US Government Authority Over Commercial AI Infrastructure

Quick Assessment

Overview

With $700B+ in AI infrastructure spending in 2026 concentrated in 5-6 US companies,⁷ a critical but underexplored question is: what legal authority does the US government actually have to direct, repurpose, or access this commercial AI infrastructure?

The answer is: considerably more than most public discourse assumes. The US possesses multiple overlapping legal frameworks — the Defense Production Act, International Emergency Economic Powers Act, CLOUD Act, FISA Section 702, and broad executive order authority — that collectively give it extraordinary power over AI infrastructure owned by US companies. These authorities have been exercised in living memory, from COVID-era manufacturing directives to post-9/11 surveillance programs, and the institutional precedents for rapid mobilization are well-established.

This page examines these legal authorities, their historical precedents, current military-commercial AI integration, constraints on government action, and implications for both domestic companies and international stakeholders. The analysis intersects with compute concentrationRiskCompute ConcentrationAll six major AI infrastructure spenders (Amazon, Alphabet, Microsoft, Meta, Oracle, xAI) are US companies subject to CLOUD Act and FISA 702, giving the US government effective legal access to the ...Quality: 70/100, geopolitics, and export controlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100.

Legal Frameworks

Defense Production Act (DPA) — 50 U.S.C. sections 4501-4568

The Defense Production Act, originally passed in 1950 during the Korean War, gives the President broad authority to direct private industry for national security purposes:

• Title I (Priorities and Allocations): The President can require companies to accept and prioritize government contracts over all other orders. Applied to AI, this could mean directing MicrosoftOrganizationMicrosoft AIComprehensive reference page on Microsoft's AI strategy covering its \$80B+ infrastructure spend, restructured \$135B OpenAI stake (~27% ownership), Azure AI growth (39% YoY with 16pp AI contributi...Quality: 58/100, Amazon, or MetaOrganizationMeta AI (FAIR)Comprehensive organizational profile of Meta AI covering \$66-72B infrastructure investment (2025), LLaMA model family (1B+ downloads), and transition from FAIR research lab to product-focused GenA...Quality: 51/100 to dedicate specific GPU clusters to government AI workloads before serving commercial customers.
• Title III (Expansion of Productive Capacity): The President can provide loans, loan guarantees, and direct purchases to expand production capacity deemed essential for national defense.
• Title VII (General Provisions): Enables voluntary industry agreements and plans of action, potentially including coordinated AI safety measures across companies.

The DPA has been invoked with increasing frequency in recent decades:

No court has successfully blocked a DPA invocation during a declared emergency. The legal threshold is a Presidential finding that the directed action is necessary for national defense — a standard that courts have consistently deferred to the executive branch on.

International Emergency Economic Powers Act (IEEPA) — 50 U.S.C. sections 1701-1707

IEEPA gives the President authority to regulate commerce during a declared national emergency. This authority:

• Can block transactions, freeze assets, and restrict access to US technology
• Has been used to justify AI chip export controlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100 to China
• Could be used to restrict who can access US-based AI compute or mandate security requirements
• Was used by the Trump administration to impose tariffs in 2025, demonstrating broad judicial deference to executive interpretation⁸

Applied to AI infrastructure, IEEPA could theoretically be used to restrict foreign entities from accessing US cloud AI services, mandate security requirements for AI infrastructure operators, or even block specific uses of AI systems deemed threatening to national security.

CLOUD Act (2018) — Clarifying Lawful Overseas Use of Data Act

The CLOUD Act requires US companies to provide stored data to US law enforcement regardless of where that data is physically located.² This applies to all six major AI infrastructure companies:

• Model weights stored in European data centers remain subject to US legal process
• Training data held in any jurisdiction can be compelled
• AI inference logs and user data are accessible regardless of data residency commitments
• Microsoft France's Director of Public and Legal Affairs acknowledged under oath: "No, I cannot guarantee French data won't be seized by US authorities"²

FISA Section 702

The Foreign Intelligence Surveillance Act's Section 702 authorizes warrantless surveillance of non-US persons' communications and data held by US companies.³ Applied to AI infrastructure:

• Foreign governments and enterprises using US cloud AI services operate under effective US intelligence jurisdiction
• Inference queries, fine-tuning data, and model outputs are accessible
• Gag orders can prevent companies from disclosing the existence of surveillance orders
• The scale of AI infrastructure makes this surveillance authority more consequential than for traditional communications

Historical Precedents for Government Direction of Industry

World War II Industrial Mobilization

The most comprehensive precedent for government commandeering of private industry. The War Production Board directed virtually all US manufacturing toward military production from 1942-1945. Auto companies were ordered to produce tanks and aircraft instead of cars. US manufacturing output tripled during 1940-1944.⁹

The parallel to AI is instructive: in WWII, a small number of large industrial companies (GM, Ford, Boeing, Lockheed) were redirected from civilian to military production through executive authority. Today, a small number of large technology companies (Amazon, Alphabet, Microsoft, Meta, Oracle, xAI) control AI infrastructure that has both civilian and military applications. The legal precedents for directing such a transition are well-established.

Post-9/11 Telecommunications Surveillance (PRISM)

The NSA's PRISM program, revealed by Edward Snowden in 2013, required cooperation from major technology companies including Microsoft, Google, Apple, and Facebook in providing access to communications infrastructure.¹⁰ This precedent is directly relevant because:

• The same companies that participated in PRISM now control AI infrastructure
• The legal frameworks (FISA) used for PRISM apply equally to AI data
• The operational precedent for compelling tech company cooperation on national security exists
• The companies have institutional experience complying with classified government requests

COVID-19 DPA Invocations

The normalized use of DPA during COVID-19 lowered the political and institutional barriers to future invocations. The government directed 3M, GM, and other manufacturers to produce specific products, demonstrating that DPA mobilization is politically feasible and operationally practical in modern emergencies.¹

Current Military-Commercial AI Integration

Government authority over commercial AI infrastructure is not merely theoretical — the integration is already underway through commercial contracts:

The military AI market is growing at 19.5% CAGR, from $11.5B (2025) to an estimated $28.7B by 2030.¹¹ The autonomous weapons market is projected to grow from $15B to $27B over 2025-2029.¹² US DoD AI R&D spending reached $1.8B in FY2024.¹³

These pre-existing commercial relationships create institutional channels that would enable rapid mobilization if the government invoked emergency authorities. The companies already have security clearances, classified computing environments, and working relationships with defense and intelligence agencies.

Constraints on Government Authority

While US government authority is extensive, it is not unlimited:

• Fourth Amendment: Domestic surveillance generally requires warrants, though significant exceptions exist for national security and foreign intelligence
• Takings Clause (Fifth Amendment): Government must provide just compensation if it effectively takes private property — redirecting $700B in infrastructure would require compensation or a narrow national security justification
• First Amendment: AI-generated content may have speech protections, though this is largely untested in courts
• Political constraints: Technology companies collectively spend hundreds of millions on lobbying and have significant political influence
• Practical constraints: Government may lack the technical expertise to effectively direct AI infrastructure operations; the complexity of AI training means that government direction could reduce rather than enhance capability
• International backlash: Aggressive use of authority over US companies could push allies toward non-US alternatives, undermining the geopolitical advantage that US AI dominance provides

The Compensation Question

A key legal question is whether government direction of AI infrastructure constitutes a "taking" under the Fifth Amendment. Under DPA Title I, the government can require priority contracts but must pay fair market prices. Full commandeering of infrastructure would likely require emergency declaration plus compensation. The scale — potentially hundreds of billions of dollars — makes this a significant practical constraint even if the legal authority exists.

Implications for International Stakeholders

For Allied Nations

European and other allied governments that rely on US cloud AI services operate in a complex sovereignty position. The CLOUD Act means their data is ultimately subject to US jurisdiction regardless of where it is physically stored. The EU's attempts at "digital sovereignty" through regulations like GDPR are structurally undermined by dependence on US AI infrastructure.

This creates a strategic dilemma: allies need access to the most capable AI systems (overwhelmingly US-developed), but that access comes with implicit acceptance of US legal jurisdiction over their AI infrastructure. France's Mistral and other European AI efforts are orders of magnitude smaller than US companies.

For Adversarial Nations

China's response to US AI dominance has been to invest heavily in domestic alternatives. China's Big Fund III allocated $47.5B to semiconductor development,¹⁴ and Chinese companies like Huawei are building alternative AI chips. DeepSeek demonstrated that efficiency innovations can partially offset compute disadvantage, achieving GPT-4-level performance at reportedly 1/10 the compute cost.¹⁵

However, US export controlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100 on advanced chips have been only partially effective — an estimated 140,000 GPUs were smuggled to China in 2024.¹⁶

For Global AI Governance

US government authority over commercial AI infrastructure effectively gives the US veto power over global AI governance. International proposals that conflict with US national security interests can be blocked by directing US companies to non-compliance. This creates a structural tension between the aspiration for multilateral AI governance and the reality of unilateral US control over most frontier AI infrastructure. The geopolitics analysis scores international AI governance effectiveness at only 4.4 out of 10.¹⁷

The "Soft Nationalization" Thesis

A common observation is that "any property owned by US companies is sort of also owned by the US government." This overstates the legal reality but captures an important practical dynamic:

• The US government has more legal tools to direct US companies than any other government has over its own companies (due to DPA, IEEPA, CLOUD Act, FISA)
• In a genuine national security emergency, the precedent for comprehensive government direction of industry is firmly established (WWII, post-9/11, COVID)
• The concentration of AI in 5-6 US companies makes coordination far more feasible than if infrastructure were globally distributed — the government needs to work with only a handful of CEOs
• Pre-existing military contracts create institutional relationships and security clearances that would facilitate rapid mobilization

The question is not whether the US government could direct commercial AI infrastructure in an emergency, but at what threshold of AI capability or geopolitical crisis it would choose to do so — and whether the results would be beneficial or harmful.

Key Uncertainties

• Threshold for invocation: Under what circumstances would the US government actually use DPA or similar authorities for AI? A national security crisis, competitive pressure from China, an AI safety emergency, or a combination?
• Effectiveness of government direction: Would government-directed AI development be more or less capable and safe than commercially-directed development? Government historically excels at focused mobilization but struggles with innovation.
• Allied response: Would US allies accept overt government direction of AI infrastructure they depend on, or would it accelerate the development of non-US alternatives?
• Corporate resistance: Could companies legally or practically resist government direction? Historical precedent suggests limited resistance is possible (Apple vs. FBI in 2016) but difficult to sustain against determined executive action.
• Safety implications: Would government control of AI infrastructure improve safety (through mandated evaluations and deployment restrictions) or reduce it (through prioritizing capability over safety for national security)?
• Offshore incentives: Could government authority create incentives for companies to move AI infrastructure outside US jurisdiction? The scale and power requirements make this difficult but not impossible for inference workloads.

Sources