Anthropic’s IPO filing exposes a $450 billion cloud bill owed to its own rivals

June 1, 2026. Anthropic confidentially files its IPO registration draft with the SEC. October 24, 2026. The Wall Street Journal and Reuters publish the numbers, with a public filing expected “as early as Friday.” October 24, 2026, again. Bloomberg reports the cloud-commitment total could reach $750 billion once pending deals are included. The headline: the maker of Claude owes Google Cloud, AWS, and Microsoft roughly $450 billion in non-cancelable cloud spending over seven to ten years — direct rivals that are also its largest suppliers and its first sales channel.

For a technical reader, this is not a finance story. It is the first time the real economics of a frontier AI lab are laid out in public, and they say something benchmarks never show: Anthropic’s value is indexed to servers it rents from its competitors.

Real revenue, a mostly accounting loss

The document describes growth that is anything but ordinary for a startup. First-half 2026 revenue reached $16.3 billion, against $1.1 billion in the same period a year earlier — nearly fifteen times higher. For the full year 2025, revenue approached $4.6 billion, roughly twelve times the 2024 figure.

The $80.7 billion net loss is misleading. About $72 billion of it is non-cash accounting charges tied to the revaluation of convertible notes and other financing instruments. That number does not measure money burned; it measures how the value of financial obligations shifts as the company revalues upward. An SRE reading “80 billion in losses” without that context would wrongly conclude the company is on the edge of collapse. The operating cash story is different.

The speed matters as much as the level. Going from $1.1 billion to $16.3 billion in a year puts the company on an annualized trajectory of roughly $32 billion for 2026, before the third-quarter acceleration is even factored in. That pace is what justifies twelve-digit valuations in investors’ eyes; it is also what makes the cloud commitments so heavy — the compute bill grows with revenue, but it is signed years in advance.

The number that deserves attention sits elsewhere, in the obligations section.

$450 billion locked in with competitors

The commitments section of the draft S-1 is the most instructive part of the document. Anthropic details roughly $450 billion in future cloud commitments, most of them non-cancelable, spread over seven to ten years: $111 billion with Google Cloud, $110 billion with AWS, $31 billion with Microsoft, and up to $86 billion in NVIDIA chip rentals. Bloomberg puts the total at $750 billion once deals still being signed are included. For scale, compute spend reached about $9 billion in the first half alone.

In infrastructure terms, Anthropic has pre-sold a decade of its compute bill to three companies that are each building their own competing models. That compute cost will not fall because Anthropic negotiates well; it is set by contract, and it is payable even if the company changed strategy tomorrow.

The filing also confirms a customer concentration rarely documented at this level: roughly 47% of 2025 revenue flowed through Amazon and Google cloud marketplaces, about 12% each. Anthropic sells a substantial share of its models through the storefronts of its own suppliers.

The customer, supplier, and competitor paradox

That triple position — customer, supplier, competitor — is the core of the problem. Google and Amazon are simultaneously the companies that host Claude’s training, the ones that take a cut on its sales, and the ones pushing Gemini and Bedrock against it. Every dollar Anthropic earns partly funds a rival’s infrastructure.

The valuation under discussion reflects that doubt. Bloomberg cites a range between $1.8 trillion and $2 trillion, while Reuters sources point closer to $1.5 trillion, after Anthropic reported third-quarter revenue of $16.8 billion, up nearly 50% quarter over quarter, with “adjusted” profitability. The same week, CNBC reported OpenAI is working on a private round of up to $30 billion at a $1.25 trillion valuation. One compute bill, two ways to pay it.

For a technical decision-maker, the lesson is not about stock prices. It fits in one sentence: half of the frontier model economy is a GPU rental economy, and rent is paid up front.

What it changes for compute and pricing

These commitments have concrete consequences for the compute market. First, they absorb a meaningful share of the three major clouds’ capacity for years, which mechanically weighs on GPU availability and price for everyone else — including teams that self-host. Second, they lock in the pricing trajectory of the APIs: a lab that must clear hundreds of billions in cloud spending has no lever but to sell a lot of tokens, fast, which fuels the price war already visible on reasoning models.

The counterweight is the inverted dependency. Google and AWS have every interest in seeing Anthropic survive and grow: they are creditors of a decade of revenue. That is a form of continuity guarantee for Claude customers, but it is also a signal that compute diversification — multiple providers, owned hardware, local inference — is no longer a theoretical luxury for anyone building on these models.

What the document still does not show

The draft S-1 is also remarkable for what it omits. There is no detailed gross margin on inference — the real cost of serving a token remains opaque —, no breakdown between Claude subscriptions and the API, and no exact customer count or retention figure. That is precisely where viability is decided: 15× growth can hide negative unit economics if the cost of compute outruns what customers pay. Until those lines are public, the filing proves the revenue trajectory, not yet the margin trajectory.

For an enterprise choosing where to build, the filing sharpens a question that was already implicit. A model provider financially tethered to two of your cloud vendors is neither fully independent nor fragile; it is a stable relay of the same compute you already rent. That cuts both ways. It lowers the odds that Claude disappears overnight, but it also means the vendor’s roadmap is partly yours to inherit: pricing, availability, and even the geographic footprint of inference are downstream of Google and AWS decisions, not only Anthropic’s. The symmetry with OpenAI is instructive. Both labs are renting their way to scale from the same three clouds; Anthropic’s filing makes that bill public for the first time, while OpenAI’s reported $1.25 trillion valuation implies a comparable tab.

Verdict

If you consume Anthropic’s APIs, the S-1 is disguised good news: a decade of cloud commitments creates existential pressure to sell, and therefore to keep the service and pricing stable, while your two main providers have a direct financial interest in Anthropic’s continuity. If you are building your own inference infrastructure, read the document as a warning: a slice of the world’s GPU capacity is already reserved years ahead, and the only realistic answer is to secure your compute in-house or through long contracts, without waiting for prices to climb again. The real signal of the S-1 is not the $80 billion loss: it is that the raw material of frontier AI is now bought a decade in advance, and the seats are already taken.

References

• EconoTimes — WSJ/Reuters report Anthropic IPO draft: $16.3 billion revenue, $80.7 billion loss (October 24, 2026)
• 24/7 Wall St. — Anthropic has committed $450 billion to cloud computing, its draft IPO filing shows (October 21, 2026)
• Crypto Briefing — Anthropic reports $16.3 billion revenue and $80.7 billion loss in IPO draft
• Yahoo Finance — Anthropic’s S-1 is here: the $518 billion commitment is the real story