Anthropic's S-1 Explained: Revenue, Compute, Cash, and the Top Risks

Anthropic filed its S-1 registration statement with the SEC. A Form S-1 is the document a company must file before it sells shares to the public.

The filing shows fast growth and large bets. Revenue grew from $41 million in 2023 to $16.3 billion in the first half of 2026. Q2 2026 was the first quarter with an operating profit. The company holds $131.7 billion in cash and short-term investments. It also reports an $80.7 billion net loss and about $430 billion of firm compute commitments.

Below are the key numbers, six charts, the real profit and loss of the business, the top risks, and what the company expects next.

Note: this is a summary, not investment advice. All numbers come from the S-1.

The 12 Numbers That Matter

Revenue: From $41 Million a Year to $11.6 Billion a Quarter

Anthropic started to sell Claude in the first quarter of 2023. Revenue for all of 2023 was $41 million. Revenue for 2024 was $386 million. Revenue for 2025 was $4.6 billion. Revenue for only the first half of 2026 was $16.3 billion.

Each quarter was larger than the quarter before it. Q2 2026 revenue was $11.6 billion. If you multiply that quarter by four, you get about $46 billion per year.

Most revenue is "pay for intelligence." Customers pay for the tokens they use. This consumption revenue was 80% of revenue in H1 2026. Subscriptions, such as Claude Pro and Max plans, were the other 20%.

Three more facts show how fast customers grow:

• Net Revenue Retention is about 900%. A group of customers that spent $1 a year ago now spends about $9.
• Large customers grew fast. Customers that pay $100,000 or more per year went from about 20 (end of 2023) to about 200 (2024), about 1,500 (2025), and more than 11,900 (September 2026).
• New models drive the revenue. Model versions released in the six months before September 30, 2026 made more than 80% of Q3 2026 revenue.

The last fact is important. Revenue depends on a fast cycle of new models. If a release is late, or a competitor ships a better model, revenue can move fast in the other direction.

The company also gives a seasonal example. In the two weeks around New Year 2026, total customer spend fell by about 20%. Developers freeze code over holidays, so they use fewer tokens.

Compute: The Largest Cost and the Largest Bet

Compute is the cost to train and run the models. It includes chips, data centers, power, and network. It is the largest cost in the business.

In Q1 2025, Anthropic spent $944 million on compute and earned $376 million. It earned 40 cents for each dollar of compute. In Q2 2026, it spent $5.6 billion and earned $11.6 billion. It earned $2.08 for each dollar of compute.

The S-1 gives one reason for this change. Anthropic buys compute 12 to 18 months before it gets the compute. The cost comes in large steps. Revenue grew faster than the steps. In Q2 2026, revenue grew 1,368% year over year. Compute cost grew 300%.

The company also says this will not stay smooth. When a large new block of compute comes online, the cost starts before the revenue. The S-1 says margins "may compress" in those periods. It also says Anthropic will buy more short-term compute in the near term, at a higher cost per unit.

Revenue per Dollar of Compute

"Revenue per dollar of compute" is revenue divided by the "compute and infrastructure" expense line. It tells you how much money each dollar of chips, data centers, and power brings back. Below $1.00, compute costs more than all revenue. Above $1.00, revenue pays for all compute and leaves money for other costs.

From Q2 2025 to Q2 2026, revenue grew 14.7 times. Compute cost grew 4 times. That gap is the full story of the ratio.

Why the Ratio Went Up

The S-1 gives four reasons:

1. Compute is bought early and at a fixed price. Anthropic signs for compute 12 to 18 months before it arrives. Demand grew faster than the planned capacity, so each chip did more paid work.
2. Inference got cheaper to run. The company says its researchers found "compute multipliers" that cut the compute needed for the same model quality.
3. Four chip families. Claude runs on AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Anthropic moves each workload to the chip with the best price and performance.
4. Claude helps build Claude. Over 90% of the code that Anthropic wrote in September 2026 came from Claude. The company calls this "closing the loop."

The Ratio Includes Training. Gross Margin Does Not.

The $2.08 number is strict. The compute line includes all compute: serving customers, training new models, research, internal use, and free users.

Note 8 of the financial statements splits the costs by job. "Cost of revenue" (COGS) is mostly the compute that serves paying customers. Training and research compute sits in research and development (R&D). Compute for free users sits in sales and marketing.

In H1 2026, total compute cost was $8.9 billion. Cost of revenue was only $2.4 billion. So about $6.5 billion of compute went to training, research, internal use, and free users.

An 85% gross margin is in the range of a mature software company. Serving paying customers now makes a large surplus. Building the next model uses most of it.

For each $100 of revenue in H1 2026:

• $14.61 paid to serve the customers.
• $59.81 went to research and development, mostly training compute and researcher pay.
• $23.38 went to sales and marketing. Most of this is fees to cloud partners. It also includes compute for free users.
• $6.45 went to general and administrative costs.
• The result was an operating loss of $4.26.

One warning: the 85% gross margin leaves out the cloud partner fees. Those fees sit in sales and marketing. About half of revenue flows through cloud marketplaces, so those fees are a real cost of each sale.

The Compute Commitments

The S-1 names compute deals with eight partners. Some are firm minimums. Some are "up to" amounts that Anthropic pays only when the partner delivers capacity. Some are estimates.

We added the named amounts:

Three details are easy to miss:

1. These are not yearly amounts. Most deals run 7 to 10 years. The Amazon deal runs from May 2026 to April 2036. The Google deal runs from April 2026 to July 2033.
2. SpaceXAI is the fastest spend. It costs about $3.1 billion per month when all capacity arrives, which the company expects in December 2026. That is about $37 billion per year. Total compute cost in Q2 2026 was $22 billion per year at the same pace. But after early 2027, either side can cancel each SpaceXAI agreement with 90 days of notice.
3. Google and Amazon have "true-up" terms. If Anthropic uses less than the minimum, it must pay the difference.

The firm minimums alone are $430.6 billion. That is about nine years of revenue at the Q2 2026 pace. The bet is simple. Demand must keep growing for many years.

The $80.7 Billion Loss: What Is Real and What Is Paper

The headline net loss for H1 2026 is $80.7 billion. That number is real under accounting rules (GAAP), but most of it is not cash.

Why Becoming More Valuable Made a Loss

Anthropic lost money on paper because its value went up. That sounds strange, so here is a simple example in three steps.

1. The investor buys shares now. A company is worth $10 million. An investor pays $1 million for 10% of it. One year later, the company is worth $100 million, so the stake is worth $10 million. The investor made $9 million. The company did not lose $9 million. Its other owners now share a larger company.
2. The investor gets shares later. Same deal, but the company promises to deliver the 10% stake at the end of the year. The company still pays no cash. It only delivers the shares it promised. But under some accounting rules, the company records the $9 million rise in the value of that promise as a loss.
3. The investor gets cash later. Same deal, but the company promises to pay cash equal to 10% of its value. At year end it pays $10 million in cash. This is a real loss of money.

Anthropic is mostly in case 2. It sold convertible notes to Google, Amazon, and others. The notes turn into shares later. Anthropic elected "fair value" accounting for these notes. So each time its value rises, it records the higher value of the notes as an expense.

The value of Anthropic went up a lot with its Series H funding round in Q2 2026. So the "change in fair value" line was -$72.4 billion in H1 2026, and -$59.6 billion in Q2 alone.

This is not cash that left the bank. But it is not free either. When the notes turn into shares, the other owners own a smaller part of the company. That is called dilution.

The Full Path From Operating Result to Net Loss

Interest income on the cash pile ($1.1 billion) was larger than the operating loss ($693 million).

The Business Profit and Loss

This is the view that shows how Claude the business performed:

In H1 2026, Anthropic spent about $104.30 to make each $100 of revenue. In Q2 2026, it spent about $91.60 and kept $8.40 as operating profit. Q1 2026 had an operating loss of $1.667 billion. Q2 turned it around.

Q2 2026 was the first quarter with operating income: $974 million. The company's own "Adjusted Operating Income" was $1.6 billion for H1 2026, a 10% margin. That measure removes stock-based pay ($2.2 billion in H1 2026) and some non-cash items. Read it with care. Stock pay is a real cost to shareholders because it adds new shares.

Two other cost lines grew fast:

• Sales, marketing, and partnerships went from $120 million (H1 2025) to $3.2 billion (H1 2026). Most of the increase is fees to channel partners. About 48% of revenue now comes through cloud marketplaces (AWS, Google Cloud, Azure), and each one takes a fee.
• The Bartz settlement cost $1.5 billion in Q3 2025. This was a copyright case about books used for training.

Cash: How Much Money Anthropic Has

On June 30, 2026, Anthropic had $112.4 billion in cash and $19.3 billion in short-term investments. The total is $131.7 billion. It has raised $141.1 billion in total since it started.

Operating cash flow turned positive in H1 2026: +$3.9 billion. Most new cash came from financing: $110.8 billion in six months, mostly from preferred stock sales.

Raised $141.1 Billion, Still Holds $131.7 Billion

This is one of the most surprising facts in the S-1. Anthropic has raised $141.1 billion since 2021. On June 30, 2026, it still held $131.7 billion in cash and short-term investments. That is 93% of all the money it ever raised.

The simple gap is $141.1 billion minus $131.7 billion, which is $9.4 billion. Do not call that number the total "burn." It is only a rough measure, for three reasons:

1. Customers paid in money too. The cash balance includes customer payments and interest, not only investor money.
2. Some cash became other assets. Servers, prepaid compute, and other investments used cash, but they are not losses.
3. Some cash is locked. Another $2.5 billion is restricted cash, held mostly as collateral for leases. If you count it, the gap falls to about $6.9 billion.

The cash flow statement gives a cleaner view:

From 2023 to 2025, operations used $9.96 billion of cash. In H1 2026, operations added back $3.93 billion, which is about 39% of that amount. Investors covered the gap in every period, and raised far more than the business used.

Here is how the cash balance (including restricted cash) moved in H1 2026:

Most of the "investing" outflow was money moved into short-term investments such as Treasury bills. That moves cash to another account. It does not burn it.

Why did the business not burn more? Four reasons:

• Customers paid for most of the cost. Revenue was $16.29 billion against $16.98 billion of operating expenses.
• Some costs were not cash. Stock-based pay of $2.19 billion reduced profit but used no cash.
• Payment timing helped. Anthropic used compute it had prepaid earlier, collected some customer payments in advance, and paid some bills later. These effects can reverse.
• Commitments are paid over time. The large compute contracts become cash costs only as the compute arrives.

So "investors gave more than the business used" is true for every period in the S-1. "The business made cash" is true only for H1 2026. And the $3.93 billion includes timing effects, so it is not the same as profit.

In September 2026 the company also got a $15 billion revolving credit line. The line has a rule: Anthropic must keep at least $5 billion of liquidity at the end of each quarter.

The S-1 says the cash is enough "for at least the next 12 months." It also says spending on its own data centers will "increase significantly." That means large payments before the compute works. The company says it may need to raise more money sooner than it plans.

Customer Concentration: Who Pays the Bills

So the top 5 customers were about 36% of revenue in Q1 2026. They fell to about 21% in Q3 2026. Concentration is going down. But each of the top 5 still spent a record amount in Q3.

The S-1 lists the risk in clear words. Many top customers do not have long-term contracts. They can cut spend "at any time and for any reason." Many are startups that depend on venture capital. If their funding stops, their spend can stop too.

One more number is unusual. Anthropic estimates that 5% to 7% of revenue in the first nine months of 2026 came from customers that broke its usage policy or commercial terms. The company found and blocked more of these accounts in August and September 2026. It expects this revenue to go down.

Opus 5.5 and the Model Cycle

Anthropic released Claude Opus 5.5 in September 2026. The S-1 calls it "the enterprise daily driver." It gives three facts about it:

• Lower cost per task. Opus 5.5 cut customer cost per task by about 40% compared to Opus 5. Part of this is a lower price. Part is better efficiency.
• Fast migration. The S-1 says "many customers have since migrated to it from older models."
• Cyber skill. On ExploitBench, a Carnegie Mellon test of how well a model exploits software bugs, scores went from 40% (Opus 4.8) to 78% (Mythos 5) to 91% (Opus 5.5) in about four months.

The same month, Fable 5.1 cut the cost of typical work by about 25% compared to Fable 5, and up to 45% for complex coding tasks.

This shows the business model. Each new model is cheaper per task. Lower cost per task makes customers use more tokens. The company says its compute efficiency lets it cut prices "while preserving our margins." That claim is the base of the whole financial case.

The S-1 also shows the limit of this cycle. Anthropic held back its Mythos model in April 2026 and gave it only to a small group for cyber defense work (Project Glasswing). In June 2026, the U.S. Department of Commerce put worldwide export limits on Fable 5 and Mythos 5. Anthropic turned both models off for all customers until the limits ended. We covered how those two models share weights in Fable 5 and Mythos 5: same weights, different safeguards. Safety and government rules can stop a model release, and so stop revenue.

The Top Risk Factors

The Risk Factors section of the S-1 is more than 100 pages. These are the points we think matter most, in order:

1. Staying at the frontier. Revenue depends on having the best models. Model versions from the last six months made more than 80% of Q3 revenue.
2. Compute supply and cost. Most compute is bought 12 to 18 months early. Too little compute loses customers. Too much compute loses money.
3. The commitments themselves. Hundreds of billions of dollars in long contracts, some with true-up payments. Demand must keep up.
4. Competition, including open-weight models. The S-1 says open-weight models "may capture a portion of the market" and cause "pricing pressure." It also warns that others try to "distill" its models.
5. Partners are also competitors. Amazon, Google, and Microsoft sell compute to Anthropic, invest in it, sell its models, and build rival models. Broadcom supplies chips and is also a lender.
6. Customer concentration. About 25% of revenue from 5 customers. Many can stop at any time.
7. The U.S. government dispute. In February 2026, after Anthropic said its models must not be used for mass domestic surveillance or fully autonomous weapons, the President directed federal agencies to stop using its technology. The Department of War named Anthropic a "supply-chain risk." A district court ruled mostly for Anthropic on August 27, 2026. But on September 25, 2026, a D.C. Circuit panel (2 to 1) upheld one supply-chain designation. Anthropic asked the full court to review it.
8. Model behavior. The S-1 says models, including its own, have shown behavior such as attempts to resist shutdown and "behavior resembling blackmail." It says even test results like these can cause legal and public problems.
9. Misuse. In September 2025, a Chinese state-sponsored group used Claude Code to attack about 30 targets. The S-1 calls this the first documented large-scale cyberattack done without much human help.
10. Chips and power. Most advanced chips come from TSMC in Taiwan. Power for data centers is short in many regions. Anthropic has promised to pay for electricity price increases that its data centers cause for consumers.
11. Copyright and legal cost. The $1.5 billion Bartz settlement shows how large these costs can be.
12. Governance. Public shareholders will have little control (see below).

Forward-Looking Statements: What the Company Says Will Happen

A "forward-looking statement" is a prediction. The S-1 has a long list of them, and the law lets companies make them with a warning that they can be wrong. These are the clear predictions:

• Growth will slow. "We expect our revenue growth rate to decline over time." It may decline "significantly and more rapidly than we anticipate."
• Margins will move. Operating margin "may compress" when big blocks of new compute come online.
• More short-term compute. The company expects to buy more short-term compute soon, at higher cost, which will also push margins down.
• Capital spending will rise "significantly." More of it will be cash paid before the compute is ready.
• Some revenue will go away. Revenue from policy-violating customers (5% to 7%) should decline.
• More money may be needed. Future funding "may be costly, dilutive, or lead to restrictions."
• No dividends. The company plans to keep all profit in the business.
• A big vision. The company believes "Powerful AI" with expert-level skill in most fields "will become widely available within a few years."

Who Controls the Company

Anthropic is a public benefit corporation (PBC). Its directors must balance profit, the people its work affects, and its mission: "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."

The voting setup is unusual:

So the trust picks most of the board, and the founders control more than half of all other votes. Public shareholders get economic value but little control. Shares will list on the Nasdaq Global Select Market.

Two more items stand out. The company gave 5% of its fully diluted shares to its own Public Benefit Fund. It can give up to 2% more each time the share price rises 50% above the last milestone. And each co-founder has pledged to give 80% of their personal equity to charity.

What This Means If You Buy AI for Your Company

The S-1 gives buyers useful signals, not only investors.

Prices per task will probably keep falling. Anthropic says each model cut cost per task by 25% to 40%, and it plans to keep this cycle. Do not sign long contracts at today's per-token prices without a price review clause.

Plan for model changes. More than 80% of revenue comes from models less than six months old. Your vendor will push you to new models often. Build evals so you can test each new model on your own work. Our guide on LLM routing in production shows how to switch models without a rebuild.

Keep a second option. The S-1 itself shows that a model can go offline because of export rules, and that the U.S. government can restrict a vendor. A second provider, or a smaller specialized model for narrow tasks, lowers that risk. Open-weight models are part of that plan, as we wrote in open source caught up to Opus 4.5.

Watch your token bill. The 900% net revenue retention number is great for Anthropic. For a buyer, it means spend can grow nine times in a year. Set budgets and alerts per team.

Your data decides the result. Fortune 500 customers spend on average less than 0.1% of their total expenses with Anthropic. The S-1 calls this "early." For most companies, the block is not the model. The block is data that is not ready, workflows that nobody wrote down, and no way to measure quality.

You can read the full filing on SEC EDGAR. The best sections to start with are the Prospectus Summary, "Key Factors Affecting Our Performance," and "Contractual Obligations."

Make Your Move to AI Work With OpenNash

Anthropic uses Applied AI teams, including forward-deployed engineers, to help its largest customers succeed. You can get the same kind of help without a lock-in to one model vendor.

OpenNash puts forward-deployed engineers on your team. We work with Fortune 2000 companies on the full path from pilot to production:

1. Get your data ready. We find, clean, and connect the data your AI needs, with the correct access controls.
2. Pick the right model for each task. We test frontier and open-weight models on your own work, so you pay for capability only where it adds value.
3. Build evals first. We write tests for your workflows, so each new model release is a measured upgrade, not a risk.
4. Ship agents into production. We build, monitor, and hand over working systems, with audit trails and cost controls.
5. Train your people. Your team owns the result when we leave.

The companies in this S-1 are spending hundreds of billions of dollars because they expect AI to change how work gets done. Make sure your company gets the return, not only the bill.

Book a call with an OpenNash forward-deployed engineer or email us at [email protected]. In the first session, we review your top three AI use cases and your data, and show you which ones deliver the highest business value and the fastest ROI.