The Pentagon Just Signed Eight AI Companies for Its Classified Networks. Anthropic Wasn't Invited.

The press release went out from war.gov on Friday morning. It carried the official letterhead of the U.S. Department of War, the rebranded successor to the Department of Defense, and ran four short paragraphs. The headline read "Classified Networks AI Agreements." The text named eight companies: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle.

It did not name Anthropic.

That single absence is the story. Two months ago, Claude was the only large language model approved on Pentagon classified networks at scale, with forward-deployed Anthropic engineers embedded across multiple military programs. On Friday, every other major American AI company was approved for the same networks, and the company that built the original capability was excluded by name.

The deals authorize the eight named companies to deploy their AI hardware and models on the Pentagon's Impact Level 6 (IL6) and Impact Level 7 (IL7) environments, the two highest tiers of the Defense Department's cloud security architecture. IL6 covers classified data up to the Secret level. IL7 covers compartmented intelligence and the most sensitive operational systems. Together they are where the U.S. military runs its actual war-fighting decision support, not just back-office paperwork.

The Eight Companies and What They Got

The Pentagon's announcement framed Friday's news as the culmination of a multi-month diversification push. The earlier deals with Google, SpaceX, and OpenAI had already been signed and reported separately in March and April. The new contracts revealed Friday added NVIDIA, Microsoft, Amazon Web Services, Reflection AI, and (announced separately on X later in the day) Oracle.

In a press statement, the War Department said: "These agreements accelerate the transformation toward establishing the United States military as an AI-first fighting force and will strengthen our warfighters' ability to maintain decision superiority across all domains of warfare."

The Pentagon's existing generative AI platform, GenAI.mil, has now been used by 1.3 million Department personnel across "tens of millions of prompts" in five months, per the official release. The classified-network deals extend that capability into IL6 and IL7, where the user base is smaller but the operational stakes are higher.

The most striking inclusion on the list is Reflection AI. The startup, founded in March 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, has not yet publicly released the open-weight frontier model at the center of its pitch. Its code research agent Asimov is still on a waitlist. As of early March, the Financial Times reported Reflection was raising at $20 billion despite shipping no public model. The Pentagon contract is the first time a U.S. agency has bought into the company at the operational tier.

How Anthropic Lost the Pentagon

The Anthropic exclusion is the structural climax of a fight that began in mid-February.

Anthropic had two firm conditions on its Defense Department use: no autonomous weapons targeting decisions and no domestic mass surveillance. The Pentagon, under Secretary Pete Hegseth and Pentagon CTO Emil Michael, demanded unrestricted access for "all lawful purposes." The talks broke down on February 27 at a 5:01 PM deadline. Within hours President Trump posted on Truth Social directing every federal agency to immediately cease all use of Anthropic's technology. Hegseth designated the company a supply chain risk, a label normally reserved for foreign adversaries like Huawei. No American company had ever received it.

Anthropic sued in two federal courts on March 9, alleging First Amendment retaliation and Administrative Procedure Act violations. On March 26, U.S. District Judge Rita Lin granted a preliminary injunction. Lin's order found the supply chain risk designation was likely "pretextual" and that the government's "real motive was unlawful retaliation" for Anthropic's public stance on safety guardrails. Pentagon CTO Emil Michael responded the same day on X that Lin's order contained "dozens of factual errors" and that the Section 4713 designation was still "in full force and effect."

On April 8, the D.C. Circuit Court of Appeals denied Anthropic's request to pause the supply chain risk designation while the appeal proceeds, with oral argument set for May 19. The court wrote that "the equitable balance here cuts in favor of the government" because the harm to Anthropic was "relatively contained" while the government interest was "judicial management of how, and through whom, the Department of War secures vital AI technology during an active military conflict." Friday's announcement is the procurement consequence of that ruling.

The Companies Anthropic Used to Compete With Are Now Inside the Tent

Two months ago, the companies that beat Anthropic to the Pentagon's first contract spillover were OpenAI and xAI. Sam Altman had to walk back his initial Pentagon comments after 1.5 million users left ChatGPT in protest. By late April, the parade had widened. Google signed a Gemini-on-classified-networks deal over the public objection of 600 of its own employees on April 28. Friday's release wrapped the rest of the field.

What it confirms is the structure the Pentagon has been building since Anthropic walked out: a redundant, multi-vendor architecture deliberately designed to remove any single AI lab's ability to set ethical terms on military use. Pentagon CTO Emil Michael had told Bloomberg in March, "I need redundancy." He now has it eight times over.

The official statement on Friday repeated the language: "The Department will continue to build an architecture that prevents AI vendor lock and ensures long-term flexibility for the Joint Force."

For Anthropic, the architecture is the punishment. The injunction prevents the Pentagon from formally banning Anthropic. It does not require the Pentagon to buy anything from Anthropic. Friday's announcement is the policy expression of that gap.

What This Means for Practitioners

The eight-company classified deal lands at a strange moment for AI engineering. Anthropic, the company excluded from the Pentagon's procurement list, just disclosed a $30 billion run-rate revenue in early April.

Investors are now reviewing offers that would push the company past 900 billion dollars in valuation. The pricing implies Anthropic is the most valuable private AI lab on Earth, even as the U.S. government is treating its products as procurement-prohibited.

For ML engineers working in defense, intelligence, and federal contracting, the immediate practical change is the model lineup approved for IL6/IL7. The list of approved models is now long enough that defense contractors can reasonably standardize on multiple of them. For practitioners not in defense, the more interesting development is the inclusion of Reflection AI alongside the established frontier labs. The startup's open-weight strategy, modeled on a deliberately American answer to DeepSeek, just received the implicit endorsement of the Pentagon's procurement organization. That is a strong signal for open-weight AI's commercial viability in regulated industries, even before Reflection has shipped a publicly downloadable model.

The deal also makes the GenAI.mil platform an unusually large managed-AI deployment. With 1.3 million users and tens of millions of prompts in five months, it is on a per-month rate roughly comparable to mid-sized SaaS AI rollouts. The fact that it was built and scaled to that level on government cloud infrastructure inside five months is a useful data point for any enterprise considering a similar internal platform.

The Other Side

Not everyone reads Friday's announcement as a clean win for the Pentagon's diversification strategy.

Defense procurement analysts at Breaking Defense have argued for weeks that diversification has a real cost: the Pentagon trades the depth of Anthropic's forward-deployed engineering for the breadth of eight vendors with smaller embedded teams. Pentagon Under Secretary Emil Michael himself acknowledged the tradeoff in February when he told Fortune about a "whoa moment" inside Defense leadership when officials realized how dependent classified operations had become on Anthropic's engineers. None of the new entrants are believed to have matched that level of in-environment support.

Civil liberties groups have pushed back from the opposite direction. The ACLU and Electronic Frontier Foundation have filed comments arguing the Pentagon's "all lawful purposes" framing leaves the door open to exactly the kinds of autonomous-weapons and domestic-surveillance use cases Anthropic refused to authorize. Their concern is that a list of eight willing vendors makes those use cases easier to pursue, not harder.

Anthropic itself has not commented publicly on Friday's announcement. The company's statement of record remains the lawsuit and the May 1 status of the injunction, both of which will return to court in coming weeks. CFO Krishna Rao's earlier filing warned that the supply chain risk designation could cost Anthropic "multiple billions of dollars" in 2026 revenue. The injunction blunted that, but the procurement consequence on display Friday is something the courts cannot reverse.

The Bottom Line

For four months, the question hanging over the Pentagon-Anthropic dispute was whether the U.S. government had the technical depth to replace Claude on classified networks at all. Anthropic's lawyers wagered it did not. The Pentagon's CTO insisted it did, and started signing deals to prove it.

On Friday, the proof arrived. Eight AI companies, three of the world's largest cloud providers, the dominant chip vendor, the largest open-weight frontier startup, and a satellite-internet operator are now contracted to put advanced AI on the same classified networks Claude used to dominate. The Pentagon is no longer betting its operational AI stack on one company that imposes ethical conditions. It is betting on eight that do not.

What the procurement list does not resolve is the underlying constitutional question. Judge Lin found in March that the Pentagon's retaliation against Anthropic was likely unlawful. The D.C. Circuit allowed procurement to proceed during appeal. Friday's contracts will be signed and operational long before the underlying case is decided on the merits. By the time the courts settle whether the government had the right to ban Anthropic from federal procurement, the question may be moot, because the government will already have built a stack that does not need it.

As Pentagon CTO Emil Michael told Bloomberg in March: "I need redundancy." He has it now. The cost was the company that built his original capability.

Sources

• Classified Networks AI Agreements (U.S. Department of War, May 1, 2026)
• Pentagon inks deals with Nvidia, Microsoft, and AWS to deploy AI on classified networks (TechCrunch, May 1, 2026)
• Pentagon strikes deals with 7 Big Tech companies after shunning Anthropic (CNN Business, May 1, 2026)
• Seven AI firms agree to deploy tech in Pentagon classified networks (The Hill, May 1, 2026)
• Pentagon makes agreements with 7 companies to add AI to classified networks (Nextgov/FCW, May 1, 2026)
• Pentagon announces deal with seven AI companies for classified systems (Al Jazeera, May 1, 2026)
• Pentagon partners with major AI companies for use in classified networks (Washington Times, May 1, 2026)
• Pentagon reaches agreements with leading AI companies (NBC News, May 1, 2026)
• Microsoft, Amazon Hand Pentagon More Control Over AI Systems Use (Bloomberg, May 1, 2026)
• Judge blocks Pentagon's effort to 'punish' Anthropic by labeling it a supply chain risk (CNN Business, March 26, 2026)
• Anthropic loses appeals court bid to temporarily block Pentagon blacklisting (CNBC, April 8, 2026)
• Inside Reflection AI: The $20B Open-Model Startup That Has Yet to Ship (Turing Post, March 2026)