Hey Reader,

Anthropic, the creator of Claude, is going public.

Yes, that means the company’s employees and investors are going to make a lot of money, but that’s not what I want to talk about today.

Instead, I want to look at what Anthropic’s first-ever public financial disclosure (called an S-1) tells us about the company and about the AI industry as a whole.

(There’s some fascinating stuff buried in this stuffy financial document. 🤑)

Then I’ll show you how the big money Anthropic is making compares to the big debt I talked about last week: the money CoreWeave and lots of other companies are borrowing to build the data centers that power AI models.

In short, Anthropic just posted its first operating profit. It’s a five-year-old company that, last quarter, brought in more revenue than Starbucks and about the same operating profit.

But… even though the product is revolutionary… the business model isn’t. So I’ll also walk through the risks ahead for Anthropic over the next few years.

Anthropic is about as big as Starbucks

Anthropic’s filing shows a net loss of $80.7 billion for the first six months of 2026, or about $13 billion a month. But $72.4 billion of that is an accounting charge. “Operating profit” leaves that charge out, so for our purposes it’s a simpler and more accurate picture of how the business is doing.

Here’s the big money it’s been making recently:

• From April through June, Anthropic made a $974 million operating profit, its first profitable quarter in the six the filing shows. That’s about what Starbucks earned over the same three months: $980 million. (Claude and coffee, anyone? ☕)
• In the first half of the year, it generated about $3.4 billion in free cash flow, the cash left over after paying its bills and buying equipment. In all of 2025, it burned $2.6 billion running the business.
• It has $131.7 billion in cash and short-term investments in the bank.

My read: as of last quarter, Anthropic is basically a profitable company sitting on a mountain of investor cash.

That’s one quarter, not a trend yet. January through March was still a loss. And Anthropic books its computing power 12 to 18 months ahead, so when big new batches come online, the filing warns, the costs “land all at once.”​

So where is all that revenue coming from?

Our flat-fee plans barely matter. 80% of the money comes from enterprise and API customers paying by the token.

Anthropic’s revenue went from $386 million in 2024 to $4.6 billion in 2025, then $16.3 billion in just the first half of 2026. April through June alone brought in $11.6 billion, up 1,368% from a year earlier.

Most of that doesn’t come from people like us paying $20 or $200 a month. In the first half of this year, $13.1 billion (80%) came from usage-based fees, where customers pay for every chunk of text the model reads and writes. That’s how Anthropic’s Enterprise plan and its API both charge, so I’ll call these enterprise customers. Meanwhile, flat-fee subscriptions brought in $3.2 billion.

My take: those enterprise customers are making the flat-fee plans possible, at least for the power users. When Anthropic added weekly limits to its subscriptions last year, it said one person had “consumed tens of thousands [of dollars] in model usage on a $200 plan.” I’d like to think I use my $200-a-month Claude Ultra plan enough to be getting a good deal. 💰

Remember the Pentagon fight? Revenue grew 14x anyway.

In March, after the Defense Department labeled Anthropic a “supply chain risk,” Anthropic’s own CFO told a federal court the move “could reduce Anthropic’s 2026 revenue by multiple billions of dollars.” At the time, I called the whole fight a PR stunt by Pete Hegseth and his team.

The S-1 backs that up. Revenue for January through June was $16.3 billion, up from $1.2 billion a year earlier. That’s roughly 14x growth in the middle of the fight.

The label is still in place. A federal appeals court upheld it 2-1 in September, and Anthropic is weighing more appeals. But financially, it hasn’t mattered.

Anthropic owes $400 billion+ for computing power. That’s mostly good news for the data center landlords.

Anthropic has signed up for more than $400 billion in computing contracts it must pay whether or not it uses the capacity, mostly with Broadcom, Google, Amazon, Microsoft and AMD. Last week, I pointed out that CoreWeave’s good credit comes mostly from its big, financially strong customers. When Meta signs the lease, lenders treat CoreWeave’s loan almost like Meta’s own debt. Anthropic became a CoreWeave customer in April, and a tenant with $131.7 billion in the bank and positive cash flow is the tenant every landlord wants.

So in general, I think the S-1 is good news for the companies selling chips and data centers. But there are two catches.

Catch 1: the timing problem from last week hasn’t gone away. Anthropic signed long, fixed bills. Its customers didn’t. In the filing, Anthropic says customers can cut usage or cancel “with limited or no notice,” and it has only $10.8 billion of signed future revenue. Anthropic itself warns of:

“periods of excess or underutilized compute and committed cash outflows that do not produce immediate revenue.”

CoreWeave owes its lenders for years. Anthropic owes its landlords for years. Anthropic’s customers owe it nothing beyond what they use this month.

Catch 2: Anthropic is winning because it has the best models right now, but that can change fast. Switching AI providers is easy. In October, the investment bank Jefferies estimated that Kimi K3, an open model from China’s Moonshot AI, performs at about 95% of Claude Fable 5. And on Oct. 28, Google announced Gemini 4 Carbon, which edged out Claude Opus 5.5 for the top spot on Artificial Analysis’s closely watched leaderboard, 59 points to 58. (For now, only a small group of cybersecurity teams can use it.) The nightmare scenario for Anthropic is a cheaper open model that beats Claude outright.

So beyond the timing mismatch, the data center landlords could also end up with a customer mismatch. If the leaderboard suddenly changes, the demand for computing power follows the new winner, but the long-term contracts stay with the old one.

The S-1 proves an AI lab can make money. It doesn’t prove the same lab will still be on top in the 2030s, when many of those data center bills are still coming due.

The biggest invention since electricity still has Netflix’s business model

Here’s my bigger take. I think AI models like Claude are the most exciting invention since electricity. And yet, under the hood, Anthropic’s business works like an ordinary software-as-a-service company, the kind you pay for monthly or by usage.

That means it has the two problems every subscription business has.

1. Churn: customers can leave whenever they want.
2. A thin moat: not much stops customers from switching to a competitor with a better product.

Netflix lives with this every day. It has to keep making new shows, because subscribers cancel if the good stuff dries up. Anthropic is in the same race, except the new shows are new models, and the race never ends. And old models become obsolete within months, whereas old shows at least have rewatch and sequel potential. If Claude falls behind for even six months, I think a big chunk of that $16.3 billion in revenue could move somewhere else fast.

You’d think the company behind the most important invention in more than a century wouldn’t have to worry about the same thing as the company behind Squid Game. But here we are. 😅

Use the best model today, and make switching cheap tomorrow

For your business, the lesson comes from the customer side of the S-1. Anthropic’s biggest risk is that customers can walk away with little notice. That’s also your leverage.

Use whichever model does the job best right now. (Today, for a lot of work, that’s Claude.) But avoid long contracts you don’t need, and build your prompts and workflows so you could move them to a different model. (We call this the “platform-agnostic” approach in our programs.) If a cheaper model catches up next year, you want to be the customer who switches, not the one stuck paying for last year’s favorite.

Until next time,

Rob Howard
Founder, Innovating with AI

👋 Thanks for being a part of Innovating with AI. You’re in good company – some of our 175,000 readers work at Google, Apple, Microsoft and IBM… plus Mazda, Puma and JPMorgan Chase.

❤️ Did you love (or hate) today’s newsletter? Reply to this email to tell me. I personally read every single message (without AI).

📬 Share this post (no paywall). If a friend forwarded you this email, subscribe here.