Anthropic filed a $965 billion IPO. The same week, the Pentagon said it still can't be trusted.

$965 billion.

That's what Wall Street thinks Anthropic is worth. On 1 June 2026, the company confidentially filed to go public, and that was the valuation attached to the S-1. The most valuable private AI company on earth, by most measures.

The same week, the Pentagon filed a document in federal court saying it still considers Anthropic a supply chain risk, won't reverse its position, and wants the DC Circuit to uphold the ban.

So we have two documents, both filed within days of each other, both about the same company. One says "most valuable AI company on the planet." The other says "national security threat." Both are currently true simultaneously. Nobody seems to know what to do with that.

I use Claude Code every day. I've written about Anthropic's models, their leaks, their product decisions. I thought I had a rough picture of how the company operated. Turns out there was a layer I hadn't followed closely enough, a months-long legal battle, a presidential ban, a federal judge calling the whole thing "Orwellian," and a subplot involving the same company the Pentagon won't trust being embedded inside the NSA.

I'll try to catch you up.

The Refusal

In late January 2026, Anthropic was in negotiations with the Pentagon over a contract worth roughly $200 million. The specific capabilities being discussed involved autonomous weapons systems and mass surveillance of US citizens.

Dario Amodei said no.

His position, grounded in Anthropic's Responsible Scaling Policy, was clear: Claude wouldn't be used for autonomous weapons or for mass domestic surveillance. Full stop. Not "we'd need to review the specifics." Not "let's find a framework that works for both sides." No.

$200 million is a number that clarifies most ethical positions very quickly. Anthropic apparently has unusually good posture.

What's worth understanding, though, is that Anthropic's position wasn't "we won't work with the military." That's the version that got simplified in a lot of the initial coverage. The actual position was more specific: they were willing to support defensive cybersecurity applications, intelligence analysis, logistics, and non-weapons use cases. They drew the line at weapons development and domestic surveillance specifically.

There's a real difference between those two things. One is a blanket anti-military stance. The other is a specific ethical limit with a clear rationale. "We won't build tools designed to kill people autonomously" is actually a more defensible position than most headlines managed to capture.

Whether you agree with where Anthropic drew the line is a separate conversation. But the line itself was specific, reasoned, and expensive. Most companies would have found a way to say yes. Some would have found a consultant to write a report explaining why building autonomous weapons was actually fine, ethically speaking, given the right governance framework. Anthropic found a way to say no. And then things got complicated.

The Retaliation

In late February 2026, the Trump administration ordered all federal agencies to stop using Anthropic products. The Pentagon formally designated Anthropic a "supply chain risk." Government employees across dozens of agencies lost access to Claude overnight.

Being designated a "supply chain risk" by the Pentagon because you won't build autonomous weapons is genuinely one of the more Orwellian things I've read in a while. You're not a supply chain risk because your products are unreliable or your company is financially unstable. You're not a supply chain risk because there's a vulnerability in your API or a problem with your data handling. You're a supply chain risk because you declined to help build something you think is dangerous. The designation is doing a lot of work there that the words don't quite cover.

On 9 March 2026, Anthropic sued the Pentagon.

On 26 March, Federal Judge Rita Lin blocked the ban. Her ruling described it as "classic First Amendment retaliation." She also wrote: "Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government." (A federal judge actually used the word "Orwellian" in a ruling about the Pentagon's conduct. I didn't add that for effect. It's in the text.)

The Trump administration appealed on 2 April to the Ninth Circuit, which handles appeals from the Northern District of California. Separately, Anthropic had filed a petition in the DC Circuit under a different statute (the Federal Acquisition Supply Chain Security Act) challenging the supply chain designation directly. The DC Circuit denied a stay in that proceeding on 8 April, which meant the designation stayed in place while both cases proceeded in parallel. On 17 April, Dario Amodei met with senior White House officials, including Susie Wiles and Scott Bessent, in a meeting described as "productive." In late April, Axios reported the White House was workshopping an executive order to restore civilian access to Anthropic products. That order has not been issued.

On 19 May, the DC Circuit heard oral arguments. Three judges. They appeared divided. Midway through the hearing, Pentagon lawyers offered to reconsider their position. That offer sat on the table for two weeks.

On 4 June, Pete Hegseth formally rejected Anthropic's request for reconsideration. He "reaffirmed the original conclusions." The Pentagon isn't moving.

So the DC Circuit must now rule.

The NSA Contradiction

Here's where the story gets genuinely strange.

While Anthropic is suing the Pentagon for barring Claude as a supply chain risk, approximately six Anthropic engineers are embedded inside the NSA. They're helping deploy Mythos, Anthropic's more advanced model, for offensive cyber operations. The work involves operations directed at China and Iran.

Pentagon tech chief Emil Michael confirmed publicly in May 2026 that Mythos was being treated as "a separate national security moment" from the ban, not a formal legal carve-out in the court record, but a practical distinction the government drew between supply contracts and classified intelligence partnerships.

In May 2026, CNBC reported that Pentagon tech chief Emil Michael confirmed Anthropic was still blacklisted, but described Mythos as "a separate national security moment," effectively acknowledging that the same company the Pentagon won't trust as a supplier is being used for classified intelligence operations.

Let's sit with that for a second. The Pentagon says Anthropic can't be trusted with a supply contract. The NSA apparently disagrees strongly enough to have Anthropic's engineers working inside the building, on offensive cyber capabilities, against two of the United States' primary strategic adversaries.

One commenter wrote: "Anthropic said no to the Pentagon. Wall Street priced the refusal at $965 billion. The word 'no' is now a line item in a prospectus." That's a sharp way to put it, and it captures the financial dimension accurately. The Anthropic IPO S-1 discloses the Pentagon ban as a material investor risk. You're now being asked to invest in a company whose ethics are simultaneously its selling point and a disclosed liability.

The same organisation that the Pentagon won't trust with a supply contract is embedded inside the nation's top signals intelligence agency. I don't have a neat way to resolve that. I'm not sure anyone does.

What it does complicate is the "ethical company" framing that got applied to Anthropic after the original refusal. They're not refusing all government intelligence work. They're drawing a specific line: no autonomous weapons, no domestic mass surveillance. But offensive cyber operations against foreign adversaries, conducted from inside the NSA? That's a different line. The line exists. It's just not where some people assumed it was.

One observer put it this way: "Blacklisted by the Pentagon as a supply chain risk, filing for an IPO, expanding Mythos inside the NSA, and now cheering a potential White House executive order. That's not a leadership strategy. That's a launch window." That's probably unfair. But it's also difficult to argue with entirely.

Where It Stands Now

Three clocks are running simultaneously.

The first is the DC Circuit ruling. Hegseth's 4 June rejection means there's no negotiated resolution. The three-judge panel (Rao, Katsas, Henderson) has to decide. The judges appeared divided at oral arguments on 19 May. There's no timeline on when the ruling comes.

The second is the IPO. The S-1 is filed. The roadshow is coming. Whatever the DC Circuit decides will land somewhere in that process. If the court upholds the ban, it becomes a material risk event in the middle of Anthropic's public market debut. If the court strikes it down, it becomes a headline in the roadshow deck. Either way, the ruling is now effectively a valuation event.

The third clock is the late August Pentagon phase-out deadline. Federal agencies were given until September to fully transition away from Claude. That deadline hasn't moved.

There's also the White House wildcard. The workshopped executive order on civilian access hasn't appeared. The Amodei meetings described as "productive" haven't produced anything visible. The White House and the Pentagon appear to be in two different conversations about this, and they haven't reconciled yet.

As of early June 2026, the DC Circuit hasn't ruled. No executive order has been issued. The September deadline is approaching. The IPO is in motion.

The thing that strikes me about this situation is how many different people can look at it and reach completely different conclusions. The Pentagon sees a supplier who refused a legitimate contract and poses a security risk. Investors see a company with principled leadership worth nearly a trillion dollars. Civil liberties advocates see a First Amendment victory. The NSA apparently sees a capable partner for offensive cyber work. Anthropic's critics see a company that draws its ethical lines selectively. Anthropic's supporters see a company that held its line when it really cost them something.

All of those readings are, to varying degrees, accurate. That's what makes this more interesting than a simple good-guys-versus-bad-guys story.

What This Means for AI Vendor Selection

When you choose an AI vendor, you choose their values under pressure.

Not their values in a press release. Not their values in an ethics policy document. Their values when the most powerful customer on earth is standing in front of them with $200 million and an implicit "or else."

Anthropic just ran that test in public. The result is interesting, and it's worth paying attention to even if you're not making decisions at the Pentagon's scale.

Here's the thing: most organisations that use AI tools haven't thought about what their vendor has said they won't do. They've thought about pricing, API limits, context windows, and whether the model can handle their specific use case. Vendor ethics is an afterthought, if it features at all.

I'm not going to pretend our choice of Claude at Webcoda was this philosophical. We use it because it's good at code. Claude Code has genuinely changed how quickly we can move on certain kinds of work, and that's why we use it. Not because of a vendor ethics audit.

But it does matter what a company does when it's actually under pressure, not just what it says in a policy document. Anthropic walked away from $200 million, got banned by the federal government, won in court, got appealed, and is now filing an IPO with the outcome still unresolved. That's a real data point about how the company operates.

The NSA angle makes it more complicated than a simple "company with principles refuses bad contract" story. They're drawing a line, but it's a specific line, not a blanket one. Offensive cyber work against China and Iran is apparently on the right side of that line, from Anthropic's perspective. Autonomous weapons and domestic mass surveillance aren't.

You might agree with where Anthropic drew the line. You might not. But knowing where the line is, and knowing they held it when it cost them real money, is more useful than any ethics statement they've ever published.

For organisations evaluating AI vendors right now, the questions worth asking have shifted. It's not just "what can this model do?" It's "what has this company refused to do?" and "what happened when someone tested that refusal?" The answers are now public record.

Anthropic's answer is: we refused autonomous weapons work, got punished for it, fought it in court, won (so far), and kept operating. The NSA carve-out means the picture isn't entirely simple. But the core answer is reasonably clear.

That's worth something. Possibly $965 billion worth, if you believe the IPO filing.

For more on the Claude product itself and how Mythos fits into the picture:

Related Article10 min read

Anthropic shipped Opus 4.8 in 41 days. The internet can't decide if it's a big deal.

Anthropic's newest model dropped 41 days after the last one. Developers are split three ways: some say it cured Claude's laziness, some say the...

Read full article

And if you want context on Anthropic's operational transparency (or occasional lack of it):

Related Article11 min read

512,000 lines of Claude Code leaked via npm packaging error

Anthropic shipped their entire Claude Code source in every npm install. A missing config line exposed KAIROS, 44 hidden features, and a stealth mode...

Read full article

The question "can AI companies say no to the most powerful customer in the world?" has been answered once. The answer is legally yes, practically complicated, and commercially very expensive. Anthropic said no to the Pentagon, got banned, won in court, got banned again via appeal, and is now filing an IPO while the next ruling is still pending. At some point this stops being a story about AI ethics and starts being a story about timing. But the ethics are real. The $965 billion valuation suggests the market thinks they're worth something too.

Key Takeaways

The three-act story:

• Anthropic refused Pentagon autonomous weapons work (roughly $200M contract) in January 2026
• Trump ordered a federal ban, the Pentagon designated Anthropic a "supply chain risk," Anthropic sued, and Judge Lin called the ban "Orwellian" when she blocked it in March
• As of early June 2026: Hegseth formally rejected reconsideration on 4 June, Anthropic IPO filed at $965B valuation on 1 June, DC Circuit ruling is pending, late August phase-out deadline approaching

The NSA contradiction:

• While suing the Pentagon, Anthropic has approximately six engineers embedded at the NSA helping deploy Mythos for offensive cyber operations
• The Pentagon's tech chief described Mythos as "a separate national security moment" (a practical distinction, not a formal legal carve-out)
• Anthropic's ethical line is specific (no autonomous weapons, no domestic mass surveillance) rather than a blanket refusal of government intelligence work

The vendor selection lesson:

• Ethics documents tell you what a company says
• $200 million decisions under pressure tell you what they actually do
• That information is now available for Anthropic, and it's worth factoring in

What's still unresolved:

• DC Circuit ruling (no timeline)
• White House executive order on civilian access (workshopped but not issued)
• September Pentagon phase-out deadline
• IPO roadshow timing relative to the ruling

---