Who Controls AI in America

This file examines that stack as a power system rather than a branding story. The issue is who gets to define what AI is for, how fast it should be deployed, what risks matter, which firms become indispensable, and how much of the resulting power remains visible to the public.

The cleanest answer is not a single sovereign actor. It is a system of checkpoints. The White House controls direction. OMB controls implementation discipline inside the federal government. NIST controls a substantial part of the vocabulary. Congress controls spectacle and pressure. Defense institutions control strategic demand. Frontier firms control practical access to scale.

When the defense establishment signals that AI is central to decision advantage, cyber defense, intelligence analysis, logistics, or autonomous systems, it changes the incentive structure for the whole ecosystem. Companies stop pitching only productivity. They start pitching strategic indispensability.

A small number of companies sit near the center of the American AI stack because they occupy several positions at once: frontier models, cloud infrastructure, advanced chips, distribution platforms, and government-facing contracts. That does not amount to a hidden cabal. It does amount to a recognisable power map.

GAO’s recent reporting makes that visible. In its 2025 review of generative AI use and management at federal agencies, GAO found that reported AI use cases across selected agencies nearly doubled from 2023 to 2024, while generative AI use cases rose from 32 to 282. The same report noted that agencies were struggling to keep up with evolving federal policy and the pace of the technology itself.

OMB and the White House matter because the federal government is not just observing AI growth. It is trying to organise it. The 2025 OMB guidance on federal AI use requires agencies to maintain reporting and AI use-case inventories, apply minimum risk-management practices to high-impact AI, and protect privacy, civil rights, and civil liberties.

Any serious answer must include the companies closest to the technical frontier. Public debate often treats these firms as market actors waiting for government decisions. In practice, they help define the menu of decisions government thinks are available.

Capability concentration is the reason. Building and maintaining frontier AI systems requires extraordinary capital expenditure, specialised engineering teams, large-scale cloud infrastructure, and access to advanced chips. Once the frontier becomes expensive enough, it becomes easier for government to talk to a relatively small circle of firms and easier for those firms to present themselves as national assets.

Congressional hearings are one visible sign of that narrowing. In the Senate Commerce Committee’s 2025 hearing on strengthening U.S. capabilities in computing and innovation, the witness list included OpenAI, AMD, CoreWeave, and Microsoft: models, chips, cloud capacity, and enterprise distribution in miniature.

The defense and intelligence side matters because it changes the stakes, the language, and the speed. Once AI is cast as a national-security capability, arguments for caution are more likely to be filtered through competition logic rather than democratic oversight logic.

DARPA’s AI Next campaign announced a multi-year investment of more than $2 billion and framed AI as central to security-clearance vetting, system resilience, and next-generation applications. The Department of Defense’s 2023 Data, Analytics, and AI Adoption Strategy pushed the same direction in administrative terms: better decisions, faster, from the boardroom to the battlefield.

This does not mean the Pentagon secretly controls all American AI. It means defense demand, national-security urgency, and contractor relationships pull the ecosystem toward reliability, speed, integration, strategic advantage, and public-private coordination.

Control over AI does not only sit with those who train the largest models. It also sits with those who define acceptable process. Standards bodies, advisory committees, agency offices, and policy networks help manufacture legitimacy.

NIST is the clearest example. Its AI Risk Management Framework is not a binding code, but it has become part of the institutional toolkit that agencies and organisations use to show they are governing AI responsibly. Frameworks can function as sorting devices: they define maturity, diligence, compliance, and the harms that become legible.

The National Artificial Intelligence Advisory Committee and the wider AI.gov ecosystem consolidate expert input, federal initiatives, and strategic direction into something that looks coherent from Washington’s perspective. The National AI Research Resource debate is presented as democratising access, but it also reveals how much access to advanced AI capability depends on infrastructure concentration.

The hardest part is distinguishing legitimate governance from soft capture. Governments need frameworks for federal AI use. Congress should hold hearings. Standards bodies should exist. The question is whether those mechanisms are sufficiently independent from the firms and strategic interests they are supposed to oversee.

Some patterns are verified: the field is concentrated, federal use is growing, national-security framing is powerful, a small number of firms dominate frontier capacity and infrastructure, and advisory and standards structures help stabilise the language of governance. Other claims require more caution. The public evidence does not support a single coordinated command structure.