Anthropic's filing shows an $80.7 billion loss and a $1.6 billion profit. Same six months.

Introduction

Anthropic, the company that makes Claude, has asked permission to sell shares to the public. If it goes ahead, anyone with a share-dealing account will be able to own a piece of it for the first time. The document it filed to do that runs to fourteen million characters and reports, for the first six months of this year, both the largest loss we have ever seen in a listing document and — on a slightly different measure of the same six months — a profit. Neither number is a trick. They are measuring different things, and knowing which is which is most of what there is to understand here. One thing the document does not say is what the shares will cost.

Start here: the words you need

An S-1 is the document a company must file with the Securities and Exchange Commission before it can sell shares to the public in America. It is written by the company and its lawyers, it is a selling document, and it is also the most heavily policed thing the company will ever publish: lying in it is a federal offence. That combination is why it is worth reading.

An initial public offering, or IPO, is the sale itself. Anthropic will print new shares and sell them; some existing owners will sell shares they already hold. Afterwards the shares trade on an exchange and anybody can buy them from whoever wants to sell.

The underwriters are the investment banks that run the sale — here Goldman Sachs and Morgan Stanley are the two named as the places to get a prospectus. They canvass large investors, find out what those investors would pay, and set a price. That process is called building a book.

The price range is the band the banks expect to price within, printed on the prospectus cover a couple of weeks before the sale. It is not in this filing. Nor is the number of shares. Both are left as literal blanks, which is normal at this stage and is also the single most important fact in this article.

A lock-up stops insiders selling for a set period after the listing — 180 days here. When it lifts, a lot of shares can become available at once.

Share classes let a company give different owners different rights. The public is being sold Class A. There are four other classes, and one of them matters enormously; we come to it below.

A public benefit corporation is a company whose directors are legally permitted — and in Anthropic’s case directed — to weigh a stated public mission alongside shareholder returns. It means a board can lawfully choose something other than the most profitable option.

What the trend looks like

Steeply, almost absurdly, up. Six quarters ago Anthropic was booking $376m of revenue in a three-month period. In the three months to 30 June 2026 it booked $11,555m — about thirty-one times as much. Revenue in the June quarter was 14.7 times what it had been in the same quarter a year earlier. The company states that growth rate as 1,368%, and our arithmetic on its own quarterly table agrees.

Growth like that is not normal and the filing does not pretend it is. It says plainly that results for the half year “are not necessarily indicative” of the full year, and that it expects its growth rate to fall.

The crossover that made the operating loss almost disappear

The biggest cost of running an AI company is compute: the chips and data centres the models are trained and served on. For most of Anthropic’s life that one line cost more than the entire company earned. In the first quarter of 2025 compute and infrastructure cost $944m against $376m of revenue — two and a half times what came in.

In the first quarter of 2026 the two lines crossed. Revenue was $4,731m and compute was $3,387m, so compute fell to 72% of revenue. By the June quarter it was 48%. Over the full six months compute was 55% of revenue, against 201% in the same period a year before. Revenue grew about fifteen-fold year on year; compute spending grew about four-fold.

That gap is the whole reason the operating loss collapsed from $2,135m in the first half of 2025 to $693m in the first half of 2026, on revenue fourteen times larger.

Three answers to one question

Here is where a reader gets misled, and it is worth being slow about. For the six months to 30 June 2026, the filing reports a net loss of $80,748m. That is eighty billion dollars and change, against $16,286m of revenue. It is a larger loss than any company has reported in a listing document, and taken at face value it would suggest a business burning money at a catastrophic rate.

It is not doing that. The cash flow statement shows operating activities generating $3.9bn over the same six months.

The explanation is one line of the income statement: a $72,388m charge for the “change in fair value of convertible notes, derivative liabilities, and SAFEs”. Anthropic funded itself partly with convertible promissory notes — loans that turn into shares later. Accounting rules require it to re-measure what those notes are worth at every reporting date and run the change through profit. The filing’s own footnote says these changes were “primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation”.

Read that twice. The loss got bigger because the company got more valuable. No cash left the building. And at the listing those notes convert into shares, after which the charge stops appearing. It is 89.6% of the reported net loss; add the other non-operating line and the two together come to $81,145m, which is more than the entire net loss.

The net loss and the operating loss differ by a factor of 117. Our reading is that the $693m operating loss is the one that describes the business, and the $974m operating profit in the June quarter is the more useful fact still, because it is the most recent.

But do not accept the adjusted figures uncritically either. The gap between the $693m operating loss and the $1,618m of adjusted operating income is almost entirely $2,200m of stock-based compensation — shares handed to staff instead of cash. That is 13.5% of revenue. It is a real cost to you as a shareholder, because it creates more shares and your slice gets smaller. It just is not a cash cost to the company. Companies prefer the measure that excludes it. You do not have to.

What the public would actually be buying

A share normally bundles two things: a claim on the profits, and a vote on who runs the place. Anthropic is proposing to sell the first and withhold most of the second.

Immediately before the offering the company will issue one single share of a new Class F stock to a limited liability company owned by its seven co-founders, including Dario Amodei, the chief executive, and Daniela Amodei, the president and chair. That one share is entitled to “an aggregate number of votes that equals 50.1% of the total voting power”. It carries no economic rights at all — no dividends, no claim on assets.

The arithmetic is absolute rather than approximate. Because the Class F share takes a fixed 50.1%, every Class A holder in the world put together is limited to 49.9%. Buying more shares cannot change that. The filing states the consequence directly: the voting power of Class A “is limited to an aggregate of 49.9% of the total outstanding voting power”. Anthropic also confirms it “is considered a ‘controlled company’” under Nasdaq’s rules, an exemption that lets it skip some board independence requirements.

Whether this is good or bad depends on what you think you are buying. Founder control has protected long-horizon strategies at other technology companies and it has entrenched bad management at others. What is not in doubt is that a Class A holder who dislikes a decision has no mechanism to overturn it. That is a permanent feature, not a transitional one.

The number that would worry us, if we were worried

Not the loss. The commitments. Anthropic has signed multi-year contracts for compute that oblige it to pay whether or not it needs the capacity. The three biggest use the phrase regardless of actual usage:

• Google, at least about $114.5bn between April 2026 and July 2033.
• Amazon, at least $110.0bn between May 2026 and April 2036.
• Microsoft, at least $44.9bn between November 2026 and May 2033.

That is $269.4bn of unconditional obligation. Against trailing twelve-month revenue of $19,708m, it is 13.7 times a year’s sales. Against the $131,651m of cash and short-term investments on the balance sheet, it is slightly more than double. Both the Google and Amazon contracts require a true-up payment if Anthropic spends less than it promised.

Add the agreements with softer or capacity-dependent terms — a further $125bn to $150bn of Google TPU capacity for 2028 to 2031, up to $120.5bn with SpaceXAI, up to $44.6bn with Nscale, more than $20bn with AMD — and the named total runs to between $579.5bn and $604.5bn, with payments extending to 2045.

We should be precise about what that figure is: it is our sum of the individually named agreements. The filing’s own aggregate commitment total is one of the blanks, left as “more than $ ” pending completion. So treat $269.4bn as the firm floor and the $600bn as an order of magnitude.

There is a second concentration worth naming. Nearly half of Anthropic’s revenue — 48% in the first half of 2026, up from 11% in 2023 — arrives through cloud providers’ marketplaces, which bill the end customer and remit the money onward. Those same channels account for 58% of money owed to the company. Several of the firms Anthropic depends on to reach its customers are also the firms it has committed hundreds of billions of dollars to, and in some cases they compete with it.

So can you buy it, and should you?

Two different questions, and only one of them has an answer today.

Can you? Probably not at the offer price. Shares in an IPO are allocated by the underwriting banks, overwhelmingly to institutions and their best clients. Some brokers run retail allocation programmes, and some companies ring-fence a slice for customers, but the realistic expectation for an ordinary account is that you buy on the first or second day of trading, at whatever price the market opens at — which in a hot deal is well above the offer price. That distinction matters more than it sounds, and the historical record below is measured the honest way: from the first traded price, not the offer price.

Should you? Unanswerable from this document, because the price is missing. This is not a hedge. A company earning $19.7bn over the last twelve months is a different investment at a $200bn valuation (about ten times sales) than at a $700bn valuation (about thirty-five times sales). The filing gives you the numerator of that fraction in considerable detail and leaves the denominator blank. Anyone telling you today whether ANT is cheap is working from a number nobody has published.

What the record can tell you is how this category of offering has behaved. The figures below are from Jay Ritter at the University of Florida, whose IPO dataset is the standard academic reference, updated 5 October 2026. They cover 9,253 American IPOs from 1980 to 2024, with returns measured to December 2025 over three years from the first closing price.

The headline base rate is discouraging: the average IPO beat nobody, trailing the market by 20.5 percentage points over three years across more than nine thousand cases. But the averages hide the thing that matters here. Anthropic falls into the two groups that did least badly — companies with over $1bn of sales before listing trailed the market by only 2.1 points, and companies with more than one share class by 6.7 points rather than 22.0.

The one positive cell in the table is technology companies with multiple share classes, which beat the market by 13.8 points. We would not lean on it. That group contains 309 IPOs, which is small enough that a handful of outcomes move the average, and the period includes the late-1990s technology boom. A result drawn from one window is a property of the window. Note also that large companies pop less on day one — up 9.8% on average against 19.9% for small ones — so the thing most retail buyers are actually chasing is weakest in precisely the category Anthropic belongs to.

What to watch next

Four things will settle the questions this filing leaves open. We are not predicting which way any of them breaks.

• An amended S-1 carrying the price range and share count, usually published a week or two before trading starts. Until it lands, every valuation of ANT is an invention. Divide the implied market value by $19.7bn of trailing revenue and compare it with what listed software and chip companies trade at; that ratio, not the headline price, is the comparison that means something.
• The first quarter reported as a public company. The June quarter’s 8% operating margin is one quarter. The question it raises is whether compute costs keep falling as a share of revenue — they went 251%, 177%, 148%, 144%, 72%, 48% — or whether the $269.4bn of minimum commitments start arriving faster than the revenue does.
• The pro forma columns being filled in. Every one of them is blank in this draft: pro forma loss per share, pro forma cash, the tax bill on settling employee share awards. Those blanks are where the dilution lives, and they are computed from the offer price.
• Day 180 after listing, when the lock-up ends and insiders may sell. The exact date depends on the prospectus date, which is not yet set.

What this means for your money

Nothing has to be decided today, and that is the practical point. There is no price, no date and no share count, so there is no transaction available to anyone reading this. The decision becomes real when the amended filing carries a range.

When it does, three figures from this document are the ones to hold onto. Trailing revenue of $19.7bn is what any valuation should be divided by. The $269.4bn of compute payable regardless of usage is the obligation that does not care how the models sell. And the fixed 50.1% vote is the thing that will not change no matter how the shares perform.

It is also worth being honest about why a reader might want this particular stock, which is rarely the balance sheet. If the appeal is exposure to AI rather than to Anthropic specifically, note that the cost side of this filing is revenue for companies you can already buy — the $269.4bn goes to Microsoft, Google and Amazon — and that a good deal of the listed AI complex has already had its drawdown.

Data appendix

Method. Every Anthropic figure was read from the Form S-1 filed by Anthropic, PBC (SEC Central Index Key 0002130357) on 26 October 2026, accession number 0001628280-26-067168, retrieved from EDGAR and converted from the 14.2-megabyte source document to text locally. The financial statements in this filing carry no XBRL tagging — the only tagged data is the registration fee — so no figure here came from a machine-readable feed and each was transcribed by hand from the statements and reconciled where possible: the quarterly compute lines sum to the half-year totals ($944m + $1,390m = $2,334m, and $3,387m + $5,558m = $8,945m), as do personnel and sales and marketing. Trailing twelve-month revenue is our sum of the four quarters to 30 June 2026. Ratios were computed, not estimated. Compute commitment totals are our own addition of the individually named agreements in “Infrastructure Service Arrangements”, because the filing’s aggregate figure is left blank; where an agreement gives a range we used the lower bound. IPO base rates are from Jay R. Ritter, University of Florida, “Initial Public Offerings: Updated Long-run Statistics”, dated 5 October 2026, Tables 24 and 16a, covering 9,253 IPOs from 1980 to 2024 with returns through December 2025 measured from the first closing market price. Figures are a snapshot of a document that will be amended; nothing here is advice, and we hold no position.