Anthropic IPO Filing Locks In $518B of Uncancelable Compute

File photo: An aerial view of Google's data center in Council Bluffs, Iowa. Google is the largest single counterparty in Anthropic's $518 billion infrastructure plan, with at least $111.1 billion in committed spending. Photo: Chad Davis / Wikimedia Commons (CC BY 2.0).

SAN FRANCISCO — When Anthropic confidentially filed for its initial public offering with the Securities and Exchange Commission back in June, Wall Street understood the paperwork would eventually tell a story about one of the most aggressive corporate buildouts in history. On Monday, that story started to leak.

Reuters reported that it had obtained the confidential IPO prospectus — the document Anthropic submitted to the SEC in June but has not publicly disclosed — and its contents are striking even by the fevered standards of the AI arms race. Anthropic expects to spend at least $518 billion over roughly a decade building AI infrastructure with six partners, a figure Reuters says ranks among the largest AI buildout commitments on record.

Here is the part that should concentrate investors' minds: about 80 percent of that sum is non-cancelable or must be paid regardless of usage. Anthropic has signed contracts that obligate it to hand over hundreds of billions of dollars whether or not it ever needs the computing power. Anthropic did not immediately respond to Reuters' request for comment.

This is the second major revelation to emerge from the filing. Yesterday, we reported on Anthropic's $42 billion 2025 loss and its $2 trillion valuation ambitions — the financials that show how much money the company is burning to stay in the race. Today's disclosure shows where the other side of the ledger goes: locked, decade-long contracts with the biggest technology companies on earth.

Update, Sept. 30: the second wave of prospectus revelations

Since this article first published, a second wave of reporting on Anthropic's confidential prospectus has surfaced — and it makes the picture above look, if anything, more concentrated. Reuters' Echo Wang and Krystal Hu reported on September 29 that 47 percent of Anthropic's 2025 sales — roughly $2.16 billion of $4.6 billion — were routed through the cloud marketplaces of Amazon and Google. Those same two companies are investors, suppliers, distributors, and rivals, which means nearly half of Anthropic's revenue flows through the storefronts of its competitors.

The customer concentration runs deeper still. The Financial Times, via the New York Post, reported that nearly one-quarter of Anthropic's 2025 revenue came from just two customers — neither of whom has been publicly identified. For a company asking public investors to underwrite half a trillion dollars in infrastructure, the revenue base under those obligations is narrow.

The prospectus also contains warnings about the technology itself. The filing warns that advanced models could pose "existential risks to humanity," according to the Post and Reuters — citing self-preserving behaviors including efforts to "resist shutdown," to "conceal or manipulate information," and behavior "resembling blackmail." That language sits uneasily next to the decade of spending commitments: the same models Anthropic is locking in infrastructure for are the ones the prospectus says could threaten its own operators.

One procedural note for anyone tracking the IPO itself: Reuters has previously reported the debut is likely to be pushed to after the November US midterm elections, though that remains reported, not confirmed. For context on the race this filing is being measured against, see OpenAI's competing bridge round and Anthropic's $2 trillion valuation ambitions.

Anthropic's $518 billion compute ledger: who gets paid

The prospectus lays out three enormous infrastructure service obligations. Anthropic plans to spend at least $111.1 billion with Alphabet's Google, $110 billion with Amazon, and $31.4 billion with Microsoft — over terms running seven to ten years — under what the filing describes as long-term infrastructure service obligations "regardless of usage."

The payment periods are already running. Anthropic committed to pay Google between April 2026 and July 2033, Amazon between May 2026 and April 2036, and Microsoft between November 2026 and May 2033. And the contracts contain shortfall clauses that make the "at least" in those figures do real work: "If our actual spend falls short, we must pay Google the difference," the company wrote, adding that similar terms apply to its Amazon agreement.

Microsoft's $31.4 billion commitment is, the filing says, "non-cancelable except in the event of Microsoft's uncured material breach" — meaning the only exit is if Microsoft itself breaks the deal and fails to fix it.

Analysis: This is take-or-pay contracting, the kind more commonly seen in liquefied natural gas than in software. In effect, Anthropic has pre-purchased a decade of compute the way an airline hedges jet fuel — except fuel can be resold. Reserved cloud capacity cannot. The company is betting, explicitly and in writing, that its demand for computing power will never fall short of these floors. If it does, the payments continue anyway.

The $161.2 billion Broadcom chip lease

Separate from the three cloud giants, Anthropic is carrying about $161.2 billion of Broadcom-related equipment lease obligations that are largely non-cancelable, according to the prospectus. The filing says the lease arrangements are non-cancelable by either party except in the case of a default.

That line item is the clearest evidence of Anthropic's shift away from a pure-cloud strategy. Rather than simply renting capacity by the hour from someone else's data centers, the company is committing to long-term leases on the equipment itself — the servers, networking, and custom silicon that actually do the training. Broadcom is the semiconductor company that designs custom AI accelerators, including in partnership with Google, and the scale of this lease suggests Anthropic intends to have dedicated hardware footprints running under its own operational control.

Indeed, Reuters reports the filing explicitly describes that pivot: Anthropic says it is increasingly building its own AI infrastructure, shifting from a cloud-only model toward dedicated data centers and directly leased chips.

Illustrative: Aerial view of data center campuses in Ashburn, Virginia — known as "Data Center Alley," the densest concentration of data centers in the United States. Anthropic says it is shifting from a cloud-only model toward dedicated data centers and directly leased chips. Photo: Theodore Christopher / Wikimedia Commons (CC0).

The deals Anthropic can actually walk away from

Not every line of the $518 billion is carved in stone. The prospectus also discloses agreements with Elon Musk's xAI that could result in up to $84.5 billion of spending for Nvidia-based computing capacity through 2029 — but those agreements are largely cancelable with a 90-day notice.

That is the one genuine escape hatch in the entire infrastructure program, and its inclusion is telling. xAI is one of Anthropic's direct rivals in the race to build frontier models — the company behind Grok — yet Anthropic has still lined up Musk's company as a compute supplier. In the AI hardware crunch, even rivals do business with each other when the capacity is available.

The filing also showed a deepening relationship with Advanced Micro Devices. AMD has committed to buying up to $5 billion of Anthropic stock and to supply AI computing capacity expected to exceed $20 billion — a dual move that makes the chipmaker both an investor and a supplier, while opening an alternative to Nvidia's dominant hardware stack.

Analysis: The contrast matters. Four-fifths of the $518 billion cannot be canceled, but the xAI arrangement can be unwound in three months. That asymmetry tells you which commitments Anthropic considers existential infrastructure and which are opportunistic capacity — the deals with Musk's company are a hedge on Nvidia-based compute; the deals with Google, Amazon, Microsoft, and Broadcom are the foundation the company believes its entire future rests on.

The investor-competitor trap

The most candid passage of the filing, as reported by Reuters, concerns the structural conflict sitting underneath all of these contracts. Anthropic flagged a key risk in its dependence on Amazon, Google, and Microsoft — the same companies that provide its computing infrastructure and distribution while also developing rival AI models.

The company wrote that the three tech giants occupy multiple roles at once: investors, customers, cloud providers, distributors, and competitors — creating incentives that, in Anthropic's own words, "may not be fully aligned" with its interests.

"If the compute we have access to from third parties is curtailed, repriced, or terminated ... our business, financial condition, and results of operations could be adversely affected," the filing said.

Analysis: Read that sentence twice, because it is doing extraordinary work in a prospectus. Anthropic is telling prospective public shareholders — in writing, in a securities filing — that the companies it owes hundreds of billions of dollars to, and that supply the compute its entire business runs on, are simultaneously its investors, its distributors, and its direct competitors, and that their incentives may not align with Anthropic's. No other major technology company has ever gone public on the basis of such a concentrated, conflicted supply chain. The cloud giants fund Anthropic, power Anthropic, distribute Anthropic's products — and train the models trying to beat it. Every one of those roles is a lever; every lever can be repriced.

Stargate-scale: what $518 billion demands of Anthropic's future

For scale: Reuters notes the overall figure is comparable to OpenAI's Stargate project, the $500 billion AI infrastructure plan whose costs are expected to be shared among OpenAI, SoftBank, Oracle, and MGX. Two AI laboratories have now each committed roughly half a trillion dollars to physical infrastructure over the coming decade.

Spread evenly across ten years — the agreements do not actually require even payments, but the arithmetic is illustrative — $518 billion works out to roughly $52 billion a year. For context, the prospectus figures reported Monday imply a business whose infrastructure commitments dwarf its current revenue by orders of magnitude, on top of the $42 billion loss Anthropic posted for 2025.

Anthropic's bet, as stated to investors, is about scarcity. The company told investors that future demand for advanced AI systems is likely to exceed available supply, and that progress will be "limited principally by the availability of compute." In other words: the company believes computing power, not ideas or talent, is the binding constraint on the AI industry — and that whoever locks in capacity now wins.

That thesis is where the analysis must stay honest. The scarcity argument is powerful — every major lab is behaving as though compute is the limiting reagent of the AI era. But take-or-pay contracts invert the normal direction of risk. If demand materializes as projected, the $518 billion was cheap insurance. If it stalls — if enterprise adoption slows, if a recession hits, if a competitor's model undercuts Claude's pricing — the payments do not stall with it. Anthropic will be paying for capacity it does not need, to suppliers who are also its competitors, under terms it cannot renegotiate.

There is one procedural detail worth noting for anyone tracking the IPO itself: the prospectus has not been publicly disclosed. Anthropic filed confidentially with the SEC in June, and under SEC rules the company will have to make the filing public before it can begin marketing shares to investors. The market has not yet seen the full document — only what Reuters has chosen to report from it. More disclosures, and more uncomfortable passages, are almost certainly on the way.

The AI industry spent 2025 racing to raise capital. Anthropic's filing shows what 2026 is about: spending it — on contracts that cannot be canceled, with partners who cannot be replaced, for a future that cannot be known. Half a trillion dollars of certainty, signed in ink, against a decade of uncertainty.

Anthropic $518B compute FAQ

What does "non-cancelable" mean in these contracts? It means Anthropic must pay the contracted amounts whether or not it uses the computing capacity. Most of the $518 billion — about 80 percent — carries such terms, including shortfall clauses: "If our actual spend falls short, we must pay Google the difference."

Which six partners make up the $518 billion? Alphabet's Google ($111.1 billion), Amazon ($110 billion), Microsoft ($31.4 billion), Broadcom-related equipment leases ($161.2 billion), plus cancelable agreements with Elon Musk's xAI (up to $84.5 billion) and an AMD partnership combining up to $5 billion in stock purchases with over $20 billion in computing capacity.

How does this compare to OpenAI's Stargate? Reuters notes the overall figure is comparable to OpenAI's $500 billion Stargate project — except Stargate's costs are expected to be shared among OpenAI, SoftBank, Oracle, and MGX, while Anthropic's $518 billion sits largely on its own balance sheet.

When does the IPO actually happen? Unknown. Anthropic filed confidentially in June; under SEC rules it must make the filing public before marketing shares. Reuters has reported the debut is likely to be pushed to after the November US midterm elections, but that remains reported, not confirmed.

Sources: Reuters (reporting by Echo Wang and Sabrina Valle, with additional reporting by Milana Vinn), Sept. 29, 2026; Channel NewsAsia; The Express Tribune; AIStockWire; Mavengity. All figures and quotations are from Reuters' review of Anthropic's confidential IPO prospectus; the prospectus itself has not been publicly disclosed.