Compute has become a geopolitical asset. That sentence would have sounded abstract in 2020. In 2026 it is operationally concrete: which chips you can buy, in which quantities, to deploy in which countries, is now a question that involves export license applications, national security reviews, and bilateral diplomatic negotiations. Technology leaders operating globally cannot treat AI infrastructure as a purely technical decision. The geopolitical layer is now load-bearing.
This post covers the U.S. export control regime and its evolution, what the EU AI Act’s compute thresholds mean in practice, the national AI infrastructure programs that are reshaping chip demand, and what enterprise technology leaders operating across multiple jurisdictions actually need to do about it.
Figure 1: Global AI chip access tiers under U.S. export control regime (2026). Access is stratified across three tiers, with national AI programs emerging across Tier 1 and Tier 2 markets.
The Export Control Regime — A Five-Year Escalation
The U.S. government’s AI chip export control regime has escalated in five distinct steps since October 2022. Each step has closed loopholes that chip vendors and foreign buyers found in the previous version.
Figure 2: U.S. AI chip export control escalation timeline. Each regulation has closed the loopholes created by the previous one, with the April 2026 H20 ban representing the most aggressive restriction to date.
- October 2022: NVIDIA A100 and H100 banned for export to China and Russia. This was the opening move — targeting the chips that were already the dominant AI training platform.
- October 2023: Expanded controls close the A800/H800 loophole. NVIDIA had created China-specific variants of their flagship chips with reduced interconnect bandwidth to stay below the export threshold. The 2023 rules updated the threshold criteria to eliminate this workaround.
- April 2024: The H20 — a further reduced China-spec chip — was initially permitted. NVIDIA attempted to design a chip that could be exported while still being commercially useful. This permission was temporary.
- January 2025: The Biden administration’s AI Diffusion Rule establishes a three-tier framework covering 120+ countries. Tier 1 (close U.S. allies): unrestricted. Tier 2 (most of the world): licensed with compute caps. Tier 3 (China, Russia, and others): banned.
- April 2026: H20 banned. Huawei Ascend 910C added to the restricted list. NVIDIA recognized a $5.5 billion inventory write-down. This is the most aggressive restriction to date — eliminating even reduced-capability chips designed specifically for the Chinese market.
What the EU AI Act Compute Thresholds Mean in Practice
The EU AI Act establishes regulatory obligations that kick in at specific compute thresholds. The key number: models trained using more than 10^25 FLOPs (floating point operations) are classified as ‘general-purpose AI models with systemic risk’ and face enhanced obligations including mandatory adversarial testing, incident reporting, and transparency requirements.
Mapping this to hardware: training a frontier-scale model (GPT-4 class or larger) on an NVIDIA H100 cluster for six months crosses this threshold. Organizations training models at this scale within the EU — or deploying models trained at this scale to EU users — need legal and compliance functions that understand this framework.
The EU AI Act’s compute threshold creates a two-tier model landscape: models below 10^25 FLOPs face lighter regulation; models above face requirements comparable to pharmaceutical safety obligations. Where you train and where you deploy are both regulated questions.
For most enterprise organizations, this is not an immediate concern — few enterprises are training frontier models. But for organizations building or deploying models trained by third parties (including GPT-4 class models via API), understanding whether those models carry systemic risk classification under the AI Act affects your vendor due diligence requirements.
National AI Programs — A New Demand Driver
The export control regime has had an unintended structural consequence: it has accelerated national AI infrastructure programs in every Tier 1 and Tier 2 country that can afford to build them. The strategic calculation is simple — nations that believe AI capability is a sovereign interest cannot accept dependence on U.S. chip export policy as a constraint on their AI development.
Japan — Fujitsu MONAKA and the NVIDIA Partnership
Japan has committed to sovereign AI infrastructure development. Fujitsu’s MONAKA custom CPU is designed for Japanese sovereign AI workloads and is a NVLink Fusion partner — meaning Japanese sovereign AI infrastructure integrates with NVIDIA’s ecosystem rather than competing with it. Japan’s approach is cooperative: secure domestic chip design capability while maintaining access to the global NVIDIA ecosystem.
UAE and Saudi Arabia — $100B+ in Committed Compute
The Gulf states have announced combined AI infrastructure investments exceeding $100 billion. The UAE’s G42 and Saudi Arabia’s HUMAIN are building hyperscale data centers with direct NVIDIA partnerships. Both countries are Tier 2 under U.S. export controls, requiring licenses for the highest-performance chips — a complication that has led both to negotiate government-to-government arrangements that include U.S. oversight provisions in exchange for access.
India — IndiaAI Mission
India’s IndiaAI Mission is targeting 10,000+ GPU deployments through public and private sector coordination. India is Tier 1 under U.S. export controls — unrestricted access — giving it a structural advantage over Tier 2 competitors in AI infrastructure buildout. NVIDIA has made India a priority market, with Jensen Huang committing to significant partnership announcements in 2026.
China — The Decoupled Stack
China is the most significant case. With H100, B200, and now H20 banned, Chinese AI development is proceeding on an alternative hardware stack centered on Huawei’s Ascend 910B and 910C series. Huawei claims Ascend 910C performance comparable to NVIDIA H100 for specific workloads. Independent benchmarks suggest a meaningful performance gap, but the Chinese government’s willingness to mandate domestic chip procurement for state-related AI projects means Huawei has a captive market at scale. The medium-term question is whether domestic demand is sufficient to fund the R&D needed to close the performance gap with NVIDIA Blackwell and Rubin generations.
What Enterprise Technology Leaders Need to Do
If your organization operates across multiple jurisdictions, the geopolitical layer requires the following operational adjustments:
- Export license review — Before ordering NVIDIA H100, B200, or equivalent chips for deployment in non-Tier-1 markets, involve legal counsel. The license application process for Tier 2 markets can take 60–120 days and may include usage restrictions and audit rights.
- Map your compute footprint to the AI Act — If you deploy AI to EU users, understand which models in your stack have been trained at scales that approach the 10^25 FLOP threshold. This affects your vendor due diligence requirements and your own liability exposure.
- Sovereign cloud options for regulated markets — AWS European Sovereign Cloud, Azure sovereign regions, and OVHcloud offer EU-resident compute. For organizations with EU data residency requirements, these platforms — powered by NVIDIA Blackwell — provide compliance-grade AI inference without building on-premises infrastructure.
- China strategy requires separate hardware planning — 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.
This is Part 8 of an advanced series on AI chip economics. Follow @VamsiTalksTech for updates.
Featured image designed by Freepik