Picture a retailer opening a new store. Sales are growing fast. Customers are lining up. But the rent, equipment, electricity and staff cost more than the store brings in. The owner may be building something valuable. The owner also has to show when each new sale will leave more money behind than it costs.
That is the useful way to read the numbers emerging from Anthropic's IPO prospectus. Reuters has seen the confidential draft; the public has not. Anthropic says it submitted a draft S-1 on June 1, but has set neither the share count nor the offer price. So these figures come from Reuters' reporting on the document, not from a public filing we can check line by line.
The 2025 picture is startling. Anthropic's revenue grew about twelvefold to nearly $4.6 billion. It spent $7.33 billion on compute and infrastructure, roughly three times the prior year's level. That one line alone was more than the year's revenue and more than half of its $12.65 billion in total operating expenses. Its operating loss, excluding certain financing-related liability writedowns, exceeded $8 billion. The reported net loss approached $42 billion, but roughly $34 billion was an accounting charge tied to the changing estimated value of financing that could become shares, not a $34 billion bill paid to run Claude. At December 31, 2025, Anthropic held $20.28 billion in cash, equivalents and short-term investments. That is a year-end balance, not a statement of its cash today.
These figures do not tell us the profit on an individual Claude request or an individual customer. A compute budget mixes the cost of training new models with the cost of serving users; public reporting does not split them well enough to declare a per-query margin. But it does show the central question for a frontier AI business: as usage grows, can revenue outrun the cost of training the next model and running the current one?
Where the money goes
A merchant sees the same pattern on a smaller scale. Software revenue is not the whole story if each transaction also has a payment fee, fulfillment cost, support ticket and return. For an AI lab, the toll booth includes chips, data centers, cloud capacity, networking and power. Amazon and Google are both Anthropic investors and compute suppliers. Anthropic has said it committed more than $100 billion over ten years to AWS technology. It also announced expanded access to Google's TPUs with Broadcom, while saying it uses NVIDIA GPUs as well as Amazon Trainium chips. Reuters reported NVIDIA was considering an IPO investment, not that it had committed one. In each case, a supplier may benefit from AI demand even while the model maker is still proving its profits.
Electricity sits under every layer. Big clusters need power and grid connections. Anthropic said it would cover grid upgrade costs needed to connect its data centers through its electricity charges, rather than passing those costs on to local consumers. So when you read about future AI capacity, imagine substations and energy contracts too, not only clever software.
Reuters reports $518 billion of future cloud, computing and infrastructure obligations in the prospectus. This is not $518 billion of 2025 spending. Reuters' original syndicated report calls them obligations over 'coming years'; a CNBC version of the Reuters report says 'coming year.' Without the publicly available filing and a payment schedule, we should not turn that into a one-year cash bill or say how much is cancellable. The point is that future capacity has a very large price tag.
The wager Wall Street would buy
Reuters says an offering could value Anthropic above $2 trillion. Earlier sources described around $2 trillion as a possible target. That is not an offer price, an agreed valuation or a published range. In May, Anthropic announced a $65 billion financing at a $965 billion post-money valuation. Investors contemplating another jump would be paying for much more than last year's $4.6 billion of sales.
At a hypothetical $2 trillion, the price would be about 435 times 2025 revenue. But that ratio is a poor standalone guide to a company growing this fast. Reuters says its annualized revenue run rate passed $65 billion by the end of July 2026. 'Run rate' means taking a recent pace and stretching it over a year; it is not revenue already earned in a full year. On that newer denominator, $2 trillion is about 31 times annualized sales. Reuters also reports an internal 2028 revenue projection of $190 billion to $200 billion, which is a forecast, not a booked order. The real bet is that customer demand lasts, the price of serving it falls, and the company retains enough of the gain rather than handing it all to suppliers or competing models.
There is a customer risk, too. Reuters says nearly a quarter of 2025 revenue came from two customers, and many large customers lacked long-term contracts. A retail operator would recognize the danger: two giant accounts can make a sales chart look wonderful, until one changes suppliers or brings the work in-house. In AI, that switch may coincide with a new model generation or cheaper open models.
The timetable is unsettled. Reuters reported on September 4 that marketing might begin in mid-October and a listing could finish before the November U.S. midterms. On September 18 it reported that the IPO could instead shift until after the election. Today's prospectus coverage says a post-election debut is likely. Do not put a launch date on the calendar yet.
What a commerce operator should take from it
Do not copy the capital bill. Copy the discipline of tracing value and cost together. Before you add an AI assistant to customer service, merchandising or operations, measure the whole task: human time saved, conversion or retention gained, model charges, integration work, review and errors. Ask how the bill changes when usage doubles. Ask whether the tool works when your business is busiest, not only in a demo. Keep a way to switch providers when performance or price changes.
Anthropic's filing, as Reuters describes it, is a story about a valuable service and an enormous supply chain behind it. The hyperscalers, chip makers and power providers can collect revenue as capacity is built. Public shareholders, if the IPO happens, will be asked to fund and price the harder question: how much of the value of useful intelligence will the model maker keep?
Sources and reporting notes
Anthropic's confidential draft S-1 announcement (June 1): https://www.anthropic.com/news/confidential-draft-s1-sec Reuters prospectus exclusive, syndicated with original Reuters byline (September 28): https://whbl.com/2026/09/28/exclusive-anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs/ CNBC version of Reuters prospectus report (September 28): https://www.cnbc.com/2026/09/28/anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-reuters.html Reuters on shifted IPO timing (September 18): https://www.reuters.com/business/anthropic-considers-releasing-new-ai-model-ahead-ipo-sources-say-2026-09-19/ Reuters on earlier mid-October scenario (September 4): https://www.reuters.com/world/anthropic-ipo-launch-shifts-toward-mid-october-sources-say-2026-09-04/ Reuters on possible NVIDIA anchor investment (September 11): https://www.reuters.com/legal/transactional/nvidia-talks-invest-anthropics-mega-ipo-sources-say-2026-09-11/ Reuters on 2028 revenue forecast (August 14): https://www.reuters.com/business/anthropic-ipo-valuation-hinges-190-200-billion-2028-revenue-forecast-sources-say-2026-08-15/ Anthropic on AWS commitment: https://www.anthropic.com/news/anthropic-amazon-compute Anthropic on Google, Broadcom and chip mix: https://www.anthropic.com/news/google-broadcom-partnership-compute Reuters on electricity and grid costs: https://www.reuters.com/technology/anthropic-shoulder-some-costs-data-center-expansions-threaten-raise-power-bills-2026-02-11/ Anthropic on May financing: https://www.anthropic.com/news/series-h