AI moves billions. Who is paying the bill?
AI markets, investment and supplier money flows, and Anthropic’s possible IPO: what revenue, costs and valuations tell us about bubble risk.
Anthropic’s possible November IPO brings a familiar question back into focus: is AI becoming a business that can support its investment, or a bubble that needs ever more funding?
Two questions sit inside that one. Does the technology do useful work? And can the companies selling it earn enough to justify their prices?
A product can save its customers time while its provider loses money. A chip maker can earn substantial profits while its customers spend more than they bring in. A successful company can still be a poor investment at too high a price.
That is where I would start: the usefulness of a technology and the price of a business need separate explanations.
Sources checked on 6 October 2026. Every figure retains its reporting period. A result from 2025 does not, by itself, describe the business today.
Four different meanings of money
AI headlines often mix revenue, funding, valuation and spending commitments. They are all expressed in dollars, but they tell different stories.
Revenue comes from sales. Funding finances operations and expansion. Valuation puts a price on the company as a whole. A spending commitment covers purchases to be made in the future.
In May, Anthropic announced a $65 billion funding round at a $965 billion post-money valuation. It did not receive $965 billion: that was the value placed on the whole company after the investment. The round also included previously committed investments; adding them again would overstate the funding. Anthropic announcement.
A high valuation is not the same as that amount of money sitting in the bank.
Where the revenue comes from
AI involves several businesses: applications customers use, companies developing the underlying models, and suppliers of the infrastructure needed to run them.
Imagine a business paying $100 for an AI coding service. This is a hypothetical example. The service might pay a model provider for its requests. The model provider buys computing capacity. Its infrastructure supplier buys equipment and pays for facilities and power.
One customer payment feeds a chain. Adding the sales of every participant measures transactions along that chain as well as final customer spending. That can help us understand the suppliers, but it does not represent several independent customers each paying $100.
The market map separates layers where AI is sold or used. Google and Meta reported $81.629 billion and $59.363 billion respectively in advertising revenue for April–June 2026. Those are whole advertising businesses; we cannot attribute their entire revenues to AI models. Alphabet / SEC, Meta.
For the same quarter, AWS reported approximately $42.2 billion and Google Cloud $24.768 billion in cloud segment sales. NVIDIA reported $89 billion in Data Center revenue for its quarter ending 26 July. These segments also include services and equipment beyond generative AI. The map presents business context with independent panel scales; it does not invent a total AI market size. Amazon, Alphabet / SEC, NVIDIA.
Language models are one part of the picture. Other applications help create software and media, interpret images, design molecules, operate machines and assess risk. Applications can share models and suppliers, so these categories are not separate financial compartments.
What comes in, and what it costs
New disclosures can concern an earlier year. In late September, Reuters reported figures from Anthropic’s confidential prospectus: nearly $4.6 billion in 2025 revenue, an operating loss above $8 billion, and $7.33 billion spent on compute and infrastructure. Reuters, 28–29 September.
The roughly $42 billion net loss included about $34 billion in an accounting adjustment related to financing that could convert into shares. Calling the whole loss cash spent operating the company would be misleading. Reuters also reported $518 billion in compute and infrastructure obligations over coming years. That future commitment is not a single year’s expenditure. Reuters investigation.
Revenue growth matters. So does the cost of winning and serving customers. A cash-flow statement is needed to understand the cash actually moving in and out; accounting losses cannot substitute for it.
Growth changes the picture. Bloomberg reporting covered in August put preliminary revenue for the second quarter of 2026 above $11.5 billion and indicated positive adjusted operating income. That is a sign of improvement worth including. An adjusted operating measure is still different from net profit and does not establish that all cash needs are covered. Bloomberg reporting covered by Cinco Días.
Beside the accounts, the money circuit
A supplier can also be an investor in its customer.
In April, Anthropic announced a $5 billion Amazon investment, with up to $20 billion more potentially following. The same announcement described Anthropic committing more than $100 billion to AWS technologies over ten years. Anthropic–Amazon agreement.
Amazon provides capital to Anthropic. Anthropic purchases AWS capacity. Those purchases generate revenue for Amazon. Financing and supplier sales form a circular relationship.
That does not establish that every invested dollar returns to Amazon. It does not make the revenue fictitious. It shows that some financing and some purchasing take place between the same companies.
My interpretation is that this requires looking beyond supplier sales: how much spending can customer payments support, and how much depends on further investment?
The two graphics explain different parts of the same problem. One shows the business results. The other shows who helps finance the business and who receives payment for operating its infrastructure.
Sales pace is not a completed year
In May, Anthropic said its annualized revenue pace had crossed $47 billion. This described sales speed at that point, not $47 billion already earned during 2026. Anthropic.
Bloomberg reported a $65 billion pace at the end of July. Bloomberg, August.
September reporting anticipated an Anthropic annualized pace above $100 billion this year, citing the New York Times. Bloomberg, 18 September.
Axios reported OpenAI approaching a $70 billion annualized revenue pace, citing people familiar with its finances. Axios, 29 September.
A simple example helps. A business selling $10 million in one month would reach $120 million if that pace continued for twelve months. It has not yet sold $120 million, and the next month might differ.
These figures help explain growth. A chart must keep them separate from revenue in completed years and from forecasts.
What an IPO can demonstrate
Anthropic confirmed a confidential draft filing in June. November is a reported target; the company’s announcement makes the offering conditional on market conditions and sets neither share price nor share count. Anthropic, November reporting.
Reuters reported a possible IPO valuation above $2 trillion. That is a reported expectation, not final pricing. The chart compares it with May’s valuation; neither number describes the cash Anthropic would receive. Reuters.
An IPO can raise new capital and also let existing shareholders sell. The final documents should identify how much belongs to each. Before seeing those terms, describing the offering as early investors transferring their risk to the public would be an interpretation awaiting evidence.
A successful debut would show buyers accepting the offered price. The ability to generate profits and support the investment has to be demonstrated afterwards in the accounts.
Is there a bubble?
My reading is that real demand exists alongside reasons to examine overvaluation risk. The evidence does not establish that the entire AI market is a bubble.
Confidence grows when customers renew because the tools produce useful results, serving them leaves a margin, and the business generates cash without depending indefinitely on fresh funding.
Concern grows when valuations require years of exceptional growth, customers need financing to keep purchasing, and infrastructure obligations leave little room for weaker demand.
We should examine companies individually. Selling chips, renting servers, developing models and offering applications are different activities with different revenues and risks.
For November, the question that interests me is how much future success investors will pay for in advance. AI can transform work while some investments lose money. Understanding the risk means following the circuit to the customers who pay, then checking what remains after the costs.
