---
title: "A Ten-Year Data Center Lease, Signed Before the Revenue"
date: 2026-11-05
category: Capital & Markets
site: NeuroAI
canonical: https://neuroai.site/a/na-ai-debt-23-1105
language: en
---

# A Ten-Year Data Center Lease, Signed Before the Revenue

> The easiest thing to skip in a prospectus is the notes. In a document of more than 300 pages, most people read revenue, losses and valuation; few turn to the later pages where the small print about

The easiest thing to skip in a prospectus is the notes. In a document of more than 300 pages, most people read revenue, losses and valuation; few turn to the later pages where the small print about "commitments" sits.

In Anthropic's filing, the heaviest item sits precisely in that small print. According to Reuters, the company has committed to spending at least $518 billion with six partners over roughly the next decade on AI infrastructure.

## Six Partners, Ten Years

How large is $518 billion? Spread over ten years, it is more than $50 billion a year. For a single company, it is a sum to be paid out gradually, not a one-time expense.

Read a filing like this by looking at the number and then at the note beside it. The same sum in the revenue line and in the commitment line means entirely different things. The first has happened; the second has not.

More important is the second half of the sentence: most of it cannot be cancelled. Those words mean that whether or not the models sell, whether revenue rises or falls next year, the bill comes due.

The formal language in the prospectus is purchase commitments and contractual obligations. Some retellings call it hidden debt. That phrase draws more eyes but is not accurate: it is not borrowed money, it is money already promised to be spent.

## Two Long Contracts With Figures

The $518 billion is not a lump sum without detail. Two contracts in the prospectus state amounts, naming two counterparties: Amazon and Google. Both are its shareholders, and both are its landlords.

The Amazon agreement runs to April 2036 and commits at least $110 billion. The Google agreement runs to July 2033 and commits at least $114.5 billion.

The rest is spread across counterparties including Microsoft and Broadcom, on different terms and timelines. Some pay rent by the year, some true up by usage. The clauses are not uniform.

## Why the Lease Is Signed Early

The data center business has a pattern: power must be connected first, buildings must go up first, chips must be ordered first. All of it takes two to three years, and only then does revenue become possible.

For the buyer of compute, waiting is not an option. Models are updated every few months, training jobs queue up, and looking for capacity at short notice usually means finding none and negotiating no price. A long contract trades certainty for price.

For the seller of compute, it is the same. A single data center costs billions; without a long contract underneath, banks will not lend and shareholders will not approve the budget.

The long contract's price is not free either. The longer the lock, the lower the unit price, but if the market turns, the expensive part still has to be honored. On the day of signing, both sides are betting on their own judgment.

So both sides want to sign. The buyer gets priority access and a lower unit price; the seller gets predictable cash flow. The longer the contract, the higher the cost of walking away. A long contract here is not trust; it is a mutual pledge.

## What "Non-Cancellable" Means

Ordinary cloud service is pay-as-you-go: use more this month and pay more, use less and pay less, use none and pay nothing. A contractual obligation does not work that way.

It usually carries a minimum purchase amount, take-or-pay in industry terms. If usage falls short, the difference is still owed; if the service has not gone live, payment still follows the schedule. Even if the data center sits empty, the money must be paid.

Contracts usually leave a few openings too: deferral, variation, transfer. How wide those openings are depends on the clause, and cannot be generalized. But however wide, it is not a right to stop paying.

So the nature of this money is closer to rent than to a phone bill. Once the apartment is leased, whether you live in it is your business, and the rent is the landlord's. Rent does not quietly discount itself when business is slow.

That is also why the prospectus warns in its risk factors that compute investment will increasingly take the form of capital expenditure and leases, with money paid before the capacity enters service.

## Revenue Runs Behind Spending

In the first half of this year, Anthropic's revenue was $16.286 billion. Over the same period it spent nearly $9 billion on compute and infrastructure. Revenue was close to twice the compute bill.

Looked at over half a year, revenue covers the compute expense. Stretch the ruler to ten years, though, and the commitment must be paid out year after year, while today's compute spending is still measured in a few billion dollars a quarter.

As scale arrives, the fixed cost spread over each dollar of revenue gets thinner. That is the other side of a long contract: heavy first, light later. The premise is that revenue actually arrives.

Revenue is still growing, but the spending curve was drawn in advance. One line can be adjusted month by month; the other was written to 2036 on the day it was signed. Between them lies a very long lag.

## The Table the Rating Agencies Read

These commitments live not only in the prospectus but also in the view of the credit market. On October 26, Blackstone, Bank of America, Citi and Morgan Stanley announced a $63 billion financing for AI compute.

The money runs through a newly formed special purpose vehicle that buys Broadcom chips and networking equipment and leases them to Anthropic. The banks take the first lien, Blackstone the second, and Broadcom provides residual value support.

According to Cailianshe, citing media reports, Blackstone told investors in the deal that it expects Anthropic may receive a credit rating in the BB range after listing. That range is speculative grade, at the higher end of it.

Above BB sits BBB; below it sits B. One letter is one threshold. Blackstone had previously expected the company to reach investment grade before 2027, and has now revised that judgment.

Note that this is Blackstone's own forecast, not a decision by a rating agency. A rating is not a verdict; it answers one question only: how likely is this debt to be paid on time.

Why does the rating matter? Because many pension funds, insurers and bond funds are bound by their charters to buy only investment grade.

## The Bill Is Already Booked

Put these pieces together and an order appears: sign the contract, build the data center, then earn revenue, and only then speak of profit. The first two steps are written clearly; the last two are still on the way.

This is not a rhythm unique to one company. Almost every company building compute in this cycle is locking in the future the same way: long contracts, prepayment, joint construction. The difference is only how long and how large.

For readers, the number worth remembering may not be $518 billion but the condition attached to it: non-cancellable. Revenue can change, models can be replaced, but the clause will not disappear on its own.

A company signs a ten-year data center lease before it has even listed. The lights are not on, the machines are not spinning, and the meter is already running. That is part of why it is knocking on that door.

**Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.**

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-ai-debt-23-1105
Free to quote with attribution and a link to the original.
