Pentagon Signs Seven Major AI Firms for Classified Military Networks

The U.S. Department of Defense has closed a sweeping round of agreements that put advanced AI from seven marquee firms onto classified military networks, according to statements from the Pentagon and multiple officials briefed on the deals. Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Elon Musk’s xAI, and the startup Reflection AI are now cleared to provide models, compute, and cloud services inside the government’s most sensitive enclaves. Anthropic remains conspicuously absent; defense officials told The Verge the company is still classified as a “supply-chain risk,” a direct fallout from its earlier refusal to relax safety filters for classified workloads.

The Pentagon’s blacklist of Anthropic traces back to February, when the startup declined to modify its usage policies for classified environments, sources familiar with the negotiations told Reuters. Hours after that impasse became public, OpenAI announced its own classified deal, a sequence Scrippsnews described as “opportunistic and sloppy.” Anthropic’s insistence on keeping its constitutional-AI guardrails intact, while popular in Silicon Valley, collided with DoD demands for unrestricted deployment. The episode has turned into a case study on how corporate AI-safety stances can become competitive liabilities in the government market.

Nvidia’s role centers on GPU clusters and inference software hardened for air-gapped networks, according to Bloomberg. Microsoft and AWS will each stand up sovereign cloud regions that meet Impact Level 6 security controls, letting agencies run frontier models without touching the public internet. Google’s amendment gives the Pentagon unfettered access to Gemini for “any lawful government purpose,” but the company simultaneously walked away from a $100 million drone-swarm program after staff protests, Tom’s Hardware reports. OpenAI’s agreement, published on its own blog, explicitly bars using its models to spy on Americans or conduct “mass surveillance,” a concession added after employee and civil-liberties backlash. Reflection AI and xAI’s exact deliverables remain classified, though industry observers expect them to supply smaller, fine-tuned models for edge deployments.

Commercial customers will feel the ripple effects almost immediately. Sovereign-cloud capacity earmarked for DoD workloads is capacity that can’t be sold to Fortune 500 firms, which means tighter GPU supply and longer lead times for everyone else. Nvidia’s hardened drivers, originally built for classified use, will likely trickle into the company’s enterprise AI Enterprise suite within 12-18 months, giving regulated industries new compliance shortcuts. More subtly, the deals cement Microsoft and Amazon as the default secure-cloud gatekeepers for both the Pentagon and Wall Street, making it harder for Google Cloud to pitch itself as the privacy-first alternative.

Employee revolts are already reshaping scope. Google staff delivered a petition to CEO Sundar Pichai on April 29 demanding the company exit lethal-autonomy projects, prompting the firm to abandon the drone-swarm contract while keeping the Gemini access deal intact, according to Tom’s Hardware. At OpenAI, policy head Chris Lehane spent weeks negotiating language that forbids “targeted surveillance of U.S. persons,” a clause that survived the final contract after an internal firestorm. These compromises show that even trillion-dollar defense budgets still bow to Silicon Valley labor markets when key engineers threaten to quit.

Contracts award letters have been issued, but the real work starts now: accreditation teams must port models onto classified enclaves, strip telemetry, and retrain on sensitive corpora without ever touching the open internet. That process typically takes six to nine months, putting live deployment around early 2027. In parallel, the DoD is drafting a second tranche of agreements aimed at AI agents capable of coordinating logistics and battlefield sensor fusion. Expect the same cast of companies, plus whoever replaces Anthropic in the next competitive round, to bid again by year-end.