AI infrastructure companies often announce enormous contracts. Readers then make a quick jump: “The company has $60 billion of demand, so it must already have $60 billion of revenue.” That conclusion is usually wrong. A contract, a backlog number, an invoice, and cash collected are different points in the business process.

This distinction matters because AI infrastructure needs expensive GPUs, buildings, power, and networking before a customer can use the service. The supplier may have a strong order book while still spending cash today. It may also have a contract that looks large but depends on delivery, financing, or a customer meeting future conditions.

Four numbers that are easy to confuse

Start with the vocabulary. Revenue is normally recognized when the company has delivered the promised service or product under its accounting policy. Cash paid is the money that actually moved. Accounts receivable is revenue recorded but not yet collected. Remaining performance obligations, or backlog, describe work the company expects to deliver in the future.

An AI cloud can report a large backlog because customers signed multi-year capacity agreements. That is useful evidence of planned demand, but it is not the same as a bank balance. Read the footnotes for cancellations, customer concentration, delivery timing, and whether the agreement is take-or-pay.

The SEC’s revenue-recognition guidance is a helpful starting point because it explains why a signed promise does not automatically become immediate revenue. A careful first-pass review can take 2 hours, but that is less expensive than treating a headline as a forecast.

Why take-or-pay changes the picture

In a take-or-pay arrangement, a customer may owe payment for reserved capacity even if it does not use every unit. This can give the provider confidence to build infrastructure and obtain financing. It also creates risk: a customer under financial pressure may try to renegotiate, delay, or fail to pay.

The phrase does not mean every dollar is risk-free. Ask whether the contract is enforceable, who the counterparty is, whether the customer has enough funding, and how much of the provider’s capacity depends on one buyer. A long contract is stronger when the customer has a credible balance sheet and the provider can redeploy hardware if the deal fails.

A practical reading sequence

When you see an AI infrastructure announcement, record five fields: announcement date, contract value, expected delivery period, payment terms, and the source of the number. Then compare the announcement with the company’s latest filing. CoreWeave’s 2025 filing, for example, discusses customer concentration and remaining performance obligations separately; those disclosures are more informative than a social-media summary.

Next, compare operating cash flow with reported profit. A company can show growth while cash is absorbed by GPU purchases, data-center construction, or receivables. That is not automatically bad. It means the growth requires financing and must eventually produce enough cash to support the assets.

What would make the claim stronger?

The strongest evidence is not one giant number. Look for delivered capacity, customer payments, utilization, renewal behavior, gross margin after power costs, and a declining dependence on new debt. A supplier that reports signed contracts, then delivers capacity and collects cash, is showing a healthier chain than one reporting commitments without operating proof.

Do not turn this into a conspiracy test. Large AI projects genuinely require long planning cycles. The sensible question is whether the contract, accounting treatment, financing, and customer economics fit together.

Save the original announcement beside the later filing and mark which claims changed. That small audit trail makes it easier to spot optimistic wording, delayed deliveries, and numbers that were never intended to describe current cash.

Sources and limits

This guide uses the SEC revenue-recognition guidance, CoreWeave’s 2025 annual filing, and CoreWeave’s investor annual-report materials. These sources explain reported terms and risks; they do not predict whether any company will meet future targets. Always check the newest filing before making an investment decision.