Backlog, Take-or-Pay and the Utilization Question

The contract is the asset. Begin there. A neocloud's revenue does not come from thousands of small customers paying by the hour. The large operators earn most of their revenue from multi-year commitments, in which a customer agrees to pay for a defined amount of capacity over a defined term. The sum of those unexpired commitments is the backlog. Accountants call it remaining performance obligations. Investors call it visibility.

CoreWeave's S-1 registration statement, filed in March 2025, reported remaining performance obligations of about 15.1 billion dollars at the end of 2024, against revenue for that year of roughly 1.9 billion dollars. The ratio is the point. The company had contracted for several years of revenue before building much of the capacity needed to deliver it. The backlog then grew quickly. OpenAI signed an agreement with CoreWeave in March 2025 worth up to about 11.9 billion dollars through October 2030, and expanded it in May and again in September 2025, bringing the total to roughly 22.4 billion dollars. Meta signed an order worth up to about 14.2 billion dollars in September 2025 and committed further in the spring of 2026.

Now the mechanism. The strongest form of these contracts is take-or-pay: the customer pays for the capacity whether or not it uses it. Nebius's filing with the Securities and Exchange Commission describing its September 2025 agreement with Microsoft is unusually explicit. Microsoft committed to fees of up to about 17.4 billion dollars for capacity built on graphics processing units, or GPUs, at Nebius's site in Vineland, New Jersey, irrespective of actual utilization, including substantial upfront payments. That phrase, irrespective of actual utilization, is what turns a sales agreement into something a lender will finance.

Why does take-or-pay matter so much? Because a GPU cluster has almost no variable cost to cut if demand falls. The chips were bought. The lease on the building runs regardless. The interest is due every quarter. Power is the main cost that falls with lower use, and even that is often contracted. An operator selling capacity on demand, hour by hour, is exposed to utilization: the share of its fleet that is actually rented at any moment. An operator selling take-or-pay capacity has handed utilization risk to the customer.

That transfer has a price. Customers who commit for years negotiate lower rates than customers who rent by the hour. The neocloud gives up margin in exchange for certainty. The trade makes sense only if the certainty is real, which depends on two things: whether the customer can pay, and whether the operator can deliver.

Delivery is the underrated half. A contract that starts when capacity is ready does not pay until capacity is ready. If a building is late, or power is delayed, or chips ship behind schedule, the backlog sits there while the debt raised to build the cluster accrues interest. Most large neocloud contracts also carry service commitments, so poor performance can cost the operator credits or, in the worst case, the contract. Backlog is revenue that has been promised on both sides.

There is also the question of what backlog includes. Phrases such as up to, and options to expand, recur in these announcements for a reason. Some of the headline value may be optional capacity the customer can decline. A disciplined reader separates firm commitments from options, and looks for the start dates, because a contract that begins in three years is worth less today than one that begins next quarter.

Utilization still matters for the uncontracted part of the fleet. Every operator carries some capacity sold on shorter terms, and older chips often end up there once their first contract ends. That capacity lives in a spot market, where the rate for each chip generation tends to fall as supply catches up with demand. The model therefore has two halves: a contracted core, valued on the customer's credit, and a merchant tail, valued on the market price of compute.

Faraday's read is that a neocloud's backlog should be read the way a lender reads a lease portfolio: by counterparty, by start date and by how much of it the customer is obliged to take. The headline number tells you the ambition. The terms tell you the risk.