The Compute Geopolitics: Who Controls the Chips Controls AI

Compute as National Security

The geopolitics of AI compute have moved from a niche technology policy concern to a central pillar of national security strategy. The logic is straightforward: AI capability depends on compute, compute depends on advanced semiconductors, and advanced semiconductor manufacturing is concentrated in a handful of locations — primarily Taiwan (TSMC), South Korea (Samsung), and the United States (Intel, NVIDIA designs). Any disruption to this supply chain — whether from geopolitical conflict, trade restrictions, or natural disaster — would cripple the AI capabilities of affected nations and their enterprises.

As Ian Bremmer observed: "Compute is the new oil — and it's being weaponized." The parallel is apt. Just as oil shaped 20th-century geopolitics — with resource-rich nations wielding disproportionate influence and resource-dependent nations navigating supply risk — AI compute is shaping 21st-century power dynamics.

The US-China semiconductor competition is the defining axis of this dynamic. Since 2022, the US has implemented progressively tighter export controls on advanced AI chips to China, restricting NVIDIA H100, A100, and their successors. China's response has been aggressive domestic investment in alternative chip architectures (Huawei Ascend, Cambricon) and workarounds (cloud access through third countries). The result: China's share of global AI compute capacity has declined from approximately 25% in 2023 to 15% in 2026 — a strategic shift with implications for every enterprise operating across these markets.

For enterprise AI leaders, the geopolitical dimension of compute is no longer something to outsource to the government affairs team. It is a direct input to infrastructure strategy, vendor selection, and deployment architecture.

Global Compute Capacity by Region

The global distribution of AI compute capacity has shifted dramatically under the pressure of policy interventions:

United States (42% of global AI compute): The dominant position, driven by hyperscaler data center investments (AWS, Azure, Google Cloud) and NVIDIA's market position. The CHIPS Act has catalyzed over $200B in announced semiconductor investments, including new TSMC and Samsung fabs in Arizona and Texas, and expanded Intel facilities in Ohio and New Mexico. However, these fabs will not reach production scale until 2027-2028, leaving a near-term dependency on overseas manufacturing. The US advantage is in AI chip design (NVIDIA, AMD, Google TPU) and cloud infrastructure, not in fabrication.

Europe (18% of global AI compute): The EU's semiconductor ambitions (European Chips Act, EUR 43B investment target) are progressing but lagging. Europe has limited AI chip design capability and minimal advanced fabrication. Its compute strategy relies heavily on US hyperscalers operating European data centers. The Gaia-X sovereign cloud initiative aims to reduce this dependency, but enterprise adoption has been slow. Key concern: European enterprises face the tension between US cloud dependence and EU data sovereignty requirements.

China (15% of global AI compute): Down from 25% in 2023 due to export controls restricting access to advanced NVIDIA chips. China has responded with domestic alternatives — Huawei's Ascend 910B chip is the leading domestic option, achieving approximately 70% of H100 performance on key benchmarks. Baidu, Alibaba, and Tencent are investing heavily in domestic compute infrastructure. Key concern: Chinese AI compute is increasingly isolated from the global ecosystem, creating a bifurcated market that enterprises must navigate.

Asia-Pacific ex-China (15% of global AI compute): Japan, South Korea, Singapore, India, and Australia are all investing in AI compute capacity. Japan's $13B semiconductor subsidy program is the most aggressive. Singapore has emerged as a neutral hub for AI compute in Southeast Asia. India is building domestic capacity but remains early-stage.

Middle East (5% of global AI compute): Saudi Arabia and the UAE have made outsized AI compute investments relative to their economies. The UAE's G42 (backed by Mubadala) and Saudi Arabia's NEOM are building hyperscale AI data centers. Key concern: US export controls apply to some Middle Eastern entities, creating uncertainty for enterprises planning to use Middle Eastern compute.

Rest of World (5%): Distributed across Brazil, Canada, Africa, and smaller markets. Compute capacity is growing but remains a small fraction of global total.

Impact on Enterprise AI Strategy

The geopolitics of compute affect enterprise AI strategy in four concrete ways:

1. Hardware procurement risk: Enterprises that depend on NVIDIA GPUs for AI training and inference face procurement lead times of 6-12 months for H100 and successor chips. The supply constraint is not a temporary shortage — it is a structural feature of a market where demand growth outpaces fabrication capacity expansion. Enterprises must plan hardware procurement 12-18 months ahead and consider alternative chip architectures (AMD MI300X, Google TPUs, Intel Gaudi) as hedges against NVIDIA concentration risk.

2. Cloud region availability: Hyperscalers are expanding AI compute capacity to new regions, but not all regions have equal availability. Enterprise workloads that require specific geographic deployment (for data residency, latency, or regulatory reasons) may face capacity constraints in less-developed regions. The practical impact: an enterprise requiring AI inference in Brazil, India, or Southeast Asia may have fewer GPU-equipped regions to choose from and higher costs per inference.

3. Sovereign cloud requirements: 42 countries now have some form of data localization law requiring that citizen data be processed within national borders. For AI workloads, this means that inference (and in some cases, fine-tuning) must occur on infrastructure physically located within the country. This creates demand for in-country compute that may not exist at the required scale, forcing enterprises to choose between compliance and capability.

4. Supply chain concentration risk: TSMC fabricates approximately 90% of the world's most advanced semiconductors at facilities in Taiwan. Any disruption to TSMC — whether from conflict, natural disaster, or trade restriction — would impact AI chip supply globally. Enterprises with mission-critical AI workloads should evaluate their supply chain exposure to TSMC through their chip and cloud providers.

The enterprises that treat compute as a strategic input — not just a procurement line item — will navigate these dynamics more effectively than those that treat it as a commodity.

Compute Risk Assessment Framework

Enterprise AI leaders should conduct a Compute Risk Assessment covering five dimensions:

Dimension 1 — Hardware Dependency: What chips power your AI workloads? How concentrated is your dependency on a single vendor (typically NVIDIA)? What is your procurement lead time? Do you have alternative chip paths tested and validated? Mitigation: qualify at least two chip architectures for your primary AI workloads. Budget for 12-month forward procurement.

Dimension 2 — Cloud Provider Concentration: What percentage of your AI compute runs on a single cloud provider? In how many regions? What happens if your primary cloud provider restricts capacity in your region? Mitigation: maintain AI workload portability across at least two cloud providers. Use containerized inference (e.g., NVIDIA Triton, vLLM) that abstracts the underlying hardware.

Dimension 3 — Geographic Exposure: In which countries do you need to run AI workloads? Do those countries have adequate compute capacity? Are there data residency requirements that constrain your deployment options? Mitigation: map your AI workload requirements against available compute in each required geography. Identify gaps early and engage with cloud providers on capacity roadmaps.

Dimension 4 — Export Control Compliance: Do any of your AI activities involve transferring models, data, or compute access across jurisdictions subject to export controls? Are you confident that your cloud provider's multi-region architecture complies with applicable export regulations? Mitigation: engage legal counsel familiar with EAR/ITAR and semiconductor export controls. Audit your AI supply chain for compliance risks.

Dimension 5 — Cost Trajectory: AI compute costs are not declining as fast as historical compute costs. GPU pricing is influenced by supply constraints, geopolitical premiums, and growing demand. What is your 3-year compute cost projection? How sensitive is your AI program's ROI to compute cost increases of 20-50%? Mitigation: invest in inference optimization (model distillation, quantization, speculative decoding) to reduce compute requirements per query. Evaluate open-weight models that can run on less expensive hardware.

The output of this assessment should be a Compute Risk Scorecard that quantifies your exposure across all five dimensions, prioritizes mitigation actions, and establishes monitoring metrics. Review quarterly — the geopolitical landscape of compute shifts faster than most enterprise planning cycles.

Strategic Recommendations for Enterprise Leaders

The geopolitics of AI compute require enterprise leaders to think about infrastructure with the same strategic rigor they apply to market strategy. Five recommendations:

1. Diversify your compute supply chain. Do not depend on a single chip vendor, a single cloud provider, or a single geography. Qualify multiple hardware platforms, maintain cloud portability, and plan for capacity in every region where you need to deploy. This adds operational complexity but reduces catastrophic risk.

2. Invest in compute efficiency. The cheapest and most geopolitically resilient compute is the compute you do not need. Model distillation (training smaller models that approximate larger ones), quantization (reducing model precision to reduce compute requirements), speculative decoding (using small models to draft and large models to verify), and inference caching (avoiding redundant computation) all reduce your compute dependency.

3. Build relationships with your compute providers. In a constrained market, relationships matter. Enterprises that are strategic customers of cloud providers and chip vendors receive better allocation, earlier access to new capabilities, and more responsive support. Treat your cloud provider as a strategic partner, not a commodity vendor.

4. Monitor the regulatory landscape. Export controls, sovereign cloud requirements, and data residency laws are evolving rapidly. Assign a team (combining legal, compliance, and infrastructure expertise) to monitor changes and assess their impact on your AI deployment architecture. Proactive compliance is orders of magnitude cheaper than reactive remediation.

5. Plan for the 2027-2028 capacity inflection. New semiconductor fabs (TSMC Arizona, Samsung Taylor, Intel Ohio) will begin production in 2027-2028, significantly expanding global AI compute capacity. Enterprises that are well-positioned — with validated workloads, multi-vendor architectures, and clear deployment requirements — will be able to scale aggressively when capacity becomes more available. Those that are not prepared will lose the window to competitors who are.

Your AI supply chain is now a geopolitical risk. Manage it accordingly.