Profitable and $66 billion in the red, in the same quarter. What Anthropic's IPO filing reveals about how AI finally started paying off

On October 26, Anthropic, the company behind the AI assistant Claude, made its S-1 registration statement public with the U.S. Securities and Exchange Commission. It plans to list on Nasdaq under the ticker "ANT."

Most first reports led with two numbers: $16.3 billion in first-half revenue and an $80.7 billion net loss. Revenue was up about 14 times from a year earlier. The loss looks like proof that AI still can't make money.

The quarterly table tells a different story. Operating income for April–June was a positive $974 million, up from a $1.7 billion loss the quarter before. The quarter it finally made money was also the quarter of its biggest loss ever. That contradiction turns out to explain almost everything about how an AI lab's finances work.

Start by breaking down the $80.7 billion first-half loss. Here is how the filing's income statement lays it out.

Source: DATA TIMES table based on the income statement in Anthropic's S-1

The operating loss was $690 million, less than the interest the company earned on its cash. Nearly the entire loss sits in the two fair-value lines.

Amazon's notes: $6.2 billion of principal, carried at $108.6 billion

Most of that markup comes from convertible notes held by Amazon.

Anthropic took money from Amazon in this form in 2023 and 2024. At the end of June, $6.2 billion of principal was still outstanding, with no interest attached. In exchange, Amazon can convert at fixed prices of $2.09 and $5.05 per share, which would give it about 1.84 billion shares.

Anthropic's shares have soared since. Its valuation jumped from $380 billion in February 2026 to $965 billion in May. The right to buy shares at about $2 each became enormously valuable.

U.S. accounting rules require Anthropic to mark these notes to their current value every period. Because the notes are Anthropic's liability, every increase in their value shows up as a bigger debt and a matching loss.

The result:

• End of 2024: $14.1 billion
• End of 2025: $41.7 billion
• End of June 2026: $108.6 billion

Notes with $6.2 billion of principal now sit on the balance sheet at $108.6 billion, about 17 times as much. The more the company is worth, the bigger its loss. The filing says so directly: increases in its valuation drive this expense up.

The other $8.8 billion line is the same kind of charge. It comes from valuing standby financing agreements Anthropic signed with Google, Amazon and Broadcom in 2026.

No cash leaves, though some of it will outlast the IPO

None of this is cash. Amazon receives no money, and Anthropic's bank balance doesn't move. At the end of June, Anthropic held about $131.7 billion in cash and short-term investments.

The notes are due to convert into shares at the IPO, and the filing's pro forma per-share figures already strip this charge out.

Some of it will stay, though. Conversion is capped so that Amazon doesn't hold more than 12% of the company (Amazon can waive the cap). Anything above the cap converts quarter by quarter afterward. Until it all converts, swings in Anthropic's share price will keep showing up as gains or losses in its results.

So what about the real business? Line up the quarterly figures and a clear trend appears.

Source: DATA TIMES table based on the quarterly results in Anthropic's S-1. The Q3 2025 loss includes a $1.5 billion copyright class-action settlement.

In early 2025, Anthropic spent $2.50 on compute for every $1 it earned. By the second quarter of 2026, it was spending 48 cents.

Over that year, revenue grew about 15 times ($0.79 billion to $11.56 billion) while compute costs grew four times ($1.39 billion to $5.56 billion). The filing itself highlights that gap as evidence of its compute efficiency.

The compute was bought 12 to 18 months ahead

There's a reason the turn was so sharp. According to the filing, Anthropic secures most of its compute through multi-year contracts, typically 12 to 18 months before the capacity becomes available. It buys first and fills it later.

All of that compute goes into one shared pool. Anthropic shifts it among research, model development, serving customers and internal use, wherever it is worth most at the moment, across four kinds of chips: Amazon's Trainium, Google's TPUs, and GPUs from Nvidia and AMD.

Think of a restaurant that signed the lease on a huge kitchen and dining room more than a year before opening. While the room is half empty, rent dwarfs sales. When the crowds arrive, rent stays the same and revenue climbs. In the first half of 2026, Anthropic was the restaurant whose tables were finally filling up.

Serving customers costs 15% of revenue

The filing also breaks costs down by purpose. For January–June 2026:

• Cost of serving customers (almost all compute): $2.38 billion, or 15% of revenue
• Research and development (compute for building the next models, plus researchers): $9.74 billion, or 60%
• Sales and marketing (including compute for the free tier): $3.81 billion, or 23%
• General and administrative: $1.05 billion

Running customers' workloads costs only 15 cents on the dollar. The heavy line is research and development for the next Claude. As long as Anthropic keeps building stronger models, that cost will stay large no matter how fast revenue grows.

Usage-based pricing carried the growth.

Of the $16.29 billion in first-half revenue, $13.06 billion, or 80%, was consumption-based: customers pay for the tokens Claude processes. Subscriptions brought in $3.22 billion, or 20%. When customers hand Claude more work, revenue rises with it.

Each time costs per task fell, usage rose even faster

The surprising part is how price cuts played out. According to the filing, Claude Opus 5.5, released in September 2026, cut the cost per task by about 40% compared with Opus 5, through both lower prices and fewer tokens per answer.

You'd expect a 40% price cut to shrink revenue.

So far, the opposite has happened. The filing points to Opus 4.5, launched in November 2025: smarter than its predecessor and cheaper per task. After the launch, quarter-over-quarter revenue growth jumped from 48% in Q4 2025 to 132% in Q1 2026. The filing credits that to several factors together, including new products and broader adoption.

Existing customers show the same pattern. Customers who were already paying a year ago (excluding individual subscription plans) now pay about nine times as much, which is what the filing's net revenue retention of about 900% means. As costs per task fell, companies handed AI many times more work. The filing also points to Claude Code and Claude Cowork being used for long-running tasks rather than one-off questions.

The customer count exploded as well. Customers paying $100,000+ a year went from about 1,500 at the end of 2025 to more than 11,900 by September 2026. More than 1,900 pay over $1 million, and about 250 pay over $10 million. Below them are roughly 40 million customers spending less than $100,000 a year.

Even big companies spend less than 0.1% of their costs on it

One more detail stands out: the filing's own argument for how much room is left.

More than half of the Fortune 500 pay Anthropic at least $100,000 a year. Yet on average, that spending is less than 0.1% of those companies' total expenses. Even its ten biggest Fortune 500 customers spend less than 1%.

For a large company, 0.1% of expenses is still pilot-project money. Anthropic's pitch is that this can grow to 1% and beyond. That's the company's view, but the numbers do show that AI spending is still a sliver of big-company budgets.

Outside the U.S., customers generated $4.52 billion in the first half, 28% of revenue. The filing doesn't break this down by country, but it names international expansion as a meaningful driver of future growth.

If the profit comes from filling compute bought in advance, the biggest risk is the mirror image. If revenue growth slows, compute costs as a share of revenue climb again.

Three contracts, at least $269 billion

The filing lists its major compute contracts. Each is a commitment to pay regardless of actual usage.

Source: DATA TIMES table based on Anthropic's S-1

Those three alone total at least $269.4 billion. Anthropic also expects to pay $125 billion to $150 billion for more TPU capacity from Google and others, delivered between 2028 and 2031.

It has also signed deals with SpaceXAI, Elon Musk's SpaceX subsidiary, for Nvidia GPU capacity: up to $120.5 billion through 2029. Once all of it is delivered, expected in December, it will cost about $3.1 billion a month. In April–June, Anthropic's revenue averaged about $3.85 billion a month. That one supplier would take the equivalent of 80% of its current monthly revenue. (After non-cancelable periods ending in January and March 2027, either side can cancel these deals with 90 days' notice.)

In our CoreWeave story, we looked at the take-or-pay contracts that hold up the compute factories. Anthropic is one of the companies on the paying end of those promises.

80% of revenue comes from models less than six months old

The second risk is how much revenue depends on the newest model. According to the filing, model versions released in the previous six months generated more than 80% of July–September revenue. When a smarter model comes out, customers switch quickly.

To keep revenue, Anthropic has to ship a better model about every six months. The research spending we saw in Part 02, 60% of revenue, is effectively the entry fee for that race.

There's seasonality too. In the two weeks ending January 2, 2026, customer spending fell about 20% from the previous two weeks, as companies froze code over the holidays. Compute costs don't take holidays, so swings like that hit profits directly.

A fight with the government, and copyright

The filing is also blunt about risks outside the business itself.

In February 2026, after Anthropic said its technology should not be used for mass domestic surveillance or fully autonomous weapons, the President directed federal agencies to stop using it. The Department of Defense (called the "Department of War" in the filing) designated Anthropic a supply-chain risk. Anthropic sued and largely won in a federal district court in August. But on September 25, the U.S. Court of Appeals for the D.C. Circuit ruled 2–1 that part of the designation was lawful. The filing warns that the dispute could lead to material revenue losses.

On copyright, Anthropic agreed in September 2025 to pay $1.5 billion to settle a class action brought by book authors, which is why its Q3 2025 loss was so large. The filing lists other suits still pending, including claims from music publishers over song lyrics and from Reddit over its posts.

EDITOR'S TAKE ── How editor-in-chief Ogasawara reads it

AI labs now buy compute first and race to fill it

Going through the filing line by line, what stood out was how much an AI lab's results depend on when, and how much, compute it bought, as much as on how good its models are.

When you see a headline about an AI company's huge loss, split it in two: paper losses that grow as the company's value rises, and losses from the business itself.

At Anthropic, the first kind came to about $80 billion and the second to less than $1 billion. The headline number can't tell them apart.

The clearest gauge of the real business is compute cost as a share of revenue. Above 100%, every sale falls further behind the compute bill. Well below 100%, money is left over for research and people. The fall from 251% to 48% is the first time an AI lab has shown a business that can make money. When OpenAI and other labs file to go public, it will be the first number worth looking for.

Anthropic's profit appeared when revenue overtook the cost of compute it had locked in 12 to 18 months earlier. That race is now permanent: new compute arriving on one side, revenue growth on the other.

From December, SpaceXAI's capacity alone costs $3.1 billion a month, and payments under the $114.5 billion Google and $110 billion Amazon deals are only starting to ramp up. If revenue grows as fast as the compute arrives, the profit holds. If it slows, the ratio climbs back above 100%, as it was throughout 2025.

The April–June profit is one quarter of unaudited numbers. What matters is whether revenue keeps pace with compute from July onward. The July–September revenue figure, still blank in the filing, will be the first test.

From the customer's side, the number that stuck with us was "less than 0.1% of total expenses." Even America's largest companies are spending that little on it so far.

At the same time, every drop in cost per task led companies to hand AI many times more work. Cheaper AI found more uses, and total spending went up, not down. The filing repeatedly describes the work shifting from single questions to long-running jobs like coding and building documents.

For anyone setting an AI budget, the useful comparison is no longer "price per request" but "price per job done." Unit prices can fall while the invoice grows, because there's more work to hand over. As long as that work costs less than having people do it, Anthropic's revenue suggests companies will keep handing over more.