Anthropic NSA deal nears clearance as AI chip push grows

White House officials are moving toward giving U.S. spy agencies access to Anthropic’s models, tying a classified National Security Agency arrangement to a broader $9 billion chip push for intelligence work, The New York Times reported. The shift makes Anthropic’s Washington effort a procurement issue as much as a policy fight, linking model access to the computing capacity agencies say they need.

The stakes run beyond one vendor. Putting frontier AI inside the intelligence budget would signal that the administration is ready to treat a small set of model developers as strategic suppliers. For rivals and investors, the question is now which companies get cleared, what review they face and how much compute Washington is willing to ring-fence.

The latest report extends Axios’s April scoop that the NSA was already using Anthropic’s Mythos model despite resistance elsewhere in the Pentagon. Axios said Anthropic had limited Mythos Preview to about 40 organisations, showing how tightly the company controlled access before a formal intelligence contract came into view. An unusual access arrangement now looks closer to a classified deal with institutional backing.

Over the past six weeks, Washington’s position has shifted. Reuters reported in April that Anthropic chief executive Dario Amodei met White House officials as concern over Mythos spread through Washington. Reuters reported again this month that administration officials briefed AI companies on a model-review process. Taken together, the reports suggest the Anthropic-NSA arrangement is being folded into a wider federal process for deciding which frontier systems can be used in sensitive settings, not treated as a one-off.

The hardware piece may matter as much as the software. The Times said the White House approved the $9 billion request so spy agencies can buy advanced AI chips, placing the Anthropic arrangement inside a broader effort to keep intelligence agencies from trailing commercial labs on compute. If the spending goes ahead, it would shape both access to Mythos and how much scarce processing capacity Washington reserves for classified work.

Competition is also building. Semafor reported that OpenAI launched a cybersecurity model meant to rival Mythos, and Axios later reported that access to that system was being widened for cyber defenders. If government buyers start pairing model approvals with dedicated chip budgets, companies that can show both technical performance and a governance structure officials can defend inside classified procurement may have an advantage. Anthropic’s run-rate revenue topped $30 billion in early April, according to the Congressional Research Service, and a durable intelligence customer would still carry strategic weight even if the near-term revenue effect stayed modest.

The policy tension

A Congressional Research Service brief captures the contradiction around the deal. The report said Anthropic has argued that it “do[es] not believe that today’s frontier AI models are reliable enough to be used in fully autonomous weapons” and that “mass domestic surveillance of Americans constitutes a violation of fundamental rights,” even as federal agencies deepen their interest in the company’s tools. The same CRS note said the Defense Department had awarded Anthropic, Google, OpenAI and xAI contracts worth up to $200 million each, showing how quickly the federal market has widened beyond a single intelligence customer.

White House backing also matters because the company is still facing institutional resistance. CRS cited an earlier Pentagon position from Defense Secretary Pete Hegseth:

“Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.”

— Pete Hegseth, cited in the Congressional Research Service brief

If the administration lets intelligence agencies proceed while expanding chip purchases, the decision has moved up a level. The issue is no longer only Pentagon resistance to one model, but a broader national-security judgment about compute, access and competition. The next test is operational: chips, secure infrastructure, model reviews and contract terms that decide whether frontier AI moves from limited testing into routine classified use.