The White House released "Winning the Race: America's AI Action Plan" on July 23, 2025, setting out the Trump administration's strategy for keeping the United States ahead in artificial intelligence. Anthropic, which had sent recommendations for the plan in March, welcomed its focus on data center and energy permitting, federal AI adoption and security testing. It also said the plan fell short on transparency rules for frontier developers, and it raised concerns about the administration's recent reversal on exports of Nvidia's H20 chips to China. For developers, America's AI Action Plan signals faster federal adoption, more government testing of frontier models and a continuing fight over who writes the rules.

Key Facts#

  • Release: The White House published "Winning the Race: America's AI Action Plan" on July 23, 2025.
  • Origin: The plan followed a request for information from the White House Office of Science and Technology Policy (OSTP). Anthropic filed its recommendations on March 6, 2025.
  • Infrastructure: According to Anthropic, the plan commits to streamlining permitting for data centers and energy projects.
  • Federal adoption: Anthropic highlighted proposals that task the Office of Management and Budget (OMB) with removing obstacles to agency AI adoption, seek information on federal regulations that impede AI innovation, update procurement standards and expand AI use in defense and national security.
  • Security research: The plan prioritizes research into interpretability, AI control systems and adversarial robustness, and it affirms the frontier model evaluations run by NIST's Center for AI Standards and Innovation (CAISI).
  • Anthropic's objections: It wanted transparency requirements for frontier developers, opposed proposals to stop states from regulating AI if the federal government fails to act, and urged the administration to keep controls on H20 exports to China.

What Happened#

Anthropic published its response on the day of the release and said many of the plan's recommendations reflected its own OSTP submission. The company singled out the commitment to faster data center and energy permitting, warning that without enough domestic power, American AI developers could be pushed to move operations abroad, where sensitive technology might be exposed to foreign adversaries. It pointed to its "Build AI in America" report, released two days earlier, which estimated that the US AI sector needs at least 50 gigawatts of electric capacity by 2028; we covered that energy report separately.

On access and skills, Anthropic welcomed the plan's continuation of the National AI Research Resource (NAIRR) pilot, which it supports as a partner, and its emphasis on rapid retraining for displaced workers and pre-apprenticeship AI programs. On security, it called the plan's focus on defending against misuse of powerful models "appropriate and excellent" and encouraged continued investment in CAISI. It said advanced models are showing concerning gains in capabilities relevant to biological weapons and urged CAISI to concentrate on the most acute national security risks.

The sharpest comments concerned chips. Anthropic quoted the plan's statement that denying adversaries access to advanced AI compute is "a matter of both geostrategic competition and national security," then argued that allowing H20 sales undercut that goal. Its reasoning was technical: as reasoning models use more compute at inference time, memory bandwidth becomes a key constraint, and Anthropic said no Chinese-made chip matched the H20 on that measure. Exporting the H20, it argued, would help offset China's shortage of AI chips just as a new phase of competition begins.

Background#

The plan grew out of an OSTP request for information issued in early 2025. Anthropic's March submission covered six areas: national security testing of domestic and foreign models; stronger export controls, including on the H20, government-to-government agreements for countries hosting large chip deployments and lower license-free thresholds; stronger security at AI labs; 50 gigawatts of additional dedicated power by 2027; government-wide AI adoption; and better economic data. In that filing, Anthropic said it expected powerful AI systems to emerge in late 2026 or early 2027.

Transparency was the main gap in Anthropic's view. Earlier in July it had proposed a framework requiring the largest frontier developers to publish safety frameworks and system cards, and in its response it said it "would have liked to see the report do more on this topic." It also repeated its view that a 10-year moratorium on state AI laws is too blunt an instrument. In an October 2025 statement, CEO Dario Amodei noted that the Senate had voted 99 to 1 to strip such a moratorium from the One Big Beautiful Bill.

Why It Matters for Developers#

  • Selling to government gets easier, and more demanding. Procurement reform and OMB's mandate matter to anyone building AI-enabled software for federal agencies. Anthropic later partnered with the General Services Administration to offer Claude for Enterprise and Claude for Government for $1 across the federal government. Expect agency buyers to ask for evaluation results, data handling details and security documentation.
  • Government testing adds another input to model releases. With CAISI evaluating frontier models for national security risks, providers have one more reason to adjust model behavior between versions. Pin model versions and keep regression suites, following evaluating AI applications in .NET, so you notice when an upgrade changes your results.
  • The rules will still vary by state. The plan did not create a national transparency standard, so state legislatures continue to fill the gap. Build jurisdiction-aware configuration and feature flags now, as described in the Responsible AI and LLM security guide.
  • Export controls reach further than chips. Chip export decisions shape compute supply, and in June 2026 a US export control directive suspended access to Claude Fable 5 for a time. Keeping model calls behind IChatClient from Microsoft.Extensions.AI makes a forced provider switch a configuration change.

What's Next#

Much of the plan depends on agency follow-through: OMB's work on adoption barriers, the review of regulations seen as impeding AI, and new procurement standards. Anthropic said it would work with the administration on implementation while pressing for attention to catastrophic risks and for strong export controls. By October 2025, Amodei said Anthropic had publicly praised the plan, while repeating its call for a single federal transparency standard. Whether Congress creates that standard, or states keep writing their own rules, remains the main open question for AI builders.

Sources#