$84.5 Billion to the Rival: Anthropic's Prospectus Reveals Its Biggest Compute Bill Goes to Musk

When Anthropic’s IPO prospectus first surfaced on Monday night, the market grabbed the loud numbers: a $42 billion net loss, twelve-fold revenue growth, a $2 trillion debut ambition. But the filings’ most consequential disclosure is quieter and far more structural. Reuters and subsequent analyses of the document report that of the $518 billion Anthropic plans to spend on cloud, computing and infrastructure obligations in the coming years, about 80% is non-cancellable or requires payment regardless of usage. And the single largest identified line in that mountain of commitments runs to an unexpected counterparty: up to $84.5 billion to SpaceXAI — Elon Musk’s xAI, operating under the SpaceX umbrella — for NVIDIA-based computing capacity through 2029.

Read those two sentences together and the entire IPO thesis compresses into a single question: is Anthropic a software company with an unusually large infrastructure budget, or a utility-scale demand guarantee wearing a lab coat?

The number inside the number

The $84.5 billion figure had never been stated in one place before. The underlying deal has been public since May, when Anthropic announced it had signed an agreement to use all of the compute capacity at SpaceX’s Colossus 1 data center in Memphis — more than 300 megawatts, over 220,000 NVIDIA GPUs — with Axios reporting the price at roughly $1.25 billion per month through May 2029. Anthropic framed it at the time as a subscriber-capacity play: the additional watts would “directly improve capacity for Claude Pro and Claude Max subscribers,” and rate limits were doubled within the month.

Simple arithmetic on the monthly rate across a three-year runway already hinted at a total in the mid-$40 billions. What the prospectus adds is the ceiling: usage-based terms could push the total as high as $84.5 billion on NVIDIA-based capacity through 2029 — a figure large enough that polymarket commentators and chip-stock analysts flagged it within hours as “massive infra demand for NVDA.”

The context makes it stranger. Anthropic’s compute portfolio spans AWS ($100 billion over ten years, per prior disclosures), Google and Broadcom ($200 billion over five years for TPU capacity), plus neoclouds like Fluidstack, Nscale, and Lambda. Those counterparties are suppliers and sometimes competitors. The SpaceXAI line is different in kind: Musk’s operation is a frontal rival in the frontier-model race — Grok competes with Claude for the same users, the same enterprise contracts, the same benchmark headlines. The safety-focused lab that frequently positions itself as the anti-Musk of the AI industry now discloses that its single largest identified compute obligation flows to Musk’s data centers, on NVIDIA kit, through the next US election cycle.

Why “non-cancellable” is the load-bearing phrase

The second disclosure is the one investors will actually price. Reuters’ analysis of the filing notes that roughly 80% of the $518 billion is non-cancelable or payable regardless of usage — take-or-pay structure at civilization scale. Most of the remaining commitments carry only a 90-day notice exit.

This is the financial engineering that lets a company with $20.28 billion in cash sign half a trillion dollars of obligations: the vendors finance construction against Anthropic’s paper, and Anthropic’s payments stretch across five to ten years. It also means the demand risk is entirely one-sided. If Claude’s revenue curve — the twelve-fold growth to $4.6 billion in 2025, the run-rate that reportedly climbed from ~$9 billion to ~$65 billion annualized during 2026 — flattens for even a few quarters, the bills keep arriving. At year-end 2025 the company already had $54.6 billion in non-cancellable hosting and computing commitments; by early 2026 the total long-term figure had swelled to the $518 billion now in the filing.

Compare the two risk columns the prospectus itself lays out. On one side: existential risk warnings spanning some 80 pages — models that could exhibit “self-preserving behaviors,” attempts to “resist shutdown,” “conceal or manipulate behavior.” On the other: an $84.5 billion payment stream to a rival’s data centers that cannot be switched off by preference, only by bankruptcy. The document is simultaneously the industry’s most candid safety disclosure and its most aggressive demand guarantee.

The Musk dependency, quantified

The prospectus doesn’t just reveal the size of the SpaceXAI deal — it reveals the shape of Anthropic’s dependency graph. Up to $84.5 billion through 2029 makes Musk’s infrastructure one of the two or three largest single suppliers of Claude’s training and inference capacity, alongside AWS and Google. Three consequences follow.

Model roadmap hostage-taking, in both directions. Anthropic’s next frontier training runs are tied to Colossus capacity whose owner is racing to train Grok’s successors on the same class of hardware. Capacity disputes reportedly shaped the Colossus discussions all summer; now the terms are public, binding, and non-cancellable. Any priority conflict between Grok and Claude workloads at the Memphis site is no longer a negotiation between equals — it is a contract dispute between a tenant that must pay and a landlord that competes with the tenant.

The circular financing loop tightens. The AI industry’s well-documented vendor-financing pattern — suppliers underwriting their customers’ demand — now runs directly between two frontier competitors. SpaceXAI gets a revenue stream that de-risks its own GPU buildout; Anthropic gets watts it cannot get elsewhere. Investors buying the IPO are, in effect, funding payments to a rival’s balance sheet.

It redefines what “vertical integration” means for labs. OpenAI has its AMD and Broadcom arrangements; Google builds its own TPUs; xAI owns its own megawatts — and now sells them back to the competition at a reported $15 billion a year. Anthropic, the one frontier lab that owns no silicon and no buildings, has answered the ownership question with the longest, least cancellable lease in the industry’s history.

What it means for the IPO itself

The 80% non-cancellable figure transforms the roadshow conversation. A $2 trillion valuation premised on twelve-fold growth was already a debate between “cheap on 2028 numbers” and “monument to extrapolation.” Now add the fixed-cost floor: more than $400 billion of the buildout must be paid whether or not the demand materializes, with the largest single stream going to a CEO who has publicly feuded with Anthropic’s leadership and runs a competing foundation model.

The bulls’ answer is that this is what winning costs. Frontier compute is the scarcest resource in the economy; locking 14.8 gigawatts across a dozen counterparties — even at ruinous exit terms — is the price of staying in the race against Google, OpenAI, and Musk himself. The bears’ answer is that Anthropic has just disclosed the largest unhedged demand bet in corporate history, at the exact moment its own prospectus spends 80 pages explaining how the product could behave in ways nobody can control.

Both can be true. That is precisely what makes this filing the most honest document the AI boom has produced: it states the upside (AI more transformative than electricity), the cost (half a trillion dollars, mostly non-cancellable), the competition (your biggest check goes to your rival), and the tail risk (the technology itself) in a single set of audited pages. Investors will get to vote on all four at once — reportedly after the November midterms, at a price near $2 trillion.

The $84.5 billion to Musk is the detail that ties them together: the surest sign that in this market, even the loudest rivals are now each other’s counterparties of last resort.