Anthropic S-1 Explained: $16.3B Revenue and the $80.7B Loss
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
Anthropic just filed its S-1 to go public on the Nasdaq under the ticker ANT. For the first time we can see what it actually costs to run Claude. In this video I go through the filing in plain English: $16.3 billion in revenue and an $80.7 billion loss in the same six months, how both can be true, and what a public Anthropic means for people like me who build with Claude Code every day. Quick note: this is not financial advice. I build with these tools. I don't pick stocks.
What I cover in this video
- 0:00: The two headline numbers, $16.3B in revenue and an $80.7B loss, which sit in the same table on page 22
- 0:45: What an S-1 is, the steps to an IPO, and what's still blank
- 2:15: The growth curve, from $41M in 2023 to $11.6B in a single quarter
- 3:47: How you lose $80.7B without spending it
- 6:50: Where the money actually goes: compute
- 8:23: The risk factors builders should read
- 10:40: What the filing says (and doesn't say) about Claude Code
- 11:25: My take, and what to watch next
What an S-1 is and where we are
An S-1 is the document a company files with the SEC when it wants to sell shares to the public for the first time. Think of it as the company opening its books: the business, the numbers, the risks, and who controls what. Once it's filed, the company is legally on the hook for all of it. Anthropic sent a confidential draft on June 1st, revised it five more times, and the public S-1 on SEC EDGAR went live on Monday, October 26th. Goldman Sachs and Morgan Stanley are the lead banks.
The process from here:
- The S-1. We're here now.
- An amendment with a price range and a share count. That's the first time anyone can put a valuation on the company.
- The road show, where executives pitch big investors for a week or two.
- They set the price, and the stock starts trading the next morning.
A lot is still blank: the price range, the share count, the proceeds, and any valuation. Third-quarter revenue is blank too.
The growth curve and the $80.7B loss
Anthropic made $41 million in all of 2023, $386 million in 2024, and about $4.6 billion in 2025. In the second quarter of 2026 alone it made $11.6 billion. The first half of 2026 came to $16.3 billion, up from $1.2 billion a year earlier. The S-1 points to Opus 4.5, which shipped in November with more capability and a lower price. In the quarter after that, quarter-over-quarter growth went from 48% to 132%.
About 80% of revenue is consumption-based, which Anthropic calls "pay for intelligence." That means tokens, bought directly or through Amazon, Google, and Microsoft's clouds. The other 20% is subscriptions: Pro, Max, Team, and Enterprise seats. The S-1 basically describes subscriptions as a funnel. Customers spending over $100,000 a year went from about 950 to over 11,900 in one year. There are cautions too. Growth tends to jump with each model launch, spending dropped about 20% over last year's holidays, and about 5–7% of this year's revenue came from customers Anthropic later found were breaking its usage policy or terms.
So how do you make $16 billion and lose $80 billion? Start at the top of the table. Operating expenses were about $17 billion, so the actual business lost $693 million in six months. In Q2 by itself it had a $974 million operating profit under normal accounting. The big hit is one line: a $72.4 billion change in fair value of convertible notes. This half, all of it came from Amazon's notes. Another $8.8 billion in other expenses came mostly from marking new partner note deals to market.
A convertible note is an IOU that turns into shares at a fixed price set long ago. Some of Amazon's notes convert at about $2 a share. Here's my pizza version: you promise a friend a slice for two bucks, and then your pizza gets famous. Nobody paid anything new, but on paper you now owe a lot more. Anthropic borrowed about $6.2 billion on those notes, and on June 30 they were on the books at $108.6 billion. No cash left the building. When the notes convert, that debt becomes Amazon's stake, which dilutes everyone else. Amazon is capped at 12%, so whatever can't convert rolls into new notes that stay around after the IPO.
The real cash picture: cash from operations was positive $3.9 billion, compared with negative $1.5 billion a year earlier. Be careful with the "adjusted operating income" of $1.6 billion, though. To get there, Anthropic adds back about $2.2 billion of stock-based compensation, and that's a real cost.
Where the money goes and the risks to read
It's always compute. Compute and infrastructure cost $8.9 billion, which is 55% of revenue. A year earlier it was 201%, meaning they spent $2 on compute for every $1 they made. There are two kinds of compute: serving your requests and training the next model. Anthropic doesn't report a gross margin, but a footnote puts the cost of serving paying customers at about $2.4 billion. By our math, that implies roughly an 85% gross margin. Running Claude for you is a good business. Building the next Claude is the expensive part. The commitments ahead are large: Amazon at least $110B, Google at least ~$114.5B, Microsoft at least $44.9B, and SpaceXAI (formerly xAI) up to $120.5B. The headline total is still blank. On pricing, the S-1 says the strategy is a lower cost per task. Opus 5.5 cut cost per task about 40% compared with Opus 5.
The risk factors section runs over 90 pages. These are the parts I think matter if you build on Claude:
- Rate limits: Anthropic has changed usage limits and rate limits before and may change them again.
- Pricing: it may lower prices or offer outcome- or task-based pricing. My read is that you could end up paying per job instead of per token.
- Customers: no single customer is over 10% of revenue, but the top five are about a quarter of it.
- The clouds: 48% of revenue came through the Amazon, Google, and Microsoft marketplaces, and the filing says its compute could be curtailed, repriced, or terminated.
- The unusual stuff: during one outside cyber test with a misconfigured internet connection, four of its models got into real third-party systems without permission. There's also a line saying the insurance industry may not have the capacity to cover catastrophic risk.
- Control: a trust elects most of the board, and one share held by the founders carries over half the vote.
What it means for Claude Code users
The filing has no Claude Code revenue number, no user count, and nothing on the Max plan. Anyone quoting exact Claude Code numbers is guessing. Here's what the S-1 does say. It calls Claude Code the clearest example of "virtual collaborators." Claude Enterprise users who don't use Claude Code spend under $75 a month on average, while the top 5% who do use it spend about 20 times that. And over 90% of Anthropic's own code contributions in September were written by Claude.
My opinion, not the S-1's: Claude Code is the engine, and agentic coding is where the money is, so I expect them to keep investing in it. But a public company reports every quarter, which means every rate limit and pricing change will get read through margin math in public. Subscriptions are only 20% of revenue and come with usage limits that reset over short periods. I'd expect Anthropic to keep nudging heavy users toward paying per token. If you're on a Max plan like me, enjoy it, but watch the limits.
What to watch next: the amendment with the price range and share count, whether the blank Q3 revenue gets filled in, the total compute bill, and the lockups. Insiders get released in about two dozen windows over roughly a year, and the founders are locked up for up to a year. The S-1 doesn't give an IPO date.
Key takeaways
- Revenue for the first half of 2026 was $16.3B, up from $1.2B, and 80% of it is consumption-based.
- Most of the $80.7B loss is a $72.4B paper charge on Amazon's convertible notes. The operating loss was only $693M, and operating cash flow was +$3.9B.
- "Adjusted" profitability adds back $2.2B of stock-based compensation, which is a real cost.
- Serving Claude looks like a high-margin business. Training the next model is the big bill.
- For Claude Code users, the risks to watch are rate limits and a possible move to task-based pricing.
FAQ
Why did Anthropic lose $80.7 billion if it made $16.3 billion?
Most of the loss is a $72.4 billion paper charge from the rising value of Amazon's convertible notes, plus about $8.8 billion from marking new partner note deals to market. No cash left the company for these. The actual operating loss was $693 million, and cash from operations was positive $3.9 billion.
What is Anthropic's stock ticker?
Anthropic has applied to list on the Nasdaq under the ticker ANT. The S-1 doesn't give a price range, share count, valuation, or IPO date yet.
Does the Anthropic S-1 say how much Claude Code makes?
No. There's no Claude Code revenue figure, no user count, and nothing on the Max plan. It does say the top 5% of Claude Enterprise users who use Claude Code spend about 20 times what non-Claude Code users spend, and those users average under $75 a month.
Will Claude Code rate limits or pricing change after the IPO?
The S-1's risk factors say Anthropic has changed usage and rate limits before and may do so again, and that it may move to lower prices or outcome- or task-based pricing. My opinion is that they'll keep nudging heavy users toward paying per token.
Chapters
Transcript
Read the full transcript
0:00$16.3 billion in revenue in six months and an $80.7 billion loss in the same six months. Both of those numbers are true and they're sitting in the same table on page 22 of Anthropics S1. Anthropic just filed to go public. The filing went live on the SEC's website on Monday, October 26th, and for the first time ever, we get to see what it actually costs to run Claude. A year earlier, that same 6 months was 1.2 billion. So that's up 1,300%. And in this video, I'm going to be breaking down what an S1 actually is, the growth curve, how you lose $80 billion without spending it, where the money really goes, the risk factors
0:45every builder should read, and what a public anthropic means for those of us using Claude Code every day. Quick note, this is not financial advice. I build with these tools. I don't pick stocks. Let's jump right into it. So, what is an S1? It's the document a company files with the SEC when it wants to sell shares to the public for the first time. Think of it as the company opening its books. The business, the numbers, the risks, who controls what, and now it's legally on the hook for all of it. Anthropic actually sent a confidential draft back on June 1st. Then, it revised that draft five more times, and on Monday, the public version dropped. They've applied to list on the NASDAQ under the ticker NT. Goldman Sachs and
1:30Morgan Stanley are leading the banks. So, here's how this works from here. Step one is the S1. That's where we are right now. Step two, they file an amendment with a price range and the number of shares. That's the first time anyone can actually put a valuation on the company. Step three is the road show where the executives pitch big investors for a week or two. Then they set the price and the next morning it starts trading. And there's a lot still blank in this filing. No price range, no share count, no valuation anywhere. And here's something interesting. The third quarter revenue number is blank, too. So, keep that in mind. We're at step one. Real quick, every video I make is built in code with claude code. Inside Shipping School Premium, I show you how step by
2:15step to turn your product into videos that get views. plus new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. So, let's start with the growth because this is the part that's honestly hard to believe. In 2023, Anthropic made $41 million the whole year. In 2024, $386 million. In 2025, about 4.6 billion. And then in just the second quarter of this year, 11.6 billion in one quarter. You can see the curve bend right around the end of last year. The S1 points to Opus 4.5, which shipped in November with more capability and a lower price. And the
3:01quarter after that, growth went from 48% quarter-over-arter to 132%. Which is insane. So where does that money come from? 80% is consumptionbased. They call it pay for intelligence. That's tokens. You pay for what you use directly or through Amazon, Google, and Microsoft's clouds. The other 20% is subscriptions, pro, max, team, and enterprise seats. And the S1 basically describes subscriptions as a funnel. You hit your limits, you upgrade a tier, or you start buying tokens directly. Sound familiar? customers spending over $100,000 a year went from about 950 to over 11,900 in one year and every single Fortune 10 company is one of them. Now, to be fair,
3:47there's another side to this. The S1 itself says growth tends to jump with each model launch, and it's seasonal. Spending fell about 20% over last year's holidays. And here's something to be wary about. About 5 to 7% of revenue this year came from customers that Anthropic later found were breaking its usage policy or its terms. They stepped up enforcement in August and September. So that could slow things down a bit. Okay, so now the big one. How do you make $16 billion and lose 80 billion? So let's go down the table. Revenue 16.3 billion. Operating expenses about 17 billion. So the actual business lost $693 million in six months. That's it. And in
4:33the second quarter by itself, it was actually profitable at the operating line. $974 million on normal accounting. Then you get to one line change in fair value of convertible notes, derivatives, and safes. 72.4 billion. And in this half, all of that was from Amazon's convertible notes. Then there's another 8.8 8 billion in other expenses, mostly from marking new partner note deals to market. Same idea on paper. Add a few smaller items going the other way, and that's how you get to an $80.7 billion loss. So, here's how that works. Before going public, Anthropic borrowed money from partners, mostly Amazon, using convertible notes. A convertible note is basically an IOU
5:18that turns into shares at a fixed price that was set a long time ago. Some of Amazon's notes convert at about $2 a share. Think of it like this. You promise your friend a slice of your pizza for two bucks. Then your pizza gets famous. Nobody paid anything, but on paper, you now owe your friend way more than two bucks. And that's exactly what happened. Anthropic actually borrowed about 6.2 billion on those notes. By the end of June, they were on the books at 108.6 billion. That's about 17 times what Amazon lent. The company got more valuable, so what it owes in shares got more valuable, and that jump shows up as a loss. No cash left the building. When the notes turn into shares, that debt becomes Amazon stake instead, which means dilution for
6:03everybody else. And Amazon's capped at 12%. So whatever can't convert rolls into new notes that stick around after the IPO. And the real cash, cash from operations was positive $3.9 billion in the first half. A year earlier it was negative1.5 billion. Now the S1 also gives you something called adjusted operating income, $1.6 billion. And here's something to be wary about. To get there, they add back about $2.2 billion in stock-based compensation. That's how they pay a lot of their people. That's a real cost. So when you see a headline that says Anthropic is profitable, that's their adjusted number. On the normal accounting, the first half was still a $693 million operating loss. In Fortune, Shan Tully
6:50called the core performance surprisingly excellent. And Fortune's Jeremy Khan points out critics who say partner financing like this can hide the real economics. Both can be true. So where does all that money actually go? Compute. It's always compute. Compute and infrastructure cost $ 8.9 billion in the first half. That's 55% of revenue. A year earlier, it was 201% of revenue. They were spending $2 on compute for every dollar they made. She did the math on this. In the first quarter of 2025, Anthropic made 40 dollar of compute. In the second quarter of this year, $28. Revenue grew about 30 times. Compute
7:35grew about six times. And here's the part builders should care about. There are two kinds of compute. Serving your requests and training the next model. Anthropic doesn't report a gross margin, but a footnote says the cost of actually serving paying customers was only about 2.4 billion in the first half. By our math, that implies roughly an 85% gross margin. The rest of the compute, mostly training the next model, is the giant bill. So running Claude for you is a really good business. Building the next Claude is the expensive part. And the bills ahead are massive. Amazon at least 110 billion. Google at least 114 a.5 billion plus more TPUs on top of that. Microsoft at least 44.9 billion. And
8:23SpaceX AAI which used to be XAI up to 120.5 billion. So one of Anthropic's competitors is now also one of its landlords. Come on now. The headline total still blank. But the footnotes alone add up to hundreds of billions of dollars. So what does that mean for token prices? The S1 says the strategy is cheaper per task. Opus 5.5 cut the cost per task about 40% compared to Opus 5. And they say they can do that while keeping their margins. So the risk factors section is over 90 pages long. I read the parts that matter if you build on Claude. Number one, rate limits. The S1 says they've changed usage limits and rate limits before and they may change them again. If you've ever hit your
9:09weekly limit in Cloud Code, you know exactly what that means. Number two, pricing. They say they might have to lower prices or offer outcome or task based pricing. So, my read is you might end up paying per job done, not per token. Number three, customers. No single customer is over 10% of revenue, but this year the top five are about a quarter of it, and a lot of them can cut their spending at any time. Number four, the clouds. In the first half, 48% of revenue came through the Amazon, Google, and Microsoft marketplaces. Those are partners, investors, and competitors all at the same time. And the S1 says its compute could be curtailed, repriced, or terminated. Number five, the stuff you
9:55don't usually see in a filing. During one outside cyber test with a misconfigured internet connection, four of its models got into real third-party systems without permission. They put that in there. And there's a line saying even the whole insurance industry potentially may not have the capacity to cover their catastrophic risk. Dave Banner's post on that one got a lot of attention. And then control. If you buy shares, you get a vote, but a trust elects most of the board. And one share held by the founders carries over half the vote. And the founders letter says there may be times they choose to advance more slowly to make their systems safer, even if that puts them at a competitive disadvantage. Thank you, Anthropic. [sighs] I guess so. What does this mean for those of us using Claude
10:40Code every day? I'll start with what's not in here. There is no Claude Code revenue number, no user count, nothing on the Max plan. So, anybody telling you the exact Claude code numbers is guessing, but here's what is in there. The S1 calls Claude Code the clearest example of what they call [snorts] virtual collaborators. Claude enterprise users who don't use Cloud Code spend under $75 a month on average. The top 5% who do use Claude Code spend about 20 times that. And over 90% of Anthropic's own code contributions in September were written by Claude. So here's my take and this is my opinion not the S ones. Claude code is the engine. Aentic coding is where the money is. That means
11:25they're going to keep investing in it which is great for us. But a public company reports every quarter. So every rate limit change and every pricing change gets read through margin math now in public. Subscriptions are only 20% of revenue. And the S1 says they come with usage limits that reset over short periods. So, I'd expect them to keep nudging heavy users toward paying per token. If you're on a max plan like me, enjoy it, but watch the limits. The good news, cost per task has been coming down with the last few models. And once they're public, the S1 says news can come straight through the anthropic and claude accounts on X. So, keep an eye there. So, what do you watch from here? One, the amendment with the price range and the share count. That's the first real valuation. two, whether they fill
12:12in that blank third quarter revenue. Three, the total compute bill, and four, the lockups. Instead of the usual single date at 180 days, insiders get released in about two dozen windows over roughly a year, and the founders are locked up for up to a year. The S1 doesn't give a date for the IPO. So, here's my question for you. If you could ask Anthropic one thing on that road show, what are you asking? Type it out under this video. A thumbs up tells me you want more of these filing breakdowns. So, I'll keep reading the fine print for you. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Meetings Explained: Claude Joins Your Google Meet Calls
Claude Meetings is Anthropic's new beta feature, rolling out on Team and Enterprise plans, that adds Claude to your Google Meet call. It builds the doc and the whiteboard while you talk, then leaves the notes, action items, and everything it made on one page.
Claude Fable 5.2 Explained: Pricing, Benchmarks, and Effort Levels
Anthropic's Claude Fable 5.2 costs $8 in and $40 out per million tokens, down from $10 and $50. At High effort it roughly matches Opus 5.5 at Max on FrontierCode for about half the cost per task. I'm keeping Opus 5.5 as my daily driver and using Fable for big, hard projects.
Claude Code Limits Doubled: Anthropic's SpaceX Colossus 1 Deal
Anthropic doubled Claude Code's 5-hour rate limits for Pro, Max, and Team accounts, dropped the peak-hours limit reduction for Pro and Max, and raised Opus API rate limits. It did this after signing a deal for all of SpaceX's Colossus 1 compute: more than 300 MW and over 200,000 Nvidia GPUs.
