The White House wants a closer look at the most powerful AI systems before they reach the public. A new executive order on advanced AI innovation and security directs federal agencies to establish a voluntary process allowing the government to evaluate certain frontier AI models before public release, when those systems could pose risks to national security or critical infrastructure. This initiative signals a major shift in Washington, expanding federal scrutiny beyond physical AI campuses and rising power demand to the frontier models running inside them.

In practice, frontier models refers to the most advanced AI systems available at a given time, specifically large foundation and reasoning models capable of sophisticated coding, scientific research, cyber operations, and autonomous task execution. Companies developing such systems include OpenAI, Anthropic, Google DeepMind, xAI, and other leading AI labs. As agencies implement the order, federal officials have focused on whether emergent capabilities could be misused against critical infrastructure, government systems, or other sensitive targets if released without adequate testing and safeguards.

The order frames advanced AI capabilities as introducing new national security considerations and calls for coordinated action across federal agencies to deploy the best and most secure technology at speed. However, it does not designate AI as a critical infrastructure sector, and it stops short of creating a licensing or permitting regime for AI development. Instead, it places frontier models within a federal review framework focused on systems whose failures, compromise, or misuse could affect critical infrastructure, financial networks, government operations, healthcare organizations, emergency services, and national security.

This new process builds on testing efforts already underway. In May, the National Institute of Standards and Technology’s Center for AI Standards and Innovation announced agreements with Google DeepMind, Microsoft, and xAI to conduct pre-deployment evaluations of frontier AI systems. The center previously established similar arrangements with Anthropic and OpenAI. According to the institute, those evaluations examine capabilities that could pose national security risks, including cybersecurity, biosecurity, and chemical weapons applications.

The White House has emphasized that participation remains voluntary, and the administration stated that the order does not authorize mandatory licensing, pre-clearance, or permitting requirements for AI model development or release. The voluntary framework leaves open questions about implementation. Kevin Frazier, an adjunct research fellow in technology and AI at the Cato Institute, said the order’s effectiveness will depend on how agencies determine which models qualify for review and how transparently those decisions are made. He noted that the order reflects that AI governance is a question that includes model development and deployment as well as the entire AI tech stack.

For data center operators and cloud providers, the implications extend beyond policy. If frontier model evaluations become routine in development, providers may see heightened demand for secure, pre-release test environments, strict access controls, telemetry systems, and audit capabilities. Questions also remain about how model access will be managed, how intellectual property will be protected during evaluations, and whether specialized secure computing environments will emerge to support government testing.

For much of the AI boom, Washington’s attention often focused on the physical demands of artificial intelligence, namely power, water, land, and interconnects. Utilities have rewritten forecasts as hyperscalers and AI developers pursue multi-gigawatt campuses. Grid operators are studying the implications of large-load growth, and developers are seeking land, power, and transmission access for increasingly capable AI systems.

Security experts argue that protecting advanced AI systems requires looking beyond the models themselves. Katharina Sommer, group head of government affairs at cybersecurity consultancy NCC Group, explained that it’s hard to have resilient and secure AI deployment without addressing model security and wider system security simultaneously. She noted that governments should build on existing cybersecurity frameworks rather than creating entirely new security regimes for each emerging technology. Approaches developed for critical infrastructure and other high-consequence systems can serve as a foundation for oversight mechanisms for frontier AI systems.

While public debate often focuses on model capabilities, many operational risks stem from the environments where those models run, including permissions, data access controls, monitoring systems, and connections to broader IT infrastructure. Infrastructure sovereignty and resilience are also receiving increased scrutiny as governments assess the systems supporting AI deployment. This dynamic is playing out globally, as the European Commission recently warned that insufficient data center capacity could limit the region’s ability to capitalize on AI development, while simultaneously emphasizing the importance of sovereign compute capacity.