Is the Pentagon about to nuke Anthropic? Yes - they pull support and government moves on them

Pentagon Nearly Done Purging Anthropic From Classified Networks — Trump Already Called Them a “Radical Left, Woke Company”

The Department of Defense is days from finishing what President Trump started in February: ripping Anthropic’s Claude models out of America’s classified warfighting stack.
Under Secretary of Defense for Research and Engineering Emil Michael told reporters this week that about 90 percent of the Pentagon’s classified AI workloads have already moved off Anthropic. All Maven Smart Systems / Palantir work left months ago. The rest is on track to be gone by the end of October.
That is not a rumor. That is the Pentagon’s tech chief, on the record, at the National Defense Industrial Association’s Emerging Technologies for Defense conference.
The official line is “multi-model diversity.” The political line is simpler. A San Francisco public-benefit corporation tried to write the rules of engagement for the U.S. military. The Commander-in-Chief told them no.
The Fight Was Never Really About “Safety”
Anthropic was first onto classified networks. Claude was wired into intelligence and warfighting systems. Then the Pentagon asked for a contract term every other serious defense vendor already lives with: the government can use what it buys for all lawful purposes.
Anthropic would not drop two red lines — mass surveillance of Americans, and fully autonomous lethal weapons. The company wrapped that refusal in the language of ethics. Defense Secretary Pete Hegseth and the White House heard something else: a private lab demanding veto power over lawful military operations.
On February 27, Trump did not mince words on Truth Social:
“THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS.”
He called Anthropic “Leftwing nut jobs,” said they tried to “STRONG-ARM the Department of War” into obeying their Terms of Service instead of the Constitution, and directed every federal agency to cease using Anthropic’s technology. Six-month phase-out for the Pentagon. “We don’t need it, we don’t want it, and will not do business with them again.”
Hegseth followed with the nuclear option in procurement language: Anthropic was designated a national security supply chain risk — the same kind of label Washington uses when it wants a vendor treated like a hostile node in the stack, not a partner. Contractors were told to get off Claude. The Air Force Research Lab even pushed industry to purge Anthropic products ahead of the department-wide deadline.
Anthropic sued. A Northern District of California fight produced one ruling the company likes to wave around. Michael says the more important case sits at the D.C. Circuit — “a higher court… more of a national security-oriented forum” — and that the supply-chain risk is still live because that court has not issued a preliminary injunction. Expect noise from the lawyers for another month or two. The operators are not waiting.
Amodei’s World: In a Few Years, Labs and Governments Are the Only Players That Matter
This is the part Anthropic’s fans never want next to the Pentagon story.
CEO Dario Amodei has spent 2026 writing and speaking as if the next few years collapse the map. If scaling laws hold “only a year or two longer,” he says, the world gets what he calls Powerful AI — “a country of geniuses in a datacenter.” Fifty million minds, all sharper than any Nobel winner, packed into racks. He has said it cannot be more than a few years before AI is better than humans at essentially everything.
He is not describing a crowded consumer app market. He is describing a new form of national power.
“If AI really will soon be ‘a country of geniuses in a datacenter,’ or anything remotely close to it, then AI is likely to be the dominant source of military and economic power for any nation,” Amodei wrote in Policy on the AI Exponential. A country with it versus a country three years behind, he argued, is Marines versus swordsmen.
Then he said the quiet part about companies. It is “somewhat awkward,” he admitted, but the next tier of risk after authoritarian governments is AI companies themselves. They run the datacenters. They train the frontier models. They sit on hundreds of millions of users. Historically, he noted, firms like the East India Company have “capture[d] the state or adopt[ed] quasi-state characteristics.” Advanced AI, he warned, may become “so powerful that it cannot safely be fully entrusted to either governments or companies.”
Read that from Frisco, Texas instead of Howard Street in San Francisco and the meaning is obvious. In Amodei’s few-year horizon, the only actors big enough to matter are governments and the handful of labs that built the machines. Anthropic’s whole pitch is that it should be the lab that survives that squeeze — because it is the “responsible” one, the one that writes the safety papers, the one that wants governments to gain the power to block or reverse model deployments it decides are too dangerous.
Washington heard a company that wants to be both the vendor and the regulator. Trump’s answer was: no private terms-of-service gets to outrank the Constitution, and no San Francisco board gets to tell the Department of War how it fights.
That is why “woke” is not just a slogan in this fight. Officials close to the administration spent months asking whether Claude was a woke model — the same concern that produced Trump’s executive order on “Preventing Woke AI in the Federal Government.” Anthropic’s Davos circuit, its Biden-era hires, its Open Philanthropy orbit, and Amodei’s habit of lecturing Washington on what the military may not do all fed the same file.
The Replacement Stack Is Already Live
The Pentagon did not sit in the dark after the breakup. It signed classified-network deals with a roster that pointedly excludes Anthropic: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle, among others. GenAI.mil has been loaded with Grok, ChatGPT, and Gemini.
Michael’s assessment of the operators who actually use the tools should sting Anthropic more than any court filing:
Warfighters on Palantir’s Maven who have seen newer ChatGPT “have liked it a lot better than Anthropic.” They found Grok “way faster than both of them.” Each model has different attributes depending on the mission. The original sin, Michael said, was being “solely reliant on one model” in the first place.
That is the operational smackdown. Not a press release. A migration.
Even the court fight has not stopped the purge of classified workloads. A judge can argue about retaliation and the First Amendment. The Department of War can still refuse to put the next targeting aid, intel summarizer, or agent workflow on a vendor that will not sign “all lawful uses.”
On X, the split is the same argument in shorter sentences. One camp says Anthropic got flagged as a “risk” for refusing to drop safety rules — proof that Washington is punishing conscience. The other camp says a lab that productizes doomerism, lobbies for licensing regimes only the richest firms can survive, and then tries to bind the U.S. military is not a conscience. It is a power play.
Both things can be true at once: Anthropic drew a line on domestic mass surveillance and fully autonomous kill chains, and it tried to make that line a condition of doing business with the most powerful military on earth. Trump decided the second part was the deal-breaker.
What “Smack Them Down” Looks Like in Practice
Here is the scoreboard as of September 11, 2026:

February 27: Trump orders every federal agency off Anthropic. Hegseth issues the supply-chain-risk designation. OpenAI announces classified-network access the same day.
Spring: Eight companies sign classified AI agreements. Anthropic is not in the photo. Air Force contractors get purge memos. Civilian agencies start getting the same guidance.
Summer: Lawyers fight. A California judge gives Anthropic a talking point. The Pentagon keeps moving metal.
This week: Michael says 90 percent of classified Anthropic workloads are gone. Maven is already off. October is the finish line. Grok and ChatGPT are what the warfighters are comparing now — and they are not asking for Claude back.

Amodei still talks as if the species has a few years before the datacenter becomes a country. The Pentagon just told him his country does not get a veto over theirs.
That is the story. A lab that spent years arguing it must remain at the frontier so the “safe” company is the one left standing just watched the U.S. government treat it as the vendor it can live without.
Trump is not having it. The Department of War is not having it. And by Halloween, if Michael’s timeline holds, classified Claude is a legacy problem — not a warfighting one.

The Whatfinger News Team: Sgt K and Ben