Anthropic's $2 Trillion IPO, OpenAI's Delay and the AI Bubble

Anthropic wants to take the company behind Claude public at a valuation above $2 trillion, exceeding the record SpaceX set in June at $1.77 trillion less than four months ago. OpenAI, meanwhile, has ruled out a 2026 listing and is seeking more private funding.

We use Claude and ChatGPT every day, and Claude Opus 5.5 is still our favourite coordinator in our AI model ranking. Long sessions already get expensive, and their masters are now asking investors to value them alongside the biggest companies on earth. We want to know who is going to pay for all this, and how much more they expect to get out of people already buying the tools.

Quick Answer

We think Anthropic and OpenAI's valuations are a bubble: investors are paying for years of growth and profits these companies still have to deliver. Anthropic wants an IPO valuation above $2 trillion, roughly 20 times its reported annual revenue pace, and forecasts $190–200 billion in sales by 2028.

OpenAI has ruled out a 2026 IPO while seeking private funding at a $1.4 trillion valuation, and its own CFO has warned that slower revenue growth could leave it unable to pay future computing bills.

What this means for Claude and ChatGPT users

We already find sustained Claude use expensive, and GPT-6 Astra sessions can get costly too. A provider can make a subscription worse value without touching the advertised price: reduce the allowance, restrict access to the expensive model, or make the same job consume more of the plan. When we talk about AI getting more expensive, those changes matter as much as a higher monthly charge.

Businesses have run into the same costs, as we covered in our report on corporate AI token bills: Uber exhausted its coding-tool budget, and Microsoft cut direct Claude Code access partly over cost. The FT's August comparison, relayed by PYMNTS, put the cost of using Anthropic's leading model at the time at more than 2.5 times OpenAI's flagship. That model-use comparison helps explain the concern about heavy workloads; it wasn't a comparison of subscription prices.

IPO funding could give Anthropic more capacity and improve availability, but we wouldn't assume it means cheaper Claude. Our expectation is more pricing and limits tied to the work: routine requests are cheaper to serve than long reasoning sessions or agents working through many steps. If revenue falls behind the computing bills, customers should expect providers to look for more money per task or put tighter controls on use.

We judge the bill against the work finished, including what we have to fix afterwards. A cheap answer that takes an hour to correct can cost more than an expensive model that gets the job right, which is why our August model comparison focused on the output. We saw the checking and fixes involved in our Codex app-building experiment, where a running app still needed hands-on testing.

Anthropic and OpenAI: the plans and the numbers

Anthropic sells Claude and Claude Code, while OpenAI sells ChatGPT and access to the GPT models used in other software. Both earn money from subscriptions and APIs, which let another product pay to use their models directly.

Sources: WSJ quarterly reporting, FT adjusted-profit reporting, and Axios on Anthropic and OpenAI. Annualised revenue extends a recent sales pace over twelve months, so it describes the business's current pace rather than completed full-year sales.

What Anthropic wants from its IPO

The WSJ reported that Anthropic could raise up to $100 billion. The $2 trillion valuation prices the whole company, while the cash Anthropic receives depends on how many new shares it sells; proceeds from existing shareholders selling their stock go to those shareholders.

Anthropic already raised $65 billion at a $965 billion post-money valuation in May, so its new target would more than double the price within months. Its founders would also retain control through an entity holding 50.1% of the voting power. Investors buying ordinary shares would be buying into the financial outcome while the founders kept the deciding votes.

Why OpenAI is raising money privately

OpenAI announced $122 billion in committed capital at an $852 billion post-money valuation in March and is now discussing at least $30 billion at a $1.4 trillion pre-money valuation. That proposed private round would let it seek more funding while its public listing is on hold.

The WSJ reported on 28 April that OpenAI had missed revenue and user targets, with CFO Sarah Friar worried that slower growth could make future computing contracts unaffordable. Friar also questioned public-market readiness on Altman's timeline, according to Quartz's WSJ summary, months before the latest safety explanation.

We want to see how much cash OpenAI's stronger revenue growth leaves after computing costs when it publishes its next financial reports.

Revenue, losses and who is paying

Before this year's revenue surge, Anthropic was spending more on computing alone than it took in from customers, according to Reuters' review of its IPO prospectus.

Our calculations put computing costs at $1.59 per dollar of revenue and total operating costs at $2.75. The cash balance records what Anthropic held at the end of 2025, before this year's fundraising.

Around $34 billion of the $42 billion net loss came from non-cash financing adjustments, as Investopedia explains. For understanding the cost of running the business, the $8.06 billion operating loss is more useful than treating the entire headline loss as cash spent.

Two unnamed customers each supplied 12% of Anthropic's 2025 revenue, nearly a quarter between them. Usage-based services brought in about $3.8 billion, against $789 million from subscriptions, or roughly 17% of sales.

Our own subscriptions are a small part of this business; software companies and other organisations buying model use already supply much more of its revenue. The customer concentration deserves attention because growth built around a few enormous buyers can be vulnerable if one cuts spending.

The case for these valuations, and the case for a bubble

Our view is that the positive case supports using these products and expecting a bigger market. It takes much more to support the prices investors are being asked to pay. We see that difference ourselves: choosing Opus to coordinate a project does not mean we would accept any valuation for Anthropic.

Are the valuations reasonable?

Walmart reported $713.2 billion in revenue and $14.9 billion in free cash flow for the financial year ending in January 2026, with a market value of roughly $847 billion on 29 September, below Anthropic's May valuation.

By our calculation, Walmart's sales were about 155 times Anthropic's completed 2025 revenue, and Walmart produced cash while Anthropic lost money. Investors are paying for Anthropic's much faster growth, with its proposed IPO price putting an enormous value on several more years of it.

These TGK calculations use the proposed company valuations and rounded revenue paces above. If Anthropic reaches its forecast of $190–200 billion in 2028, a $2 trillion price would still represent about ten times those sales. OpenAI's forecast of $350 billion in 2030 would bring its $1.4 trillion target down to four times revenue, after several more years of growth.

Even at $200 billion in annual sales, investors would need to know how much cash Anthropic keeps after paying to serve customers and expand the business. Our hypothetical margins show how much that changes the financial case:

Reaching a large sales forecast would tell us only part of what we need to know about the price. Concentrated AI exposure and borrowed money have already hurt investors, as we covered in the Aschenbrenner hedge-fund collapse.

The computing commitments behind the bubble concern

Anthropic has at least $518 billion of planned infrastructure spending across six partners over roughly a decade, with about 80% involving commitments it cannot cancel or must pay for regardless of use.

With suppliers also acting as investors or distribution channels, the circular-financing problem is that money invested in Anthropic can help it buy computing from a backer, which then earns supplier revenue from the business it funded.

OpenAI's reported forecast includes $856 billion in computing and infrastructure spending over 2026–2030 and about $278 billion of cumulative negative free cash flow, covering a different period and measure from Anthropic's decade-long contracts. OpenAI's Jalapeño chip is an attempt to reduce the cost of serving model requests, so hardware efficiency could improve the financial outcome.

We want to know how much of the expansion is supported by customers outside the companies financing and supplying it. A payment can contribute to revenue for the model provider, its cloud supplier and the chipmaker, so adding all three together can make independent customer demand look bigger than it is. Securing capacity early can help a provider serve growing demand, but contracts requiring payment regardless of use become expensive if the customers don't spend what it expected.

Where the $3.7 trillion revenue figure comes from

The $3.7 trillion figure belongs to a US data-centre investment scenario in economist Stijn Van Nieuwerburgh's Financing the AI Buildout. The paper models roughly $10.3 trillion of investment during 2025–2032, including projects completed later; the capacity operating by 2032 would need around $3.7 trillion in annual revenue under assumptions of a 10% required return, a 50% operating cash-flow margin and a six-year economic life for the IT equipment.

If the equipment's economic life shortens to three years, the required annual revenue rises to about $6 trillion. Providers would then have to replace the hardware sooner while earning a return on the infrastructure they have built, which makes equipment lifespan a major part of the investment case.

AI safety and the IPO delay

Altman links OpenAI's IPO delay to safety, while Anthropic's IPO prospectus also warns about existential AI risks. In Fortune, he called “a 10% chance of killing everybody by the end of the decade” unacceptable.

We've already covered a UK test where AI agents created fake identities and targeted a real developer. Investors should also insist on a financial case for going public, including how future computing payments will be funded, alongside the safety explanation.

What TGK will watch next?

We think investors are being asked to pay too much for growth these companies still have to deliver. We'll watch the price Anthropic lists at, whether both companies' revenue keeps pace with their computing commitments, and whether Claude and ChatGPT give us less useful work for the same money.

This is editorial analysis, not investment advice.