White House Bypasses Pentagon to Secure Anthropic AI for Spy Agencies Amid Chip Crunch

The White House has moved decisively to lock in access to advanced artificial intelligence for the nation’s intelligence agencies. A classified contract with Anthropic now takes shape. It lets the National Security Agency and its peers tap the startup’s latest models despite an earlier clash with the Pentagon.

This shift marks a pragmatic turn. Months of tension between Anthropic and the Defense Department had produced a supply-chain risk designation, contract cancellations, and a government-wide pause on the company’s technology. Yet spy agencies see the tools as too valuable to sideline. And the administration agrees.

At the center sits a $9 billion secret funding request. The White House approved it to buy cutting-edge chips and build the infrastructure America’s spies need. Those chips power the newest AI systems. Without them agencies risk falling behind adversaries who pour resources into similar capabilities. The gap is real. The timeline is tight.

New models consume far more compute than experts predicted even recently. Intelligence officials worry the shortage leaves them unable to test or deploy frontier systems on classified networks. Susie Wiles, White House chief of staff, stepped in. She authorized the NSA to keep using an advanced Anthropic model while the new contract finalizes, according to officials cited by The New York Times.

The model in question is Mythos. Released in April, it runs more efficiently on the latest hardware but works on older chips too. That flexibility matters when data centers with proper cooling and power take time to construct. Intelligence agencies rely heavily on Amazon Web Services classified clouds. Upgrades there move slowly under strict security rules.

But this isn’t simply about hardware. The contract includes deliberate limits. It bars use of the AI on Americans’ data. The language avoids the Pentagon’s demand for “any lawful use.” Earlier this year that insistence triggered a standoff. Anthropic resisted, citing concerns over mass surveillance and autonomous weapons. The Pentagon labeled the company a supply chain risk. Federal agencies began phasing out its tools.

Anthropic pushed back in court. It secured a preliminary injunction in one case. Meanwhile its CEO Dario Amodei met with senior administration figures including Wiles. Those sessions were described as productive. Now the White House appears ready to chart a different path for intelligence work than for military operations.

Officials tell The Information the emerging agreement would permit classified use with restrictions. Anthropic’s models carry a reputation for stronger safety features than some rivals from OpenAI or Google. Intelligence leaders reportedly prefer them for sensitive analysis. The deal could open the door for a multibillion-dollar broader government pact.

Recent reporting adds weight. GV Wire confirmed the $9 billion figure and Wiles’ direct role in keeping NSA access alive. It noted the contract explicitly drops the unrestricted-use clause that defined the Pentagon fight. Larger investments will likely follow. Experts say current sums only scratch the surface of what classified AI demands.

AI already shapes daily intelligence tasks. It combs intercepts, spots overlooked signals, and aids cybersecurity. Mythos gained early restricted sharing because of its skill at discovering and exploiting software flaws. Only a handful of U.S. and U.K. agencies and banks saw it at first. Its full release remains limited.

The administration’s move reflects hard realities. Spy agencies cannot gather intelligence domestically. Their rules on Americans overseas are strict. That makes them more comfortable with the safeguards Anthropic demands. Pentagon leaders wanted maximum flexibility for warfighting. The White House sees the intelligence community’s needs as distinct.

So the Anthropic contract is designed to become a template. Other AI providers could follow with similar terms. Intelligence officials hope it smooths a parallel arrangement for OpenAI. That company already holds Pentagon deals but lacks a direct NSA path yet. A successful Anthropic model might ease concerns across the board.

Chip scarcity complicates everything. Nvidia’s Grace Blackwell superchips sit at the heart of the $9 billion plan. Congress must still approve the full amount. The administration has reprogrammed $800 million to accelerate some capacity. Even with money in hand, physical build-out of secure data centers takes months or years. Liquid cooling, massive power draws, and isolation from unclassified networks add layers of delay.

Former NSA chief data scientist Vinh Nguyen captured the urgency. “Our intelligence community needs the frontier — the best A.I. chips, models, systems, talent — on a timeline that matches the threat,” he told The New York Times. The statement lands with force. Threats do not wait for infrastructure.

President Trump has signaled interest in AI leadership. He abruptly canceled an executive order signing last week, saying he disliked parts of it. That order aimed to create a process for sharing new models with government before public release, acknowledging cybersecurity risks. The cancellation leaves questions about the administration’s exact policy direction. Yet action on chips and the Anthropic deal shows concrete steps continue.

Industry demand for top chips runs red hot. Anthropic itself has throttled access to its Claude chatbot during peaks, much like a utility managing blackouts. Global competition with China, which cannot receive the latest Nvidia exports, adds strategic pressure. U.S. agencies must outpace both commercial growth and foreign programs.

The dispute’s backstory runs deep. Last summer the Pentagon awarded up to $200 million each to Anthropic, OpenAI, Google, and xAI. Those deals sought to speed AI adoption for national security. When negotiations over usage terms broke down early this year, the administration moved aggressively. Trump directed agencies to stop using Anthropic products. The supply-chain risk label followed. OpenAI quickly filled some gaps.

Anthropic maintained it had narrow exceptions only. No mass domestic surveillance. No fully autonomous lethal weapons without oversight. The company argued those positions aligned with its principles and early contracts. A lawsuit and public statements from CEO Amodei and co-founder Jack Clark kept the pressure on. Clark told Reuters in April the firm continued talks on Mythos with the administration, stressing shared national security goals.

Now the pendulum swings back. The NSA contract carves out space for continued collaboration. It sidesteps the Pentagon’s harder line. And it pairs with serious money for hardware. The combination suggests the White House wants American frontier AI inside sensitive government systems. On its own terms where possible.

Delays remain a risk. Classified networks cannot simply borrow commercial capacity without rigorous vetting. Even approved models require integration time. Training and inference at scale eat power at levels that strain even upgraded facilities. Future funding requests will almost certainly grow.

Still, the trajectory is clear. Intelligence agencies gain access to Mythos and potentially more. Anthropic regains a foothold in the most sensitive corridors of government. The Pentagon’s earlier ban loses some sting for non-military users. And the $9 billion investment signals that compute for AI now ranks alongside traditional spy hardware in priority.

Officials declined further comment citing classification. Representatives from Anthropic, OpenAI, AWS, and Nvidia offered no public statements on the developments. The silence fits the domain. Yet the pattern of authorizations, funding, and meetings tells its own story.

AI has moved from experimental aid to core analytical engine inside the intelligence community. The White House bet is that the best models, paired with sufficient chips and careful rules, deliver advantage without unacceptable risk. Whether that balance holds as capabilities accelerate will define the next chapter. For now the deal moves forward. The chips are ordered. The models are authorized. The work continues.