Three Anthropic numbers the IPO headlines get wrong, and how to read them

What changed

Anthropic's preliminary S-1, filed on 26 October 2026, is the most detailed public account yet of what a frontier AI company costs to run. It has also produced a crop of headline numbers that are being repeated without their definitions. Three of them matter to anyone briefing a board or a client on Anthropic's position, and all three are in the filing with enough detail to read correctly. We have set out the full figures, with page references, on our Anthropic IPO tracker; this briefing takes the three most misread and explains what each one measures.

A note on who is writing. The AI Consultancy is Certified in the Claude Partner Network and sells Claude-related services; we take no referral, reseller or deployment income from Anthropic or any vendor (details here). This is business analysis for people who buy and deploy Claude. It is not investment research, and we are not authorised to give investment advice.

Why it matters for UK business

Number one: the $80.7bn loss. The net loss for the six months to 30 June 2026 was $80.7bn. That is the filed figure and it is correct. But the income statement also shows what is inside it: $72.4bn is the change in fair value of convertible notes, derivative liabilities and SAFEs. These are the instruments Anthropic sold to investors over several funding rounds. As the company's valuation rose, accounting rules required those instruments to be remeasured upwards, and the increase runs through the income statement as a loss. It is an entry, not an outflow. The line that measures the business is the operating result: a $693m loss for the half, with $974m of operating income in the second quarter on $11.6bn of revenue. The cash-flow statement is blunter still: operations generated a $3.9bn inflow for the half. So the honest sentence is this. Anthropic's operations approached break-even in the first half of 2026 and turned to income in the second quarter, while the accounting value of what it owes its financiers rose by $72bn. Neither "the company is burning $80bn" nor "the company is profitable" follows from the evidence. The fair-value charge is not irrelevant either: it reflects value transferred to the holders of those instruments and dilution still to come when they convert. Both readings belong in a board paper. Six-month and quarterly figures are unaudited.

Number two: 900% net revenue retention. Anthropic reports net revenue retention of approximately 900% at 30 September 2026, and the figure has been quoted as if it described customer loyalty. The filing defines it. It compares the revenue in the latest three months from an eligible prior-year cohort, meaning customers that generated at least $1 a year earlier, with that cohort's revenue in the corresponding three months a year before. Fixed-price Claude Pro and Max subscriptions are excluded. What it says is that existing eligible customers spent about nine times as much as they did a year earlier. What it does not say is how many customers stayed, what the churn rate is, or what a trailing-year ratio would look like. And the arithmetic guarantees it will fall: the comparison base for next year is the much larger 2026 spend. The filing's other adoption figures need the same care. More than 11,900 customers generated $100,000 or more of trailing-twelve-month revenue at 30 September 2026, up from more than 950 a year earlier; that is the filing's enterprise proxy, with accounts aggregated by organisation, not a count of users or seats. The approximately 40 million customers below that threshold are not a disclosed count of active people either. For a UK buyer the useful inference is narrow: large early adopters expanded their spend very fast. It is not evidence that a mid-sized UK firm's Claude bill should, or will, do the same.

Number three: the compute commitment total. Coverage since the filing has quoted totals for Anthropic's compute and infrastructure obligations ranging from about $450bn to about $518bn, with one widely repeated figure of $498bn described as the non-cancellable share. The filing itself states no total, and its aggregate September commitment fields are blank. What it does disclose is a list of arrangements on different bases: Microsoft, at least $44.9bn from November 2026 to May 2033, usage-independent; Google, at least approximately $114.5bn from April 2026 to July 2033, with a further $125bn to $150bn of TPU capacity for 2028 to 2031 outside the fixed table; Amazon, at least $110.0bn from May 2026 to April 2036; SpaceX.AI, up to $120.5bn into 2029, roughly $3.1bn a month at full capacity, which either party can cancel on 90 days' notice after initial periods ending in January and March 2027; Nscale, up to $44.6bn through 2033, subject to delivery, acceptance and financing; Lambda, approximately $41.3bn through 2034; AMD, estimated above $20bn for initial capacity from 2027, conditional; and Broadcom equipment leases with aggregate rentals of approximately $161.2bn excluding interest. A minimum purchase, an "up to" figure with a cancellation right, a conditional capacity estimate and a schedule of lease rentals are four different kinds of obligation. Adding them produces a number with no meaning, and the range of published totals is itself the evidence: different writers added different rows. The right way to carry this into a board paper is as a table, with each amount beside its basis and its conditions, and a sentence saying that the aggregate is not disclosed. That is how we present it on the tracker, and why there is no total there.

What to do, and what not to do

Do:

• When you quote a loss figure, name the line: operating, adjusted or net, and the period. For the half-year they are $693m, $1.6bn of income, and $80.7bn respectively, and they are three different things.
• When you quote the retention figure, add its definition in the same sentence, and say that it excludes Pro and Max and will fall.
• Present compute arrangements as a table with each one's basis. If a supplier or a journalist gives you a total, ask which rows were added.
• Check the page reference. Every figure above is in the filing's results of operations, non-GAAP reconciliation, cash-flow statement, customer definitions and contractual obligations sections, and our tracker lists them.

Do not:

• Describe the net loss as cash burn, or the adjusted result as profit.
• Describe 900% as customer retention.
• Build a cash runway, a gross margin or a valuation from this preliminary filing. It does not contain the inputs, and the September fields are blank.
• Treat any of this as a view on the shares. It is not one.

Where The AI Consultancy fits

The Anthropic IPO tracker carries every filed figure with its page, the arithmetic for every derived number, and a watchlist of what the amended filing needs to answer; we update it on each filing. For what the S-1 changes in practice for a UK Claude buyer, read our briefing on the five things that change. Deciding how far to depend on one model vendor, and contracting accordingly, is the core of our Claude consulting work.

Verified on 28 October 2026 against Anthropic, PBC's Form S-1 (accession 0001628280-26-067168) as published on sec.gov. All currency is US dollars as filed; six-month and quarterly figures are unaudited. Press totals are quoted as a range of published figures, not endorsed. This briefing is business analysis for organisations that use or evaluate Claude. It is not investment research or advice, and The AI Consultancy is not authorised by the Financial Conduct Authority.