Esta nota aún no está traducida al español, así que aparece a continuación en su inglés original.
· Nacho Planas
Anthropic opens its books
Anthropic went public on 19 November in the largest IPO ever. The prospectus finally puts audited numbers behind the run-rates. The revenue is real, the business already makes cash, and it has promised to spend close to a trillion dollars on compute, much of it with its own shareholders. What the filing shows, how management explains it, and what you have to believe at today's price.
Disclaimer: Not investment advice. I may hold positions in some of the companies mentioned. Figures are from Anthropic's final prospectus, dated 18 November 2026, unless I say otherwise. Trading prices, roadshow comments and analyst forecasts come from the press, and I name the outlet each time. Where a number is my own calculation, I say so. Links to every source are in the text and collected at the end.
The largest IPO ever, in the worst week since 2020
On Wednesday 18 November Anthropic sold 980 million shares at $83 each, above the $70 to $78 range it had marketed them at. That raised $81.3 billion, the most any company has raised in a listing. Seven of its suppliers bought another $26 billion of stock privately at the same price: AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. The shares opened on the Nasdaq on Thursday at $111.50 under the ticker ANT. They closed the day at $97.17 and closed Friday at $88.91, in what CNBC called tech's worst week on Wall Street since 2020.
How much the company is worth depends on which shares you count. On the roughly 13.2 billion shares outstanding after the offering, Friday's close values it at about $1.2 trillion. Counting options, unvested restricted stock units (RSUs) and the shares Amazon can still take for its convertible notes gives about 17.6 billion. On that basis the value is about $1.6 trillion, which is the number most of the press used (CNN had $1.76 trillion at Thursday's close). Both counts are my calculations from the prospectus.
In the money loop piece in October, I said the thing to watch was the first public prospectus from Anthropic or OpenAI, because it would replace run-rates and press leaks with audited numbers. It's here. The short version: the revenue is real and growing faster than anything a public company has reported, and the business already generates cash. To keep growing, Anthropic has promised to spend about $900 billion on compute. Much of that money goes to the same companies that own its shares, and the prospectus says plainly that paying it depends on revenue continuing to grow.
The run-rates were real
Until this month, everyone outside Anthropic was working from run-rates: a recent month of revenue multiplied up to a year. They are unaudited, and the labs calculated them differently. OpenAI's chief revenue officer told staff in April that Anthropic's run-rate was inflated by about $8 billion, because it booked sales through Amazon and Google at their full price.
The prospectus settles most of that argument. Anthropic's revenue was $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025. In the first nine months of 2026 it was $33.2 billion, thirteen times the same period a year earlier. Revenue for the quarter to September was $17.0 billion, against $1.4 billion a year before.
Multiply that quarter by four and you get $67.8 billion. The run-rate at the end of September was about $71 billion, and the prospectus says it reached $81 billion by 31 October. The two methods give much the same answer, so the run-rates the press reported all year were a fair guide. The accounting choice is still there: 44% of revenue this year came through cloud marketplaces such as AWS Bedrock and Google's Vertex, and Anthropic counts those sales gross. But it doesn't change the size of the business much.
What drives the growth is a small number of large customers spending more every quarter. Net revenue retention, which compares what the same customers paid this quarter with what they paid a year earlier, is about 900%. A company whose customers merely keep spending the same scores 100%. Most of the money is metered usage (79% of revenue this year), billed per token by companies building products on Claude, with subscriptions making up the rest.
Two details here cut both ways. Concentration is falling: 79% of third-quarter revenue came from outside the top five customers, up from 64% in the first quarter, and no single customer is above 10%. That means in the first quarter of 2026, five customers paid more than a third of the bill. The other detail is how much the revenue depends on the newest models. Models released in the previous six months accounted for more than 80% of third-quarter revenue. Anthropic's lead is real, but it has to be won again every few months, against open models that trail the frontier by about four months.
The prospectus also says, in as many words, that this can't last: "We expect our revenue growth rate to decline over time, and it may decline significantly and more rapidly than we anticipate." Every prospectus says something like this. It matters more here because of what the company has committed to spend.
A business that already makes cash
AI labs are usually described as money pits. The filing shows something more interesting: a lab whose costs fell behind its revenue this year, and which has since chosen to spend ahead of it again.
Anthropic doesn't report a gross margin, because its income statement groups costs by type rather than separating the cost of serving customers. The largest line is compute and infrastructure, which covers training, research and serving customers together. In the first quarter of 2025 that line was 2.5 times revenue. By the second quarter of 2026 it was 48% of revenue.
Revenue passed the compute bill in the first quarter of 2026, and in the second quarter Anthropic made an operating profit of $974 million. The third quarter went back to a loss of $567 million. The company gives two reasons. It bought $5.0 billion of short-term compute, which "currently costs significantly more" than capacity under its multi-year contracts, because demand ran ahead of the capacity it had already contracted. And it took a $2.0 billion non-cash charge for matching employees' charitable donations in stock.
The timing explains the pattern. Anthropic says it contracts compute 12 to 18 months before it arrives, and the cost starts when the capacity comes online, "often in large, discrete steps." Revenue that the new capacity will support comes later. So margins jump when revenue catches up and fall when a big block of compute lands. The company warns this will keep happening "at times meaningfully, even if the underlying economics of the business continue to improve." Price matters too: price cuts on new model versions took about $600 million off revenue in the third quarter.
Anthropic's adjusted operating income, which adds back stock compensation, the donation charges and the $1.5 billion it paid to settle the Bartz copyright case with authors, was $2.2 billion in the second quarter and $2.5 billion in the third. That's a 15% adjusted margin in the third quarter and 12% for the nine months. More concretely, operating cash flow was positive at $6.6 billion for the nine months. After about $2.9 billion of equipment purchases and prepayments for leased hardware, that leaves roughly $3.7 billion of free cash flow (my calculation).
The headline loss is a different story. The nine-month net loss was $104 billion, which looks alarming but is almost entirely paper. Anthropic had raised money through convertible notes, debt that turns into shares at a price set in advance. Each time its valuation jumped, those notes became more valuable to the holders, and accounting rules book that gain as a loss for the company: $94.3 billion in nine months, $59.6 billion of it in the quarter of the $965 billion Series H. Most of the notes convert at the IPO, so most of that charge stops. Another one starts. When the offering closed, a large block of employee RSUs vested, and Anthropic will book about $14.7 billion of stock compensation at once in the fourth quarter. Expect the first quarterly report as a public company to show a large GAAP loss for reasons that have nothing to do with the business.
The compute bill
The number that dominates the prospectus is $900.9 billion. That's what Anthropic has committed to spend, "substantially all" on compute and infrastructure, mostly between now and 2045.
The total has more parts than it first appears. Only $320 billion sits in the standard table of contractual obligations as of 30 September. The largest piece, $451.2 billion, is leases for data centres and equipment that have been signed but haven't started yet, most of which begin between 2026 and 2028 and run for around ten years. Another $105.7 billion was signed after September, and $24 billion is short-term contracts. On top of all of that are contracts with no fixed price, mainly extra Google TPU capacity and AMD's Helios racks, which the company estimates at another $150 to $175 billion.
The timing is what matters for cash. About $70.7 billion falls due in the twelve months after September, and $451.6 billion is expected to be incurred before the end of 2030.
This changes a number I published in October. In the money loop piece I put Anthropic's disclosed commitments at about $225 billion, or 3.5 times its run-rate. That was all that was public. Against the full $900.9 billion and an $81 billion run-rate, the ratio is about eleven times. That's much closer to OpenAI's 14.6 times than to the gap I described then. Anthropic still has the stronger revenue base, and its revenue is counted on a comparable basis now that it's filed. But both leading labs have committed to spend more than ten years of current revenue, and my earlier framing understated that for Anthropic.
Who is on the other side
The prospectus names the largest contracts, and the names will be familiar.
- Broadcom leases Anthropic Google-designed TPU hardware for about $161.2 billion in rent over five-year terms, installed in data centres that others run for Anthropic. These leases can't be cancelled, and Anthropic has put cash into a restricted account for Broadcom's benefit.
- Google has a commitment of at least $121.3 billion for cloud hosting to July 2033. If Anthropic spends less, it pays Google the difference. Separately, it has agreed to take more TPU capacity from 2028 to 2031, which it estimates at $125 to $150 billion more. Google owns 13.4% of Anthropic.
- Amazon has at least $110 billion of take-or-pay commitments to April 2036. Amazon owns 14.6%, or 23.6% if its remaining convertible notes were converted without the ownership cap.
- SpaceXAI, the former xAI and now part of SpaceX, leases Anthropic its Nvidia-based Colossus capacity for up to $120.5 billion into 2029, about $3.1 billion a month once fully delivered in December. In October I described this deal at $1.25 billion a month. It has since grown to two and a half times that. It is also the most flexible contract in the book: after initial periods end in January and March 2027, either side can cancel on 90 days' notice.
- Microsoft has at least $47.2 billion to 2033. Nscale (up to $44.6 billion) and Lambda ($41.3 billion) are neoclouds, smaller companies that borrow to buy Nvidia GPUs and rent them out. Anthropic pays them only as capacity is delivered. Nscale still has to find financing. Lambda missed a deadline to secure a data-centre lease, so its largest contract, worth $34.9 billion, can now be terminated by either side. Neither has done so.
- AMD has an agreement for Helios rack systems in 2027 with no fixed amount, expected to exceed $20 billion.
Two things stand out. The first is how much of the bill could be walked away from: $186.2 billion, about a fifth of the total, depends on the other side securing financing, facilities or approvals, or carries a termination right that can already be used. That gives Anthropic room to move if demand disappoints, and it means some of the headline will probably never be spent.
The second is the loop. Amazon and Google, Anthropic's two largest compute suppliers, together own about 28% of the company. Broadcom, Microsoft, AMD, Nvidia, Micron and TSMC bought shares in the private placement at the IPO price. Their money arrives as equity and goes back to them as payment for chips, memory, cloud capacity and leases. I described this pattern in October as Anthropic's tight loop, "compute contracts paired almost one for one with equity from the same counterparty." The prospectus confirms it and extends it to almost every link in the supply chain: Micron makes the high-bandwidth memory I wrote about in memory's pizza economics, and TSMC makes nearly every chip involved.
None of this is improper, and it isn't necessarily a problem. Suppliers investing in their largest customer is common, and here the customer's revenue is real. But it means the companies best placed to judge Anthropic's demand are also the ones paid when it buys more compute. And their stakes are Class N shares, which carry no vote.
Can it pay?
After the offering, Anthropic has about $232 billion of cash and short-term investments, plus a $14.8 billion credit line and up to about $67 billion more that its strategic partners have agreed to lend through convertible notes once the private placement closes. That covers the next twelve months of commitments three times over, and about half of what is due before 2031. The rest has to come from revenue.
The prospectus is blunt about what happens if it doesn't: "Our ability to meet the substantial and growing financial obligations of these requirements to secure compute depends on our ability to generate significant revenue growth." It points out that most revenue is metered usage with no long-term contract, while the compute is committed for years. Missing a payment could set off cross-defaults under its credit agreement and some of its notes, which the filing describes as "a cascading series of adverse consequences."
That is the standard language of a risk factor, and I wouldn't read it as a forecast. But it describes the structure accurately. Customers can stop spending at any time; Anthropic can't stop paying for most of its compute.
How management explains it
Anthropic's executives don't hide from the compute question. They put it at the centre of the pitch.
In the prospectus, the company describes four drivers: adoption of Claude, compute efficiency, operating efficiency and compute investment. It says they "materialize on different timelines." Its evidence for efficiency is that revenue grew 1,127% year on year in the third quarter while compute expense grew 382%. Revenue per employee went from $1.4 million to $6.6 million over the year. More than 90% of the code contributed to its codebase in September was written by Claude. The filing uses the phrase "closing the loop" for Claude helping to build its own successors. That's the same recursive self-improvement I examined in an earlier essay, now presented to investors as a cost advantage. The filing also lists the products behind the growth, including Claude Cowork, whose office plugins set off the SaaSpocalypse in February, and Claude Security.
Much of the new spending goes into what it calls Anthropic Direct: leasing chips and housing them in dedicated data centres instead of renting capacity from a cloud provider. Its argument is that data centres take the longest to secure but are a small share of the cost. Signing for the buildings early and committing to the accelerators inside them later is a more flexible use of capital. Over time, cheaper direct capacity is supposed to replace the expensive short-term compute that hurt the third quarter.
On the roadshow, chief financial officer Krishna Rao said the company believes it can reach an adjusted operating margin above 35% in the long run, with compute falling to roughly 45 to 50% of revenue. He gave no date. The third quarter was 15%.
Dario Amodei, the chief executive, was more direct about the risk. Talking with JPMorgan's Jamie Dimon during the roadshow, he called compute "the deepest business dilemma that the company faces" (Dealroom). He said revenue had grown roughly tenfold in each of 2024 and 2025, but compute has to be bought one to two years ahead. Buy too little and you turn customers away; buy too much and you're paying for idle chips. He has told this story before: Anthropic planned for tenfold growth and got eightyfold, and had to ration usage until the Colossus capacity arrived. Asked why go public, he said the main reason "goes back to these compute bets": a cash cushion against getting them wrong. He added that the company should never be worth "a dollar more than it's really worth."
There is a tension here that I'd keep in mind. In September, Amodei asked governments to slow the frontier, two months before an offering whose arithmetic needs revenue to keep compounding. The prospectus shows why a slowdown would suit a company in Anthropic's position. A slower race means less pressure to sign ever-larger compute contracts years ahead of demand, and a four-month lead lasts longer. That doesn't make the call insincere, but it isn't free of self-interest either.
Who controls it
Public shareholders own the economics but have very little control. Anthropic is a public benefit corporation, legally required to weigh its mission alongside profit. A single Class F share held by the seven co-founders carries 50.1% of the votes on ordinary matters. The Long-Term Benefit Trust, an independent body of trustees, elects a majority of the board through Class T stock. Amazon's and Google's Class N shares carry no vote. It's a "controlled company" under Nasdaq rules, so it doesn't need a majority of independent directors. Buyers of Class A shares are signing up to back a mission they can't steer.
What you have to believe
At Friday's close, Anthropic is valued at about $1.2 trillion on shares outstanding, or about $1.6 trillion fully diluted. The analysts who published around the listing all work at banks that weren't underwriters. Citizens expects about $52.5 billion of revenue this year and $137 billion in 2027 (Invezz). Needham's $150 target assumes growth of 156% next year and 98% in 2028. Melius projects about $430 billion of revenue by 2029 and values the company at six times that. On the 2027 estimate, the stock trades at about nine times revenue on shares outstanding and about eleven times fully diluted. That's expensive for most companies, and cheap for one that really grows 2.6 times next year.
The more useful test is the compute bill. Here is a simple one using management's own target.
Commitments due before 2031 work out to about $109 billion a year from 2027 to 2030 (my estimate: $451.6 billion less roughly $15 billion in the last quarter of 2026, spread over four years). If compute is to be half of revenue, as Rao suggests, Anthropic needs about $218 billion of revenue a year over that period. Revenue in the year to September was about $35 billion.
On the analysts' path, Anthropic passes that line in 2028, and the bill looks comfortable. Now suppose growth halves every year from 2027: 100%, then 50%, 25% and 12.5%. That's a sharp slowdown, but it still means revenue doubles next year. In that case revenue is about $105 billion in 2027, roughly equal to that year's share of the compute bill, and doesn't reach $218 billion until 2030. The company wouldn't collapse. It would burn cash, lean on its $232 billion of cash and the roughly $67 billion its partners have agreed to lend, cancel what it can (the SpaceXAI and Lambda contracts first), and margins would stay far from 35%.
So the question isn't whether Anthropic survives a slowdown; with this much cash it survives a severe one. The question is what it's worth if growth falls to merely very fast, and the price is set for the analysts' path rather than the halving one.
The bull and bear cases
The strongest bull arguments:
- The revenue is audited and matches the run-rates. The main uncertainty in every earlier analysis, including mine, is gone.
- It already generates cash. Operating cash flow was positive for nine months, and the second quarter showed the business can be profitable when revenue catches up with compute.
- Demand still exceeds supply. Anthropic paid premium prices for short-term compute because it had customers to serve. Net revenue retention of 900% and 11,900 large customers point the same way.
- The balance sheet is built for mistakes. $232 billion of cash, about a fifth of the commitments cancellable or contingent, and partners who have agreed to lend more.
The strongest bear arguments:
- The bill is fixed and the revenue isn't. About $900 billion of commitments, much of it take-or-pay, against revenue that customers can cut at any time.
- Prices are falling. Price cuts took $600 million off revenue in a single quarter, and corporate spending on AI tokens stopped growing this summer even as usage rose.
- The lead has to be won again every few months. More than 80% of revenue comes from models less than six months old, and open models are about four months behind.
- The circle is tight. The suppliers most exposed to Anthropic's demand own much of it and are paid by it, and public shareholders have almost no say.
| Evidence for | Evidence against | My lean | |
|---|---|---|---|
| Revenue quality | Audited, matches run-rates, 11,900 large customers | 44% through cloud marketplaces, top five ~25% | Bull |
| Unit economics | Compute down from 251% to 48–58% of revenue; positive cash flow | Margin swings with each block of compute; price cuts | Lean bull |
| Compute bill | $232B cash, $186B contingent or cancellable | ~11× run-rate, take-or-pay, cross-defaults | Lean bear |
| Valuation | ~9–11× 2027 revenue if analysts are right | Priced for growth well above a halving path | Neutral |
| Governance | Mission protected by trust and founders | Public holders can't steer; suppliers hold non-voting stakes | Neutral |
On balance, the business is better than I expected and the bill is bigger than I expected. At today's price, the market is paying for the first and mostly looking past the second.
What to watch
- The first quarterly report as a public company, likely in February. Expect a large GAAP loss from the $14.7 billion stock-compensation charge. The figures that matter are revenue growth, compute as a share of revenue and how much short-term compute it bought.
- The underwriters' analysts, who can start publishing 25 days after the offering, in mid-December. Goldman Sachs and Morgan Stanley led the deal.
- The SpaceXAI notice windows in January and March 2027, and whether either side cancels the Lambda contract. Cancellations would suggest Anthropic thinks it has enough capacity.
- Price per token. Another $600 million quarter of price cuts would be a signal; a smaller one would suggest pricing power is holding.
- The lock-up, which ends 180 days after the listing, around mid-May 2027. Employees and early investors can sell from then. The suppliers' private-placement shares aren't under the lock-up but can't be resold without registration.
- OpenAI's prospectus, the first chance to compare the two leading labs on audited numbers.
Sources
Primary
- SEC, Anthropic, PBC final prospectus (424B4), 18 November 2026
- SEC, Anthropic, PBC Form S-1, 26 October 2026, and amendments of 2, 10 and 17 November 2026
Trading and valuation
- CNBC, Anthropic's record IPO collides with tech's worst week on Wall Street since 2020, 21 November 2026
- CNN, Anthropic valued at $1.76 trillion after historic first day of trading, 19 November 2026
- Fortune, Anthropic stock pops 30% to $2 trillion market cap in mega IPO, 19 November 2026
- TechCrunch, Anthropic raises $65 billion, nears $1 trillion valuation ahead of IPO, 28 May 2026
Management commentary
- Dealroom, Amodei tells Dimon Anthropic is going public for a compute cash buffer, November 2026
- FourWeekMBA, Anthropic's IPO video sees 35%+ adjusted margins; Q3 was 15%, November 2026
- Constellation Research, Anthropic's IPO roadshow video takeaways, November 2026
- Fortune, Anthropic's IPO filing shows tricky financial mechanics, 4 November 2026
- Euronews, Anthropic swings back to operating loss as compute bill looms over IPO, 3 November 2026
Analysts
- Yahoo Finance, Needham initiates coverage of Anthropic with a Buy rating, November 2026
- Yahoo Finance, Could Anthropic actually climb 63%? Or is Melius totally wrong?, November 2026
- KuCoin News, Analyst compares Anthropic to Amazon and Microsoft, sets $150 price target, November 2026
- Invezz, Anthropic stock forecast: price targets, 20 November 2026
Earlier essays on this site
- The AI money loop: who funds whom, and where it's cracking
- Can AI ever build itself? A deeper look into autonomous AI research
- The dawn of the agentic era: who gets paid when software starts acting
- The AI supply chain adventure: How minerals become LLMs
- Memory's Pizza economics
- The SaaSpocalypse, sorted: who AI is actually disrupting