Deal analysis

Anthropic's S-1: The $80bn Loss Isn't the Cash Problem. The Calendar Is.

Anthropic's headline loss is mostly non-cash. Its public S-1 reveals a bigger lesson for founders: long-term compute commitments against revenue customers can cut quickly.

By · 27 October 2026 · 13 min read

A rising glass staircase supported by heavy server racks, illustrating growth and long-term infrastructure commitments

On Monday 26 October, Anthropic publicly filed its registration statement with the SEC (Form S-1 on EDGAR). It intends to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The company's own announcement is clear that the number of shares and the price range have not been set, and that the registration statement is not yet effective. So nobody can yet tell you what Anthropic will be worth on listing day.

Our earlier piece on the leaked draft looked at customer concentration, channel dependency and adjusted profit measures. The public filing confirms most of that, explains the gap between operating and net loss that the leak coverage left open, and surfaces a bigger issue.

That issue is not the size of the loss. It is the calendar. Anthropic has signed up to costs measured in years, and in some cases a decade, while earning revenue that, by its own description, many large customers can reduce at any time. A smaller version of that mismatch sits inside many owner-managed businesses, and buyers look for it.

All references below are to the S-1 unless another source is linked.

First, most of the headline loss is a non-cash charge

The number that went round on Monday was an $80.7bn net loss for the six months to 30 June 2026. It is real, and it is not where the cash pressure sits.

In those six months, revenue was $16.3bn, up from $1.2bn in the same period of 2025, roughly fourteen times higher (Summary Consolidated Financial Data). The GAAP operating loss was $693m. Most of the gap between that and the $80.7bn net loss is a single line: a $72.4bn charge for the change in fair value of convertible notes, derivative liabilities and SAFEs, plus a further $8.8bn of other expense linked to note purchase agreements.

The filing says those charges were "primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation." In plain English: Anthropic owes instruments that convert into its own shares, so as the company becomes more valuable, the accounting records a loss.

That is not meaningless. The charge reflects value passing to noteholders, diluting everyone else. It just does not consume cash.

The same pattern explains 2025. The full-year net loss was $42.0bn, of which $34.3bn was this fair value movement. Of the $132.9bn accumulated deficit at 30 June 2026, the S-1 says $113.2bn relates to these fair value changes since 2023.

The operating picture is also turning. In the second quarter of 2026, revenue was $11.6bn and Anthropic reported a GAAP operating profit of $974m, the first quarterly operating profit in the periods shown (Quarterly Results of Operations). Cash generated from operations in the first half was $3.9bn, against cash consumed in every prior full year disclosed.

One limit matters. Q2 2026 is the latest quarter for which this preliminary filing gives revenue and operating figures, not the latest calendar quarter. The filing cites some third-quarter metrics, such as customer mix, but leaves the Q3 dollar figures blank.

Adjusted figures need the same care. For the first half the company reports adjusted operating income of $1.618bn against the GAAP operating loss of $693m. The difference is mainly $2.185bn of stock-based compensation and $126m of non-cash charitable contributions. Share-based pay is not free: existing shareholders pay for it in dilution rather than cash. Unlike the gross margin calculation below, both operating measures do include model training compute and channel partner costs.

None of that makes the business cheap or safe. The $80.7bn is simply the wrong number to use as a measure of cash risk.

The owner's lesson: losses that come from how a financing instrument is accounted for are not losses from trading. If your accounts carry anything similar, such as shareholder loans with conversion rights or deferred consideration, be ready to explain it in a page, including what it means for ownership. A buyer who has to work it out alone will assume the worst first.

The missing gross profit line

In Note 1 to the financial statements there is a smaller detail. During 2025, Anthropic changed how it presents its operating expenses. It moved from a functional layout, which showed cost of revenue and gross profit as separate lines, to a layout grouped by the nature of the expense: compute and infrastructure, personnel, sales and partnerships, and so on (Note 1, Change in Presentation).

As a result, there is no gross profit line on the face of the income statement.

The functional numbers have not vanished. Note 8 still discloses them. Taking that table, cost of revenue was $2.4bn in the first half of 2026 against revenue of $16.3bn. On my calculation, not a figure the company presents, that implies a gross margin of about 85%. The same calculation gives roughly 40% for full-year 2025 and about 12% for the first half of 2025.

The detail that matters is what sits outside cost of revenue. Compute used for model development and research sits in research and development, which ran at $9.7bn in the first half. Channel partner costs sit in sales and marketing. So the 85% is a margin before the cost of building the next model and before the fees paid to the cloud platforms that distribute nearly half the product.

The company argues that serving customers and training models are so intertwined that splitting them is artificial. That is defensible. But any buyer, of an AI lab or a £5m engineering firm, will rebuild your profit and loss account in the shape that suits their analysis.

The owner's lesson: if you change how your accounts are presented in the years before a sale, explain it before anyone asks. An unexplained change in format reads as concealment even when it is not.

The real story: long-dated costs, short-dated revenue

Here is the part of the filing I would spend the most time on.

The Contractual Obligations section of the MD&A lists Anthropic's major compute arrangements. Some examples:

  • Google: at least about $114.5bn between April 2026 and July 2033, "regardless of actual usage", with a true-up payment if spend falls short. Separately, an estimated $125bn to $150bn for additional TPU compute between 2028 and 2031.
  • Amazon: at least $110bn between May 2026 and April 2036, again with a true-up.
  • Microsoft: at least $44.9bn between November 2026 and May 2033, regardless of actual usage.
  • Broadcom equipment leases: around $161.2bn of rent, excluding interest, over five-year, non-cancellable terms.
  • SpaceXAI: up to $120.5bn into 2029, running at roughly $3.1bn a month once all capacity is delivered.
  • Nscale and Lambda: up to about $44.6bn and $41.3bn respectively, into 2033 and 2034.
  • AMD: no fixed dollar commitment, but the company expects spending to exceed $20bn if it procures the full initial committed capacity, subject to conditions.

On my arithmetic, the arrangements with a stated figure add up to more than $600bn, before the estimated TPU spend and AMD. The filing's own total is blank in this preliminary version, so treat my sum as an indication of scale. The bases differ: minimums, caps, conditional and cancellable deals. None of it is debt in the balance-sheet sense, and it falls due over each contract's term, not at once.

Now set those against the revenue side, as the company itself describes it:

  • Revenue is recognised as customers consume tokens. Subscriptions are "generally 12 months or less."
  • The top five customers made up about 25% of revenue for the nine months to 30 September, and "many of our top customers are not subject to long-term contracts with us or have contracts with near-term expiration dates and are able to reduce or in some cases cease spending at any time and for any reason" (Risk Factors).
  • Model versions released in the previous six months accounted for over 80% of third-quarter revenue.
  • Customer spend fell about 20% over the two weeks to 2 January 2026, compared with the previous fortnight, because of holiday code freezes.
  • The company estimates that 5% to 7% of revenue in the first nine months of 2026 came from customers it later found to be using the service in breach of its usage policy or commercial terms, and it expects that share to fall after tighter enforcement in August and September.

Put those two lists side by side and the shape is clear. Much of the cost base is long-dated and, in several cases, payable whether or not it is used. The revenue base is short-dated and usage-driven. The 80% figure does not prove that each model has a fixed six-month life. It does show how quickly usage moves to the newest release, which means Anthropic has to keep shipping models good enough to win that usage, at a pace it does not fully control.

The SpaceXAI deal makes the point with one number. At $3.1bn a month once fully delivered, it would cost about $9.3bn a quarter, around 80% of second-quarter revenue on my calculation. Yet it is one of the more flexible contracts: either side can cancel on 90 days' notice once the initial periods end in early 2027. The more rigid commitments are the cloud take-or-pay deals running into the 2030s.

To be fair to Anthropic, it hides none of this, and the logic is coherent. Compute has long lead times. Revenue grew 1,368% year on year in the second quarter while compute and infrastructure expense grew 300%, according to the prospectus summary. At 30 June 2026 it had $131.7bn of cash, cash equivalents and short-term investments, and it says its resources are enough for at least the next twelve months. The company also states plainly that "our net cash used in these activities may increase regardless of whether our financial performance improves."

That sentence is the honest heart of the filing. The commitments are a bet on demand years ahead, financed by money raised now.

Why this matters far beyond AI

In mid-market deals the same mismatch looks far less dramatic, which is why owners miss it. It looks like a fifteen-year lease taken for one big contract that renews every year, a minimum purchase agreement priced for last year's volumes while customers buy on purchase orders, or a personal guarantee that outlives the revenue it was meant to support.

Buyers do not just ask what you earn. They ask what you are committed to spend if revenue falls by a third, and how quickly you could get out. If the honest answer is "not quickly", the risk lands somewhere in the deal: a lower multiple, more of the price deferred into an earnout, or a request that you stay on the hook for a lease after completion.

The answer is rarely to avoid long commitments; sometimes they are what make the business work. It is to make the duration of your costs visible, match it to your revenue where you can, and know what happens to each commitment if demand falls.

Control: what the public buyer will be getting

The governance section describes a structure any owner selling a minority stake should recognise. Parts of it are proposed for the IPO rather than in place today.

Anthropic is a Delaware public benefit corporation. Its Long-Term Benefit Trust, made up of five financially disinterested trustees, holds a special Class T share that lets it elect and remove a majority of the board. Under the proposed structure, immediately before the offering a single Class F share will be issued to an LLC owned by the seven co-founders, carrying 50.1% of total voting power on general matters, including the election of the directors the Trust does not appoint (Prospectus Summary). Neither share carries economic rights. Class A shares carry one vote each, and strategic partners hold largely non-voting Class N stock.

The filing's own risk factor puts it bluntly: the structure "may effectively eliminate the ability of holders of our Class A common stock to influence important decisions, including a change in control." The company expects to be a "controlled company" under Nasdaq rules.

Whatever you think of the mission behind it, the commercial effect is simple: public investors will be buying economics with very limited control. How much, if anything, the market charges for that will only become visible once the shares are priced and trading.

The owner's lesson: if you sell a minority while keeping control, expect the buyer to reflect that in the price. If you sell control while keeping a stake, look hard at its protections. Voting rights, information rights, drag and tag provisions and reserved matters are where minority value is protected or lost.

What the filing confirms from the leak, and what it changes

The concentration points from the leak hold up. The notes to the financial statements show two customers at 12% each of 2025 revenue. A customer accounted for 11% in the first half of 2026. Sales through cloud marketplaces made up 48% of revenue in the first half and 58% of net receivables at 30 June 2026.

The company also points to improvement. It says around 79% of third-quarter revenue came from customers outside its top five, up from about 64% in the first quarter. It reports net revenue retention of around 900% at 30 September 2026.

That 900% needs its definition. The S-1 takes customers with revenue in the three months ending a year earlier, excludes Claude Pro and Max subscribers, and compares those same customers' revenue in the latest three months. It measures how much one cohort expanded over a year, not a retention rate for every customer and not a measure of profit. CNBC cited a median of 108% for public cloud software companies, based on Altimeter Capital data, but definitions vary, so treat that as context rather than a like-for-like comparison.

What changes is the emphasis. On the leaked numbers it was reasonable to ask whether the business could make money. On the public numbers it made an operating profit in Q2 2026. The open question is whether demand will be there, in the right volumes, when the bills for years of compute arrive.

Four questions to ask of your own business

For any owner thinking about an exit in the next few years:

  1. What are you committed to spend, for how long, regardless of revenue? List every lease, minimum purchase, finance agreement and guarantee, with its end date.
  2. How long is your revenue committed for, and how reliably does the rest recur? Separate contracted revenue from revenue that simply tends to recur. Both have value. Buyers will weigh contractual protection differently from a demonstrated track record of repeat business, so be ready to evidence each.
  3. What would your profit look like if revenue dropped by a third for a year? If the answer is a large loss you could not exit, a buyer will find it and price it.
  4. Can you explain every unusual line in your accounts in a page? Financing charges, one-off settlements, presentation changes. Clarity on these protects your credibility on everything else.

Anthropic can carry a very large duration mismatch towards the public markets because it has unusual growth, unusual scarcity and, at 30 June 2026, $131.7bn of cash, cash equivalents and short-term investments. Most businesses have none of those cushions, and the sale process is not the time to discover how exposed you are.


Figures are taken from Anthropic's preliminary prospectus filed on 26 October 2026, which is subject to completion and change. Interim figures for 2025 and 2026 are unaudited. Gross margin percentages, the SpaceXAI revenue comparison and the aggregate of named compute commitments are the author's own calculations from the filing. This article is not investment advice.

Free tool

How buyer-ready is your business?

Take the free 25-question Exit Readiness Scorecard to see your strengths and preparation gaps. Enter your email to access the tool, then choose where to focus first.

Take the free readiness scorecard