Anthropic S-1 Breakdown

Anthropic S-1 Breakdown

Breaking down the growth, profitability, margins and compute needs and commitments behind the largest IPO ever

I’m Tanay Jaipuria, a partner at Wing and this is a weekly newsletter about the business of the technology industry. To receive Tanay’s Newsletter in your inbox, subscribe here for free:

Subscribed

Hi friends,

The much-awaited Anthropic S-1 dropped last week which dispelled any doubts if at all people may have had about the quality of the Anthropic business.

The tl;dr is that Anthropic is a wonderful business that is growing incredibly fast at a scale we’ve never seen before from any business in the past, and has done so with very limited cash burn while being asset light and even profitable in Q2 ‘26.

The big questions and risks associated with the business come down to the large compute commitments they have made going forward, and what the next few years will look like from a demand and profitability perspective in light of these commitments and in the face of increasing competition.

In this piece, I’ll discuss:

• Revenue and demand

• Profitability and burn

• Gross Margins and Compute Usage

• Compute Commitments

• Backlog and Future Compute Usage

I. Overview

Anthropic was founded in 2021 by a group of former OpenAI researchers, led by Dario Amodei. The company has had a meteoric rise since its founding and is the fastest company to $10B in revenue (and will be to $100B in revenue) and to a $1T+ valuation.

At its core, Anthropic builds the Claude family of models, across different sizes and price points. Mythos sits at the frontier, with Opus, Sonnet and Haiku below it allowing users to trade capability for speed and cost.

We can think of the Anthropic business as having a few layers:

• Models: Mythos, Opus, Sonnet and Haiku

• Platform: The API and developer tooling around it such as tool use, MCP, Agent SDK and managed agents

• Applications: Claude, Claude Code, Cowork, Tag and newer produts such as Claude Security, Claude Design

Customers can buy either the models directly via an API or consume them via Claude’s Applications (on both subscriptions or usage-based pricing).

But enough of the overview, let’s jump into it!

II. Revenue and Demand

Anthropic’s growth curve is the steepest any company has seen at their scale, growing from $10B run-rate to ~$65B+ run-rate so far this year. In fact, I would probably argue that the H1 of 2026 is one of the best half’s of execution and performance from any company ever.

Quarterly revenue went from $376M in Q1’25 to $2.0B in Q4’25 to $11.6B in Q2’26. Q3 was disclosed as 17.0B in the amended filing earlier today, a bit higher than the $16.8B reported.

About 80% of that revenue is consumption-based (customers paying per token, mostly through the API and the cloud marketplaces). The rest is subscriptions, which are generally monthly or annual.

From a customer base perspective, some of the numbers are staggering. Net dollar retention was around 900%, again numbers that we’ve never really seen from any company at anywhere near this scale. The number of customers spending $100K+ year is up 8x in 9 months to ~11.9K.

The best stats around how relatively early it is in terms of penetration of TAM are the following:

• Only half the Fortune 500 spend >$100K per year

• Of the ones that do spend over >100K per year, Anthropic is under 0.1% of their expenses.

What the right number looks like over time is a big open question, but as one datapoint, John Collison mentioned that Stripe spends roughly 1% of opex on tokens and he expects that number to go to 5-10% over time.

Even if the average business over time nets out at 1% of spend, that’s still a 10x growth from already penetrated accounts, not to mention the other half of sizable enterprises that don’t really use Anthropic much.

Customer concentration has come down too, with the top 5 customers in Q3’26 representing 21% of revenue, and only one customer representing greater than 10% of revenue.

III. Profitability and Burn

The S-1 reports expenses by the type (compute, people, GTM, other), which we can graph below. As is visible in the chart, expenses as a percent of revenue went from in the 300s down to 92% in Q2’ 26 as revenue scaled, their first profitable quarter on a GAAP basis.

The big decline and still the biggest bucket of spend is compute, which I’ll touch more on later, but went from 250% of revenue down to 58%. The only line that hasn’t declined significantly is the Sales, marketing and partnerships line-item, which I’ll also go into shortly and is largely because of payments to cloud channel partners.

We can also look at things in a bit more traditional way of COGS, R&D, G&A and Sales and Marketing on a per-half basis. Here, we see that both R&D is dropping a lot as a percent of revenue, as is COGS itself, driven by compute efficiency. One note is that Sales and marketing does include payments that one might consider should belong in COGS such as channel partner payments and cost to serve free users.

There has been a lot of fear, uncertainty and doubt about Anthropic’s reported $80B in net losses even ahead of the S-1. While the number is true, it is largely because of a change in marks of various convertible notes issued to Amazon ($72B) and to Google/Broadcom ($9.5B). These are non-cash losses that are largely an accounting artifact, and don’t really influence the fundamentals of the business since Anthropic now has other ways to raise cash and won’t be continuing to issue more of these notes on an ongoing basis. The one caveat here is that in case Anthropic has been getting sweetheart deals on compute because of Amazon/Google’s investments that may not be representative of compute pricing going forward. They do mention their compute agreements “are subject to negotiation upon renewal, and we may be unable to secure renewals on terms as favorable as our current arrangements.”

Focusing on operating margins on a simple GAAP basis, Anthropic’s numbers look as below, with the huge improvement over time clearly visible. Even in Q3, the GAAP losses are because of a $2B charge given equity issued for matching employee donations.

The other point to discuss is that Anthropic has been extremely efficient in regards to cash burn. Over the course of its lifetime, it raised ~$141.1B and still has $131.7B on its balance sheet, burning <$10B to get to ~$67B of run-rate revenue. To put that into context, it burned less absolute dollar than Uber ahead of IPO. Additionally, from an cash burn efficiency perspective, Airbnb is the only other reasonably sized company to go public that comes close:

IV. Gross Margins and Compute Usage

It’s worth digging a bit deeper on Anthropic’s use of compute and how to think about the real margins of this business. Compute essentially has three purposes, each which go in a different area of the P&L:

• Paid user inference → COGS

• Free user inference → Sales and Marketing

• Training & Research → R&D

As one can see below, the share on (free + paid) inference is around 34% in H1’26, which is down from ~50% in H1’25.

Inference has gotten much cheaper relative to revenue for Anthropic. Revenue grew 14x y/y in H1’26, and inference spend grew about 2.7x.

Part of that is Anthropic cutting costs per task with each release. Some of this can be passed back to customers of course, but also just allows for lower cost to serve. Per the S-1:

“With the release of Claude Fable 5.1 in September 2026, we reduced the cost of typical workloads by approximately 25% compared to Fable 5, with savings of up to 45% for complex coding and highly agentic tasks. Similarly, with the release of Opus 5.5 that same month, we reduced customers’ cost per task by approximately 40% compared to Opus 5.”

Anthropic reported 85% gross margins in H1 ‘26, but perhaps to get a better sense of the true incremental costs of revenue, we may want to consider a few additional factors:

1/ Free tier inference: This represents about 7% of compute and sits in sales and marketing. While its not uncommon to do this, some may believe that ultimately it should be in COGS.

2/ Cloud Partners: A reasonable share of Anthropic’s revenue comes from AWS, Google and Microsoft. Anthropic books that revenue gross and pays the clouds a fee, which sits in “sales, marketing and partnerships” rather than cost of revenue.

The share of revenue coming through the clouds went from 11% in FY23 to 48% in H1’26, and was down slightly to 44% in Q3 ‘26 (which is good for Anthropic’s overall margins)

By estimating the value of it from the filings, the fee works out to roughly 35% of the marketplace driven revenue (but its went down in Q3 to ~31%), or about 17% of total revenue, although the amended filing notes there were “changes in the terms of our channel partner distribution arrangements”. That’s a sizable toll, and it means the clouds make money from Anthropic by supplying compute to them, by distributing their products and by being shareholders. One could argue they should only count the revenue net of distribution fees (this is what OpenAI does) or you could put the costs in COGS instead of S&M (as they have done).

3/ Training & Research:

The S-1 has a line that gives us a sense of useful life of models:

“Model versions released in the six months ended September 30, 2026 accounted for over 80% of our revenue for the three months ended September 30, 2026.”

If that’s the case, one might argue that training is a bit closer to COGS, with a shelf life of a model being a bit longer than 6 months.

We can make these adjustments to get a view of the adjusted gross margins of the business as below, depending on what you decide to include in COGS vs not.

Ultimately, perhaps a better number for the true gross margins of Anthropic might be in the 45-64% range depending on how aggressive you want to be. One important thing to note is that regardless, none of this really changes the aggregate figures (since moving these costs to Gross Margins reduces R&D/S&M by those amounts) and either way they would be profitable in Q2 ‘26. But it may give us a better sense of the true operating leverage in the business and what long-term margins could look like.

V. The Compute Commitments

The risk factors in Anthropic’s S-1 focus a lot on Compute. Anthropic faced an acute compute shortage, which led they to pay up for compute from SpaceX.

The S-1 reveals a number of compute commitments in the future, which is necessary given the revenue growth they are seeing and the time lag of the ~18 months before the compute comes online. At the end of 2025, Anthropic had ~$105B of signed obligations. Today that number is close to $900B.

These include commitments to Broadcom, SpaceXAI, Google, AMazon, Microsoft, Nscale and Lambda and doesn’t include some that are yet to be sized (AMD).

Most of these can’t be canceled unless the other side breaches the contract, with the one exception being the SpaceXAI deal which has a 90 day notice to cancel from either side.

The updated S-1 gives us a rough sense of certain commitments and a high level timing. Note that this is only a rough estimate.

Its clear that Anthropic has a steep jump in compute payments starting next year to likely >$70B. This does seem manageable given their current cash position, current gross profits on revenue, their debt capacity and the additional capital they will raise in the IPO, but is something worth paying attention to over time.

Anthropic has been very capital light so far, and has used its cloud partners capex and rented compute from them, often funded by investments from the same partners in return for equity. Many of the new named compute deals will tend to follow that model. One exception is the Broadcom deal which works a bit differently in that they are leasing TPUs directly and those will sit on its balance sheet as finance leases. Add Anthropic’s first dedicated data center and its clear we’ll start to see somewhat of a shift in terms of financing compute and capex, and the S-1 does acknowledge that it expects “an increasing portion” of its compute investment to come as capex and up-front payments.

VI. Backlog and Compute Usage Going Forward

So compute is largely committed to on a long-term basis. What about the use of that compute?

The obvious use of course is demand from the API and subscriptions businesses it already has. Anthropic has grown revenue explosively so far, with 10x y/y growth for multiple years, and 900% NDR. And we just discussed how they’re still scratching the surface of their TAM.

This demand is clearly there today, but the slowdown in Q3 (50% quarter over quarter growth is still very impressive at that scale however!) highlights that offering frontier models at competitive price points will be critical to retaining and capturing that demand.

If you consider the remaining performance obligations of the business, which is essentially the contracted backlog, it was at ~$14B in September. That’s a pretty small backlog relative to the scale of the business, as the chart below shows when compared to other usage based businesses.

Now, when a business grows this fast, it’s hard to literally sign contracts with customers for future years out, which definitely has a role to play in this. But the other is that most usage is API driven, and switching models is easy, so businesses may commit to their cloud providers but not commit necessarily to individual model providers. This means at worst you have to re-earn revenue vs competitors every new model release and at best you just have low contractual revenue visibility going forward.

I don’t think demand will fall short, but its likely it may not immediately utilize the full size of short-term compute commitments since those just have other factors like adoption and change management. Fortunately, Anthropic has many other uses of the compute, namely:

The first is what Anthropic calls “closing the loop,” using Claude to speed up its own research. Anthropic already spends more on Claude internally at list prices than headcount, and increasingly view Claude as driving its own training data, evals, and producing next versions of their models. This ultimately drives demand, but could also open up other unexpected business lines as model capabilities advance.

The second is science. The S-1 describes the idea of Claude being used in automated discovery in bio through initiatives like their Claude Science product, an in-house molecular biology lab, and Claude operating microscopes, liquid handlers and even robotics arms.

VII. Closing Thoughts

We’ve covered the salient and key things in depth, but in closing, I wanted to quickly touch on a few other interesting tidbits from the filing.

• Distillation: 5-7% of Anthropic revenue came from users who were later found to violating their terms of service, namely, trying to distill their models. They have stepped up enforcement of this and so expect this to go down in Q4 onwards.

• Data Spend: Using the S-1 to do some rough calculations, we can approximate spend on data vendors at Anthropic had a steep jump in the last quarter. This number has usually hovered around 4-6% of compute expenses.

• Donation Matching: Employees who joined prior to May 2024 get a 3:1 match on charitable donations of their equity, with 933.7M shares reserved for it. Employees who joined after gets a 1:1 match. In Q3 alone, the value of shares issued as part of the matching was $2B.

If you have any comments or thoughts, feel free to tweet at me!

Thanks for reading! If you liked this post, give it a heart up above to help others find it or share it with your friends.

If you have any comments or thoughts, feel free to tweet at me.

If you’re not a subscriber, you can subscribe for free below. I write about things related to technology and business once a week on Mondays.

Subscribed

© 2026 Tanay Jaipuria
548 Market Street PMB 72296, San Francisco, CA 94104