Anthropic IPO: S1 Breakdown

You may already have the hottest job in finance.

Did you automate accruals last month? Build an agent that scans for large transactions? Blow through your token limit twice in the past week?

Congrats, you’re a finance engineer.

Instead of booking JEs or wrangling workpapers by hand, finance engineers architect the systems that run accounting and FP&A.

Every week, Numeric’s Alec Schon (technical accountant turned product manager), writes a newsletter for finance engineers. Each send includes a new Claude skill you can download and use, plus his take on what’s changing in the field.

Whether you’re an aspiring finance engineer or seasoned one, its worth a read.

Anthropic’s S1 dropped yesterday morning, and it was such a momentous occasion that my wife made me pancakes.

(Technically she was making them for my 1 year old son, and I pleaded to be added to his order, but I like to think it was to celebrate this S1).

We’ve been together for 13 years, and the number of times she’s made me pancakes I can count on two hands. So it was a glorious and delicious human moment as I set off to work to cover a human-like technology that will either zap us from humanity or cure cancer and help us all live until 200.

For the record, we won’t be touching the third rail in this piece as to whether our existence is threatened.

I’m bringing the syrup today.

Anthropic filed to list on the Nasdaq under the ticker ANT. Their last private round was a $65 billion Series H that closed in June, at a $965 billion valuation. The price range is still blank, but bankers have told investors the deal could raise $100 billion or more at a valuation near $2 trillion.

Anthropic did $19.7 billion of revenue over the trailing twelve months ended June 30, 2026. The second quarter alone was $11.6 billion, up 1,368% y/y (so +14x lolol).

The three biggest questions from my PPOV (Pancake Point of View):

Have to admit, it’s the best shareholder letter intro I’ve read (if it was written by Claude IDK, but it slaps). It goes from we will try to cure cancer, to here’s why you might not want to buy this, to an explanation of what exponential looks like, and why they’re going to make some investments that look “irresponsibly large” on paper but are justified in this new world. It has it all.

They lead with the scientific angle: cure diseases and extend life. That’s very thoughtful positioning versus potentially killing us all.

Then they hit you over the head with four things they want any investor to understand: the potential of AI, the risks, how Anthropic is built differently, and “the ways in which our behavior may differ from that of a purely profit-maximizing company.”

It’s the same thing they do in recruiting, where candidates get an “anti-sell,” or a pitch “deliberately designed to be unappealing to those who are a poor fit.”

Like, listen, you’ll get enormously wealthy here but no one ever refills the joe and the snacks are mid.

And then they warn you of surprises coming down the pike:

They compare it to pharma, where huge upfront R&D only makes sense later in light of the returns from successful drugs, except at Anthropic “a similar process is occurring much faster and on a much larger scale.” Their stated approach is to be “bold but sane.”

They’ve also committed to keeping Claude free of advertising. And they say there may be times they choose to advance more slowly than they could “even if that puts us at a competitive disadvantage.”

They sign off asking you, as a prospective owner, “not to think of your purchase in purely transactional terms.”

If you read our Nscale S-1 breakdown, you’ve seen the five layers of AI:

Anthropic (mostly) lives at layer five. It rents lots of layers one through four from Amazon, Google, and Microsoft, and it trains and serves Claude across four chip families: Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs.

But it’s starting to reach down the stack. They call it Anthropic Direct: leasing chips directly and hosting them in dedicated facilities operated on their behalf, which now puts equipment leases and committed power costs on their P&L. The filing also mentions guarantees to developers to get data center capacity built.

There are four model classes: Mythos (including Fable), Opus, Sonnet, and Haiku. Each one sits at a different point on what they call the capability-cost-speed curve. Customers use the smart, expensive ones for coding and agents, and route the simple stuff to the cheap, fast ones.

They charge two ways:

The two models feed each other. When a subscriber runs out of the usage in their tier, they can move up a tier or buy more intelligence directly, which turns subscription revenue into usage revenue. A growing number of enterprises also sign non-cancelable annual or multi-year contracts with minimum spend commitments in exchange for a committed price list. Those mostly show up as usage revenue too.

Much of the diffusion of Claude has been organic. Prosumers use it for their own work and then champion it inside their companies, and by the time the sales team gets involved, Claude is often already in use.

The biggest benefit to their go-to-market is that it’s available through all three major cloud providers, so a customer can buy it inside the AWS, Google Cloud, or Azure relationship they already have. Anthropic keeps the contract with the customer and pays each cloud a fee / rake / vig / slice of the pie.

And then there’s the old-school channel model via consultants. Accenture, Deloitte, KPMG, and PwC all have Claude practices, and Salesforce and Snowflake embed it in their platforms.

They also see the private equity angle. The problem is how you get it into mid-sized companies who aren’t that tech forward, so they launched Ode with Anthropic, an AI services company built with PE firms to deploy Claude across their portfolios.

Sales, marketing, and partnerships came to 19% of revenue in the first half, and most of that was fees to channel partners.

A few weeks ago I wrote There are no new crimes, just new ways to calculate revenue after Higgsfield’s CEO said on 20VC that they calculate annualized revenue “the same way OpenAI and Anthropic do,” by taking the last 28 days of revenue and multiplying by 13.

They call it Run-Rate Revenue:

At the risk of being pedantic (think that’s how you use that word): it’s the last 28 days of usage x 13 + a month of subscriptions x 12.

But….. they don’t actually give us the number! They just define it. Lol.

Good thing they’ve been telling everyone and their estranged grandmother for months. Dario said on stage in April that they’d hit a $30 billion run rate, then the Series H announcement in May said $47 billion, and Bloomberg reported $65 billion at the end of July. All up from $9B to end 2025.

When the $30 billion number came out, OpenAI threw shade and said it was overstated by about $8 billion. Which actually makes sense considering they report revenue from cloud partners (that big growing segment) gross, not net of fees.

As for year end, The FT reported investors expect the run rate to land somewhere between $100 billion and $120 billion by the end of the year. And also Shane Gillis works for Kalshi and can apparently not predict the weather so good.

Net net: I think GAAP revenue is a much more honest view of what they can actually do than a run rate. Usage revenue moves around a lot. For example, during the two weeks ended January 2, total customer spend dropped about 20% compared to the two weeks before, because developer commits slow down during holidays and retail code freezes.

They also say recent run-rate growth was “driven in part by rapid initial customer adoption and expansion often characteristic of an early-adoption period,” and that they expect the growth rate to decline, possibly “significantly and more rapidly than we anticipate.”

I still bet you dollars to donuts we start hearing a lot more about 90-day rolling averages when that happens.

You can go back and forth on revenue growth expectations, but you can’t argue with the revenue efficiency. These are the highest efficiency numbers I’ve ever seen for a company.

Net revenue retention is about 900%, meaning customers who were spending a year ago are now spending about nine times as much. Snowflake was at 158% when it filed and GitLab was at 152%, and people thought those were crazy.

Some of this is a feature, not a bug, because of their pricing model. AI is pretty much synonymous with usage-based pricing now, and 80% of Anthropic’s revenue is consumption, so revenue goes up as customers use more.

While I don’t think this figure will stay at 900% forever, even if it is cut to 300% that would be S tier.

Also, it’s crazy how they can have 900% NDR but also be a commodity product with little to no switching costs.

Oh, and Revenue per employee is $6.6 million, up from $1.4 million a year ago. My rule of thumb for a public company is $450K. TTM revenue grew about 1,200% over the past year while headcount grew about 170%, from roughly 940 people at the end of 2024 to roughly 5,300 today.

I know I said everyone is talking about revenue too much, but you can kinda back into their revenue as of September, which they don’t disclose, using the revenue per employee and the employee count. They say TTM revenue per employee was over $6.6 million as of September 30, with approximately 5,300 employees, which puts trailing twelve-month revenue at roughly $35 billion. Subtract the three quarters we do have (Q4 2025 through Q2 2026, about $18.3 billion) and Q3 2026 comes out to roughly $16.7 billion, up from $11.6 billion in Q2. I love arithmetic.

They express their financial model using four drivers:

The first three have all been moving up fast. But the fourth one lags:

They typically commit to compute 12 to 18 months before it’s available, and once it comes online the expense hits the P&L whether or not the revenue it’s meant to support has shown up yet.

So every time a big tranche lands, margins take a step down. They tell us “we expect operating margin and Adjusted Operating Margin may compress in those periods.”

When demand outruns their long-term capacity, they also buy short-term compute, which costs more per unit, and they expect to lean on more of it in the near term.

Capital allocation for compute 101:

It’s one pool covering research, model development, serving customers, and internal use of Claude.

Revenue per dollar of compute is a really cool stat that companies should try to adopt internally.

Q2 revenue grew 1,368% year over year while compute and infrastructure expense grew 300%. If you run it quarter by quarter off their P&L, you can see the lag and then the catch up:

That’s a 5.2x improvement in compute efficiency over five quarters (because why not, give us a weird number of periods). So it’s like slow slow then everything all at once.

In my revenue-crimes piece I asked whether they were still selling a dollar for sixty cents, or maybe a dollar for 75 cents by now.

In Q2 they spent $10.6 billion of total operating expenses to bring in $11.6 billion of revenue, which works out to about 92 cents per dollar, or $974 million of operating income (an 8% margin) and the only profitable quarter in the filing.

Zoom out to the full first half, though, and it’s still a $693 million operating loss.

They pulled gross margin off the income statement last year when they switched to grouping expenses by type (compute, personnel, sales and marketing), but the footnotes still break out cost of revenue the old way. In the first half it was $2.4 billion against $16.3 billion of revenue, which is about an 85% gross margin, up from 40% in 2025 and negative 95% in 2024.

But the free tier compute and the cloud partners’ cut sit in sales and marketing, so the real number is lower.

Channel partner fees went up about $2.7 billion year over year in the first half, and the compute behind free usage is somewhere inside the remaining sales and marketing line, which caps it at roughly $600 million.

Load those back in and it’s more like the low 60s. And since they book cloud marketplace sales gross, the top line includes the cut that goes to the cloud providers, so the first half is closer to $13.6 billion than $16.3 billion net of it.

The accounting charges for stock-based comp need a full post. If you read our Oura breakdown, this is the same setup: most of their RSUs need two triggers to vest, time served and a liquidity event, and until the IPO is effective they can’t book the expense because the IPO isn’t considered probable.

There’s another $3.1 billion on other awards on top of that, and they’re using the accelerated method, which front-loads it. Part of the IPO proceeds is also going to cover the tax withholding when those RSUs settle.

This part is counterintuitive. Claude Opus 5.5 cut customer costs per task by about 40% compared to Opus 5, and their argument is that cheaper tasks make people use more of it, not less:

My number one question is whether the use cases keep growing as fast as the model costs are dropping.

Adoption curves are never an even path, and they’re buying compute 12 to 18 months ahead at a fixed cost, so if prices fall faster than people find new things to do with it, the compute bill doesn’t shrink with the revenue. Optimization waits for no man. The question is whether use cases keep growing once we hit a point where the models are “smart enough” for everything we do today.

Nonetheless, they launched 250 new model versions, platform features, and applications in the first nine months of 2026, and models released in the prior six months made up over 80% of Q3 revenue, so customers move to whatever is newest almost immediately.

The flip side is that the newest stuff is agentic, and agents chew through much more compute per interaction than a chat. The risk factors say that as agentic work grows, the compute needed to serve customers “could increase disproportionately relative to the associated revenue.”

Weird question: is it actually in their favor to not let the models drop too much, at a rate faster than we can diffuse them into our lives?

How much can they realistically grow wallet share at the largest companies? Over 50% of the Fortune 500 already spend $100K or more a year with them, but that spend is less than 0.1% of those companies’ total expenses on average. That actually felt low compared to how we (read: I) talk about budgeting for tokens. Even the top 10 Fortune 500 customers spend less than 1% of their total expenses on Anthropic, which is still eight times the average.

Weird question, but is there such a thing as having too many partners at the party? Amazon, Google, and Broadcom are investors, suppliers, and lenders all at once, and the filing points to how tangled it gets:

They’re also competitors, since all three sell their own AI models on the same clouds that resell Claude. The filing admits that being on their platforms gives them “visibility into our pricing and commercial terms.”

On the counterparty piece, there are large notes to Amazon, Broadcom, and Google. Amazon still holds about $6.2 billion of convertible notes that turn into non-voting stock at the IPO, up to its ownership cap. On top of that, there’s about $77 billion of undrawn financing capacity sitting with the same three:

And those are just the ones with notes. On the compute side, it’s a who’s who:

You’d normally want to diversify your counterparty risk, but when does having too many counterparties hurt you with the others? Or do they not give a crap, because Anthropic is the biggest counterparty? Does it get to a point where if you work with everyone, you can do whatever? Big dog make the rules!

It’s both leverage and cost. Remember - 48% of first-half revenue ran through the cloud marketplaces, which also make up 58% of receivables. And the risk factors admit they’ve “sought quantities of compute from CSPs in excess of what those partners were willing to provide.” So it’s not unlimited good will between them.

Being on all three major clouds is the biggest benefit to their go-to-market, and now they’re vertically integrating and building out their own compute on top of it. The total purchase commitments and lease agreements are blank, and more importantly I’d hate to be the lease accounting person at this company.

This risks section is robust, so here are the five I’d pay attention to, in plain English.

1. Everything is connected to everything.

So like a vortex of hell. Nice. Their own example: a bad copyright ruling forces them to retrain or limit their models, which hurts revenue, which makes it harder to pay for compute, which could trip covenants on their credit facility and trigger cross-defaults elsewhere. Like the broader US economy. Oh, have you seen The Big Short? That was a good movie.

2. The chips come from Taiwan, and the power comes from regulators.

It’s nuts that an island about the size of Maryland + Delaware is the lynchpin for the AI economic engine. By most estimates, TSMC makes about 90% of the world’s most advanced chips, and the filing lists the exposure: earthquakes, typhoons, tensions with China, and Taiwan’s own power grid. They say efforts to build capacity elsewhere “remain in early stages.” On power, they need deals with utilities and grid operators plus sign-off from public utility commissions.

3. Building are expensive

Most of the big compute deals are non-cancelable, and the total they’ve committed is blank.

4. Copyright is still a live wire.

Did you hear that it came out in this case that Anthropic spent millions of dollars buying print books, ripped off the bindings, cut out the pages, and scanned them one by one, then threw the originals away? Picture a first edition of The Great Gatsby going under the paper cutter. It would make my 2nd grade librarian Mrs. Walters weep. The judge actually said the scanning was fair use. The $1.5 billion they paid to settle Bartz in 2025 was actually for the pirated books they downloaded. Funny how that works The bigger risk now is fair use not holding up for AI training in other cases, and the filing says a ruling against another AI company “regardless of the specific facts” could set precedent that hits them too.

5. Your vote doesn’t count for much.

The Long-Term Benefit Trust elects a majority of the board, and the founders hold 50.1% of the vote. The filing says the structure “may effectively eliminate the ability of holders of our Class A common stock to influence important decisions, including a change in control.” Nobody is pulling a Kendall Roy here.

Good news though! They get the greenlight on the following:

There are five classes of common stock, and two of them are a single share each.

The price range is blank, but the reporting has bankers pitching about $2 trillion, roughly double the $965 billion Series H from five months ago. They’re also sitting on about $132 billion of cash and short-term investments before a dollar of IPO proceeds comes in, so enterprise value is closer to $1.87 trillion.

What does $2 trillion buy? It depends a lot on which revenue number you pick:

So the multiple is anywhere from about 17x to about 100x, depending on whether you believe the run rate or the GAAP. And if you strip out the cloud providers’ cut from the gross-booked revenue, every one of those multiples goes up by about a fifth.

For context, here’s where recent names landed in our breakdowns this year:

1. They listed their Threads account in case you need to contact them. The disclosure channels for investors include X, Instagram, LinkedIn, and Threads (@claudeai).

2. They switched auditors a year before filing. They dismissed Grant Thornton and engaged EY in October 2025. The filing says there were no disagreements.

3. They cut the vesting cliff in half. In January 2026 they changed the one-year cliff to six months on all stock options that hadn’t hit it yet, which is a nice recruiting move.

4. Their boomerang rate is in the founders’ letter. All seven co-founders still work there, and over three quarters of the first 100 employees are still there. Of the employees who left that they would have preferred to keep, 19% came back.

5. Shoutout to Societe Generale, listed in the fourth row of 22 banks. This bank caught a stray in my SpaceX breakdown. But I gotta say, you punch above your weight class.

The bull case is basically the efficiency stats. A business growing more than 10x a year with 900% net revenue retention is priced at something like 20x to 30x its run rate, which is less than what CrowdStrike and Cloudflare get on forward revenue with a fraction of the growth.

On trailing GAAP revenue, though, you’re paying about 95x, more than Palantir, for a company that buys its compute from Amazon, Google, and Microsoft, who sell competing models on the same clouds that resell Claude.

And it all hinges on my number one question, which is whether the use cases keep increasing as fast as the price per task keeps falling.

Never let your company be worth a dollar more than it is actually worth.

They said it, not me.

None of this is investment advice. I edit this while my wife watches Love is Blind Boston. I haven’t looked up yet, but based on the voices I think I went to college with one of the guys. TBD. So do your own homework. For information and entertainment purposes only.

CJ