Anthropic just passed OpenAI in revenue. The Pentagon is trying to destroy it. Both facts are connected.

The numbers that nobody predicted

I have been tracking AI company financials for two years now. I have a spreadsheet with quarterly estimates, funding rounds, headcount, compute costs, the whole thing. And I will admit something: I did not see this one coming.

Anthropic just hit a $30 billion annualized revenue run rate.

That number alone is wild. But the context is what makes it historic. OpenAI, the company that most people assume is winning the AI race by a comfortable margin, is sitting at roughly $24 billion. Anthropic passed them. Quietly. Without a Super Bowl ad. Without a consumer app that 200 million people open every month. Without becoming a household name.

And here is the part that keeps rattling around in my head: the same week Anthropic became the highest-revenue AI company on the planet, a federal appeals court upheld the Pentagon’s right to blacklist them.

The company that said “no” to the Department of Defense is the one that is winning.

I think those two facts are deeply connected, and I think most of the analysis I have read misses why.

Thirty times in sixteen months

The growth trajectory is borderline absurd. Anthropic was running at about $1 billion annualized revenue at the start of 2025. By mid-year, $4.5 billion. By December, $9 billion. In February 2026, when they announced their Series G, it was $14 billion. Now, in April, $30 billion.

That is 30x in sixteen months. The $9 billion to $30 billion jump happened in four months.

I keep reading comparisons to Slack or Zoom during COVID, but those do not feel right. Slack went from $630 million to $900 million in its breakout year. Zoom went from $623 million to $2.6 billion. Those were 1.4x and 4x respectively. Anthropic did 30x.

The closest comparison I can think of is AWS in its early years, and even that took longer.

Where the money is actually coming from

This is where the story gets interesting if you are building anything in AI right now.

OpenAI’s revenue is heavily consumer. ChatGPT subscriptions, the consumer app, and now advertising revenue. That model works. $24 billion proves it works. But it has a ceiling problem. Consumer subscriptions are price-sensitive. Ad revenue requires eyeball time. Both depend on keeping individual users engaged, and individual users are fickle.

Anthropic went the other direction. Enterprise APIs. Workflow integration. The boring stuff that does not generate headlines but generates contracts.

Over 1,000 enterprise customers are now spending more than $1 million per year on Claude. That number was 500 in February. It doubled in less than two months. And enterprise revenue makes up roughly 80% of Anthropic’s total, compared to OpenAI’s more consumer-heavy mix.

The distribution strategy is worth studying. Anthropic made Claude available on all three major cloud platforms: AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure. OpenAI, partly because of its deep Microsoft relationship, is still primarily available through Azure. If you are a CTO at a company running on AWS, the path to Claude is shorter than the path to GPT-4.

That sounds like a small thing. It is not a small thing. It is probably a $10 billion thing.

The efficiency gap nobody talks about

One number I keep coming back to. Anthropic spends approximately one quarter of what OpenAI spends on model training. By 2030 projections, OpenAI plans to spend around $125 billion annually on compute and training. Anthropic is projecting about $30 billion.

That 4x gap matters for a simple reason. Anthropic projects positive free cash flow by 2027. OpenAI anticipates $14 billion in losses this year alone, with breakeven not expected until 2030.

I wrote last week about OpenAI’s pre-IPO positioning around the “robot tax” proposal. When you read that policy paper through the lens of a company burning $14 billion a year heading toward a public offering, the framing makes a lot more sense. OpenAI needs the narrative to be about infrastructure scale, government partnerships, and civilizational importance because the current financial story is not great.

Anthropic does not need that narrative. The financial story speaks for itself.

Now let me tell you about the Pentagon

In February, Defense Secretary Pete Hegseth met with Anthropic CEO Dario Amodei. Hegseth had a simple demand: remove the safety guardrails from Claude so the military can use it for “all lawful use.” He gave Amodei a Friday deadline.

Anthropic had a $200 million Pentagon contract at stake. That is real money. And the specific guardrails Hegseth wanted removed were not trivial limitations on how many paragraphs Claude would generate. He wanted two things.

One, the ability to use Claude in autonomous weapons systems. Two, the ability to use Claude for mass domestic surveillance of American citizens.

Amodei said no.

Not “let us study this further.” Not “we need to consult our board.” No.

Anthropic’s position is that AI is not reliable enough to operate weapons, and there are no laws or regulations covering how AI could be used in mass surveillance. Both of those statements are factually correct, by the way. They are not ethical posturing. They are engineering assessments.

Hegseth responded by designating Anthropic as a “supply chain risk,” which is the government equivalent of putting someone on a blacklist. It means federal contractors cannot use Anthropic products. It means government-adjacent companies think twice before signing deals. It is designed to be economically devastating.

Two courts, two answers

Anthropic sued. Twice, in two different courts.

On March 26, U.S. District Judge Rita Lin in California blocked the Pentagon’s blacklisting. She called the administration’s moves “Orwellian” and said they could “cripple” the company. She found that the Pentagon appeared to have unlawfully retaliated against Anthropic for its views on AI safety, which is a First Amendment violation.

That was a big win. It lasted about two weeks.

On April 9, the D.C. Circuit Court of Appeals reached the opposite conclusion. The appeals court acknowledged that Anthropic would “likely suffer some degree of irreparable harm” but declined to issue its own order blocking the blacklist. The reasoning was that the “precise amount of Anthropic’s financial harm is not fully clear.”

So right now, as I write this, there are two federal courts looking at the same set of facts and reaching opposite conclusions. One says it is Orwellian retaliation. The other says the harm is unclear. A hearing with more evidence is scheduled for May 19.

Anthropic’s response to the appeals court loss was measured. “We’re grateful the court recognized these issues need to be resolved quickly.” Translation: we are not backing down.

Why “no” is the product

Most of the analysis I have read gets this wrong. The typical framing is: Anthropic is growing despite the Pentagon fight. Revenue is strong despite the regulatory headwinds. The company is succeeding in spite of its refusal to comply.

I think it is the opposite. I think Anthropic is growing because of the fight.

Think about what an enterprise CTO sees when they evaluate AI vendors. They see a company that was asked by the most powerful military on Earth to remove safety constraints, and the company said no. They see a company willing to walk away from a $200 million contract on principle. They see a company that would rather get blacklisted than build something it considers unsafe.

If you are putting Claude at the center of your customer data pipeline, your legal document review, your medical records processing, your financial analysis, that refusal is not a liability. It is the single most important signal you can get.

It tells you: this company will not do reckless things with your data either.

I do not think this is an accident. I think Dario Amodei understood exactly what signal the Pentagon fight would send to enterprise buyers, and I think the revenue numbers since February confirm it. The 1,000-customer milestone doubled from 500 in less than two months. That acceleration started right after the Hegseth meeting became public.

The OpenAI contrast

This puts OpenAI’s recent moves in a different light.

OpenAI converted from a nonprofit to a for-profit structure. It launched advertising in ChatGPT. It published a policy paper positioning itself as a partner to government on infrastructure and taxation. It is preparing for what might be one of the largest IPOs in history.

None of these moves are wrong, exactly. But they are all moves toward alignment with institutional power. OpenAI is positioning itself as the company that works with government, with advertisers, with Wall Street.

Anthropic is positioning itself as the company that tells them no when it matters.

Enterprise buyers noticed. Over 100 enterprise customers reportedly raised concerns to Anthropic about whether the Pentagon designation would affect their ability to continue using Claude. They were not calling to cancel. They were calling to check in. That is the behavior of customers who are loyal to the product and worried about losing access, not customers looking for an exit.

The compute bet

There is a hardware story here too, and it complicates the narrative in useful ways.

Anthropic just signed a massive compute deal with Google and Broadcom covering several gigawatts of next-generation TPU capacity, coming online starting in 2027. Gigawatts. Plural. That is data center scale that would have seemed science fiction five years ago.

This deal does two things. It gives Anthropic the compute runway to scale with demand that is clearly outrunning current capacity. And it deepens the Google relationship in ways that have strategic implications I have not seen widely discussed.

Google invested heavily in Anthropic. Google is now its primary compute provider. And Anthropic’s Claude is available on Google Cloud, directly competing with Google’s own Gemini models on Google’s own platform. That is a weird arrangement. It works as long as Anthropic keeps growing fast enough that Google benefits more from the partnership than from the competition. At $30 billion run rate, the math still works.

But there is a vulnerability here. If the Pentagon blacklist sticks and starts to spread to the broader defense-industrial complex, some of that enterprise revenue could be at risk. Anthropic warned the court that the designation could cost billions. Over 100 enterprise customers raised concerns. The May 19 hearing matters more than most people realize.

What I would tell a builder right now

I talk to a lot of founders building on top of AI APIs. Most of them are trying to figure out which model to bet on, which provider to integrate with, which direction the market is going. The Anthropic story teaches one thing above everything else.

Your constraints are your positioning.

The things you refuse to build are as important as the things you build. Anthropic’s two red lines, no autonomous weapons and no mass surveillance, are not marketing copy. They are product decisions that shaped the entire company’s trajectory. Every enterprise buyer who reads the CNN headline about Hegseth threatening to blacklist Anthropic over “woke AI” gets a very clear signal about what Anthropic will and will not do with its technology.

In a market where every AI company is technically capable of roughly the same things, the differentiator is trust. And trust is built by the things you refuse to do, not by the things you can do.

I would also pay attention to the distribution strategy. Anthropic being available on AWS, Google Cloud, and Azure while OpenAI is mostly on Azure is not a footnote in a press release. It is the entire reason enterprises can adopt Claude without rearchitecting their cloud infrastructure. Meet customers where they already are. This is not a new idea but it is amazing how many companies building AI products forget it.

And the efficiency numbers matter. Anthropic spending 4x less on training while generating more revenue is not just a financial metric. It tells you something about engineering culture. The company that figures out how to do more with less tends to keep doing more with less. That compounds.

The IPO question

There are reports that Anthropic could IPO as early as October 2026. At $30 billion run rate with 80% enterprise composition and a path to positive free cash flow by 2027, the public market story is straightforward. Growth investors love revenue acceleration. Enterprise investors love high-retention contract revenue. And the Pentagon fight gives it a narrative hook that consumer tech companies would kill for.

I do not know if they will go public this year. I do know that the IPO math works better than OpenAI’s right now. OpenAI at a rumored $300 billion valuation with $14 billion in projected losses and a consumer-heavy revenue mix is a harder sell than Anthropic at whatever multiple you put on $30 billion of mostly-enterprise revenue with a near-term path to profitability.

The wild card is the Pentagon. If the blacklist holds and expands, it introduces a regulatory risk factor that public market investors will price in aggressively. If Anthropic wins on May 19, that risk factor disappears and the IPO path is wide open.

One hundred enterprise customers called to check in, not cancel

I keep thinking about that detail. Over 100 enterprise customers, each spending over $1 million per year, saw that the Pentagon blacklisted their AI vendor. They called Anthropic. Not to pull out. To check in.

That tells you everything you need to know about product-market fit.

When your customers are more worried about losing access to your product than about being associated with a company the Pentagon just blacklisted, you have something that is very hard to compete with. You have a relationship where the switching cost is so high and the trust is so deep that even the federal government trying to kill your vendor does not change the calculation.

OpenAI has scale. It has brand recognition. It has ChatGPT and 200 million users. Those are real advantages. But it does not have enterprise customers calling to check in when things get hard. That is a different kind of moat entirely.

Where this goes next

May 19 is the date to watch. That is when the D.C. appeals court hears more evidence on the Pentagon blacklist case. The outcome will shape AI industry dynamics for years.

If Anthropic wins, it proves that AI companies can maintain safety standards even when the government demands otherwise. It sets a legal precedent around First Amendment protections for AI companies’ product decisions. And it removes the biggest risk factor from what could be the most consequential AI IPO of the decade.

If Anthropic loses, every AI company with government contracts faces a new question: can the Pentagon force you to remove safety guardrails by threatening your business? The downstream effects of that precedent would be enormous, and not just for Anthropic.

Either way, the revenue numbers have already answered the market question. Enterprise buyers have chosen. They picked the company that says no.

The most important thing Anthropic ever built might not be Claude. It might be the willingness to lose a $200 million contract rather than build something they do not think is safe. That decision, more than any model benchmark or API feature, is what $30 billion in annual revenue looks like.

Frequently asked questions

What is Anthropic’s current revenue run rate?

Anthropic’s annualized revenue run rate reached $30 billion in April 2026. This is up from $9 billion at the end of 2025 and approximately $1 billion at the start of 2025, representing 30x growth in roughly 16 months. The $9 billion to $30 billion jump occurred in just four months.

Has Anthropic actually passed OpenAI in revenue?

Yes. As of April 2026, Anthropic’s $30 billion annualized run rate exceeds OpenAI’s approximately $24 billion. The difference is driven primarily by Anthropic’s enterprise-heavy revenue mix, with over 1,000 customers spending more than $1 million annually, compared to OpenAI’s more consumer-weighted model built around ChatGPT subscriptions.

Why did the Pentagon blacklist Anthropic?

Defense Secretary Pete Hegseth designated Anthropic as a “supply chain risk” after CEO Dario Amodei refused to remove safety guardrails from Claude for military applications. Specifically, Anthropic refused to allow Claude to be used in autonomous weapons systems and mass domestic surveillance. Anthropic argues the designation is unconstitutional First Amendment retaliation for its views on AI safety.

What is the current legal status of the Anthropic vs Pentagon case?

There are conflicting federal court rulings as of April 2026. On March 26, California District Judge Rita Lin blocked the blacklisting, calling it “Orwellian.” On April 9, the D.C. Circuit Court of Appeals refused to block it, saying the financial harm was unclear. A further evidentiary hearing is scheduled for May 19, 2026.

How does Anthropic’s spending compare to OpenAI?

Anthropic spends roughly one quarter of what OpenAI spends on model training. By 2030, OpenAI projects spending approximately $125 billion annually, while Anthropic projects approximately $30 billion. Anthropic expects positive free cash flow by 2027. OpenAI anticipates $14 billion in losses in 2026 with breakeven expected around 2030.

Could Anthropic go public in 2026?

Reports indicate Anthropic could IPO as early as October 2026. At a $30 billion revenue run rate with 80% enterprise composition and a near-term path to profitability, the public market story is strong. The key variable is the outcome of the Pentagon blacklist case, with the May 19 hearing potentially determining timeline.