mission represents another model. The country is targeting 10,000+ GPU deployments through public and private sector coordination. Because India is classified as Tier 1 under U.S. export controls, it receives unrestricted access to advanced chips, giving it a structural advantage over Tier 2 competitors in AI infrastructure buildout.

China's situation is the most significant. With H100, B200, and H20 chips all now banned, Chinese AI development is proceeding on an alternative hardware stack centered on Huawei's Ascend 910B and 910C series. While independent benchmarks suggest a meaningful performance gap between Huawei's chips and NVIDIA's latest generations, the Chinese government's willingness to mandate domestic chip procurement for state-related AI projects provides Huawei with a captive market at scale.

What Do the EU's AI Regulations Mean for Global AI Development?

Beyond export controls, the European Union's AI Act introduces another layer of complexity through compute thresholds that trigger enhanced regulatory obligations. Models trained using more than 10^25 floating point operations, or FLOPs (a measure of computational work), are classified as "general-purpose AI models with systemic risk" and face mandatory adversarial testing, incident reporting, and transparency requirements.

To put this in practical terms, training a frontier-scale model comparable to GPT-4 on an NVIDIA H100 cluster for six months would cross this threshold. This creates a two-tier regulatory landscape: models below the 10^25 FLOP threshold face lighter regulation, while models above it face requirements comparable to pharmaceutical safety obligations.

For most enterprise organizations, this is not an immediate concern because few companies are training frontier-scale models. However, organizations deploying third-party models trained at this scale, including GPT-4 class models accessed via API, need to understand whether those models carry systemic risk classification under the AI Act, as this affects vendor due diligence requirements and liability exposure.

What Steps Should Technology Leaders Take Now?

For organizations operating across multiple jurisdictions, the geopolitical layer of AI infrastructure has become load-bearing. Technology leaders cannot treat AI infrastructure as a purely technical decision anymore. Here are the operational adjustments that experts recommend:

• Export License Review: Before ordering NVIDIA H100, B200, or equivalent chips for deployment in non-Tier 1 markets, involve legal counsel immediately. The license application process for Tier 2 markets can take 60 to 120 days and may include usage restrictions and audit rights.
• Map Your Compute Footprint to the AI Act: If you deploy AI to European Union users, understand which models in your stack have been trained at scales that approach the 10^25 FLOP threshold. This directly affects your vendor due diligence requirements and your own liability exposure under EU regulations.
• Evaluate Sovereign Cloud Options: AWS European Sovereign Cloud, Azure sovereign regions, and OVHcloud offer EU-resident compute powered by NVIDIA Blackwell chips, providing compliance-grade AI inference without requiring organizations to build on-premises infrastructure.
• Plan for Alternative Hardware in China: If you operate AI infrastructure in China, planning for Huawei Ascend-based alternatives is now a practical necessity rather than a contingency. The performance gap is real but shrinking, and the regulatory trajectory suggests it will not reverse.

The fracturing of the global AI stack is no longer a theoretical concern. It is an operational reality that affects chip procurement timelines, regulatory compliance, infrastructure investment decisions, and long-term technology strategy. Organizations that understand these geopolitical layers and plan accordingly will have a significant advantage over those that treat AI infrastructure as a purely technical problem.