Geopolitics

Key Takeaways

• AI infrastructure is viewed as a strategic national security asset comparable to oil refineries or uranium enrichment in previous eras
• US export controls restrict sales of advanced chips (H100, H200) to China and other countries
• China is developing domestic alternatives despite restrictions—Huawei Ascend chips and algorithmic efficiency breakthroughs like DeepSeek
• The Stargate announcement ($500B, 10 GW by 2029) reflects national security framing of AI infrastructure investment
• Allied coordination through Five Eyes and with Europe attempts to create technological blocs, but coherent policy remains elusive

AI as Strategic Resource

On January 21, 2025, President Donald Trump stood in the White House Roosevelt Room flanked by executives from OpenAI, Oracle, SoftBank, and NVIDIA. He announced "Stargate"—a joint venture committing $500 billion to build 10 gigawatts of AI data center capacity by 2029.

Presidential involvement in private infrastructure announcements is unusual. Presidents don't typically appear at groundbreakings for shopping malls or warehouse complexes. But this was different. "This is a resounding declaration of confidence in America's potential," Trump declared, framing the investment as essential to national security and economic leadership.

The event crystallized a shift in how governments view AI infrastructure: not as merely commercial technology, but as strategic national asset. Just as 20th century geopolitics revolved around control of oil refineries, uranium enrichment, and semiconductor fabrication, 21st century competition increasingly centers on AI computing capacity.

This reframing has profound implications. If AI infrastructure is strategic, then its location, ownership, and access become matters of statecraft. Export controls, industrial policy, and government investment follow. The trillion-dollar private sector buildout becomes entangled with national security imperatives.

The US-China Technology Competition

The strategic framing of AI infrastructure emerges primarily from US-China technological competition. Both countries view advanced AI as crucial to military capability, economic productivity, and global influence.

Why Infrastructure Capacity Matters

AI capabilities follow "scaling laws"—empirical relationships showing that model performance improves predictably with increased computing power, data, and parameters. Larger models trained on more data using more compute generally achieve better results.

This means that AI leadership isn't purely about algorithmic innovation. It requires massive infrastructure: the ability to train trillion-parameter models on thousands of GPUs for months, then serve billions of inference requests daily. Infrastructure capacity becomes a prerequisite for frontier capabilities.

The United States currently leads in this infrastructure race. American companies—NVIDIA, Microsoft, Google, Amazon, Meta—operate the world's largest AI data centers. China has substantial capabilities but faces constraints from semiconductor technology gaps and energy availability.

The Military Dimension

Advanced AI has clear military applications:

• Intelligence analysis: Processing satellite imagery, signals intelligence, and open-source data at scale
• Autonomous systems: Drones, vehicles, and weapons platforms with AI-enabled decision-making
• Cyber operations: Both offensive capabilities and defensive systems detecting attacks
• Logistics and planning: Optimizing supply chains, deployment, and resource allocation

The country that achieves decisive AI advantage in military applications could fundamentally alter strategic balance. This possibility drives both investment and restrictions on technology transfer.

The Economic Stakes

Beyond military applications, AI capabilities translate directly to economic productivity. AI systems are being deployed across:

• Drug discovery and biotechnology research
• Materials science and engineering design
• Financial modeling and risk assessment
• Manufacturing optimization and automation
• Energy system management and grid optimization

Countries leading in AI deployment gain productivity advantages that compound over time. The stakes aren't just about today's applications but decades of cumulative economic gains.

Export Controls

If AI infrastructure is strategic, then preventing adversaries from acquiring advanced technology becomes policy priority. The United States has pursued this through increasingly strict export controls.

October 2022: The Semiconductor Restrictions

In October 2022, the Biden administration announced sweeping restrictions on semiconductor exports to China. The rules prohibited:

• Advanced AI chips: NVIDIA's A100 and H100 GPUs, AMD's MI250, and similar products
• Semiconductor manufacturing equipment: Particularly extreme ultraviolet (EUV) lithography systems from ASML
• Technical support: American personnel supporting Chinese semiconductor facilities

The controls aimed to limit China's ability to train frontier AI models and to slow development of domestic semiconductor manufacturing that could eventually achieve independence from American technology.

The Performance Threshold

Export controls don't ban all chip sales—only chips exceeding certain performance thresholds. NVIDIA responded by creating "export-compliant" versions of its products: the A800 and H800, which have reduced chip-to-chip communication bandwidth to stay below control thresholds.

This created a cat-and-mouse dynamic: the US sets performance limits, manufacturers design products just below those limits, and controls are updated to close loopholes. In October 2023, expanded rules restricted the A800 and H800 specifically.

Allied Coordination

Export controls only work if allies cooperate. A Chinese company blocked from buying NVIDIA chips in the US could simply purchase from European or Japanese suppliers—unless those countries also restrict sales.

The US secured agreements from:

• Netherlands: ASML agreed not to sell EUV lithography equipment to China
• Japan: Restrictions on advanced semiconductor manufacturing equipment exports
• South Korea: Samsung and SK Hynix limited technology transfers to Chinese operations

This allied coordination attempts to create a "technological bloc"—countries aligned in restricting China's access to frontier technology. The approach mirrors Cold War export controls (COCOM) but applied to semiconductors and AI rather than nuclear or missile technology.

China's Response

Export controls create pressure for China to develop domestic alternatives. The effectiveness of restrictions depends on whether China can innovate around them or whether technological gaps prove insurmountable.

Huawei Ascend: The Domestic Alternative

Huawei, China's telecommunications and technology giant, developed the Ascend series of AI accelerator chips. The Ascend 910B, released in 2023, reportedly achieves performance roughly comparable to NVIDIA's A100 (though not the more advanced H100 or Blackwell generations).

Ascend chips use older semiconductor manufacturing nodes (7nm rather than cutting-edge 3-5nm) and face challenges with software ecosystems, developer tools, and energy efficiency. But they demonstrate that China can produce domestically-manufactured AI chips, even if they lag frontier performance by 1-2 generations.

The critical question is whether this gap narrows or widens over time. If Chinese semiconductor capabilities improve faster than export controls tighten, restrictions may prove ineffective. If the gap widens, China faces lasting technological disadvantage.

Algorithmic Efficiency: The DeepSeek Breakthrough

In December 2024, Chinese AI startup DeepSeek released a model claiming performance competitive with OpenAI's GPT-4 and Anthropic's Claude, but trained for only $5.5 million—a fraction of the $100+ million typical for frontier models.

If accurate, DeepSeek's approach represents a different response to export controls: instead of matching American hardware capacity, achieve comparable results through superior algorithms and training efficiency. The innovation demonstrates that compute restrictions don't guarantee AI leadership if algorithmic innovation can compensate.

DeepSeek's claims remain disputed, and independent verification is difficult given opacity around Chinese AI development. But even if exaggerated, the release signals China's strategy: innovate around hardware constraints through software, algorithms, and training methodologies.

The Innovation Incentive of Restriction

Export controls create powerful incentives for domestic innovation. When companies can't import technology, they must develop it themselves or go without. This pressure accelerates investment in alternatives.

History provides mixed evidence on whether such constraints succeed. Soviet-era export controls slowed but didn't prevent nuclear weapons development. Semiconductor restrictions on Japan in the 1980s contributed to rapid Japanese innovation that eventually challenged American dominance. The long-term effects of AI export controls remain uncertain.

Policy Reversals and Uncertainty

Export control policy has proven inconsistent, with reversals and contradictions undermining strategic coherence.

The AI Diffusion Rule: Proposed and Rescinded

In October 2024, the Biden administration proposed the "AI Diffusion Rule"—regulations requiring companies to obtain licenses before deploying advanced AI models internationally. The rule aimed to prevent American AI capabilities from benefiting adversaries.

Industry objected strenuously. How would licenses work for models accessible via cloud APIs? Would every customer require vetting? The compliance burden and competitive disadvantage of restrictions seemed enormous.

Within weeks, the administration withdrew the proposed rule, citing need for further study. The reversal illustrated the difficulty of regulating AI deployment in an interconnected global economy.

H200 Sales to China: Revenue Sharing Approved

In a seemingly contradictory move, the Commerce Department in late 2024 approved NVIDIA's request to sell H200 chips to China—technology theoretically restricted under export controls—if NVIDIA shared a portion of revenue with the US government.

The approval sparked immediate controversy. Congressional critics argued it undermined the entire export control regime. How could chips be simultaneously too dangerous to export freely but acceptable with revenue sharing? The decision suggested commercial considerations outweighing security rationale.

NVIDIA defended the arrangement, arguing that revenue sharing funded US AI research and that Chinese customers would simply buy alternatives (or smuggle chips) if sales were blocked entirely. Better to sell legally and capture revenue than lose markets to competitors or black markets.

Congressional Oversight Efforts

The House Select Committee on the Chinese Communist Party has attempted to impose stricter, more consistent controls through legislation. Proposed measures include:

• Tightening performance thresholds for chip exports
• Expanding controls to cover cloud computing services, not just hardware
• Restricting American investment in Chinese AI companies
• Requiring public disclosure of advanced AI development by Chinese entities

But Congressional action faces challenges. The technology evolves faster than legislation moves. Industry lobbying opposes measures that cost American companies revenue. And international coordination remains difficult when allies have different economic relationships with China.

The Challenge of Coherent Policy

Export control policy suffers from fundamental tensions:

• Security vs. Commerce: Restrictions protect advantage but cost industry revenue
• Unilateral vs. Multilateral: US-only controls are easily circumvented; allied coordination is slow
• Hardware vs. Software: Chips can be controlled; algorithms and techniques spread freely
• Present vs. Future: Today's restrictions may accelerate adversary innovation tomorrow

No policy can optimize across all dimensions simultaneously. The result is inconsistency, exceptions, and ongoing adjustments that create uncertainty for both industry and strategic planning.

The Stargate Announcement

The January 2025 Stargate announcement represented a different approach to AI geopolitics: massive domestic investment rather than restricting adversary access.

$500 Billion, 10 GW by 2029

The joint venture among OpenAI, Oracle, SoftBank, and MGX (the UAE sovereign wealth fund) committed to:

• $500 billion in capital investment over four years
• 10 gigawatts of data center capacity—roughly 10x current scale
• Facilities located exclusively in the United States
• Priority for American-manufactured equipment (particularly NVIDIA chips)

If executed, Stargate would represent the largest private infrastructure investment in American history, exceeding the Interstate Highway System in real dollars.

National Security Framing

The announcement emphasized national security benefits:

• Ensuring American AI leadership
• Creating jobs and industrial capacity
• Securing supply chains for critical technology
• Demonstrating allied investment confidence (UAE participation)

This framing positions AI infrastructure as comparable to defense production or space programs—ventures where national interest justifies government involvement beyond normal market regulation.

Bipartisan White House Support

Remarkably, both Trump administration officials and Biden-era appointees endorsed the Stargate framework. AI policy has achieved unusual bipartisan consensus: both parties view American AI leadership as crucial.

This consensus creates political space for policies that might otherwise face objection:

• Streamlined environmental review for data center projects
• Federal funding for grid infrastructure improvements
• Tax incentives beyond typical state offerings
• Priority interconnection queue status for AI facilities

Whether such policies materialize remains uncertain. But the bipartisan framing makes intervention more politically feasible than in typical commercial sectors.

Skepticism and Scrutiny

Not everyone accepts the national security framing at face value. Critics note:

• Private benefit, public cost: Companies profit while taxpayers subsidize infrastructure
• Overstated strategic importance: Is AI truly comparable to nuclear weapons or semiconductors?
• Environmental bypass: National security framing used to avoid environmental oversight
• Uncertain execution: Commitments don't guarantee construction; capital may not materialize

The magnitude of claimed investment ($500B) also invites skepticism. For comparison, total announced data center investment across all companies in 2024 was approximately $250 billion. Stargate alone claims to double that, creating questions about feasibility.

Allied Coordination

The United States can't pursue AI geopolitics unilaterally. Allied coordination shapes technology flows, investment patterns, and strategic positioning.

Five Eyes Alignment

The Five Eyes intelligence alliance (US, UK, Canada, Australia, New Zealand) has extended coordination to AI technology:

• Sharing threat intelligence on foreign AI development
• Coordinating export control policies
• Joint research initiatives on AI safety and security
• Aligned policies on Chinese technology in critical infrastructure

This coordination leverages existing intelligence relationships but faces challenges when commercial interests diverge. UK companies may prioritize different markets than American firms, creating pressure to deviate from aligned policies.

European "Strategic Autonomy"

The European Union pursues "strategic autonomy" in AI—developing domestic capabilities to reduce dependence on both American and Chinese technology. This includes:

• AI Act: Comprehensive regulation of AI development and deployment
• Data sovereignty: Requirements for data to be processed within Europe
• Industrial policy: Subsidies for European AI hardware and software companies
• Research funding: Billions allocated to AI research institutes

European strategic autonomy creates tensions with American policy. The EU's AI Act imposes regulatory requirements that American companies view as burdensome. European data localization creates market fragmentation. And European companies compete with American firms for global market share.

Yet Europe and the United States share fundamental interests: democratic governance, rule of law, and concern about authoritarian uses of AI. The challenge is balancing competition and coordination.

The Geography of Advanced AI

Advanced AI development currently concentrates in:

• United States: OpenAI, Anthropic, Google DeepMind, Meta, Microsoft (dominant position)
• China: Baidu, Alibaba, Tencent, DeepSeek (substantial but constrained)
• United Kingdom: DeepMind (pre-Google acquisition), university research (significant but smaller scale)
• Europe: Mistral AI, research institutes (emerging but capital-constrained)

This concentration means that AI geopolitics involves relatively few countries with frontier capabilities. Unlike manufacturing or agriculture, AI development doesn't distribute widely. The technology and capital requirements create natural concentration.

Whether this concentration persists or whether capabilities eventually diffuse remains an open question with profound geopolitical implications.