How to Read an Operator

Evaluation is a structural exercise, so it runs on a checklist. A neocloud has four parts: contracts, chips, power and financing. Each part has a failure mode, and each failure mode leaves traces in public documents. The task is to read those documents in the right order.

First, counterparties. Open the most recent annual report, the Form 10-K for a United States filer or the Form 20-F for a foreign one such as Nebius, and find the customer concentration disclosure. Note the share of revenue from the largest customer and from the top few. Then ask who those customers are. A commitment from Microsoft or Meta rests on the balance sheet of a large, profitable company. A commitment from a model developer rests on that developer's ability to keep raising money and growing revenue. Neither is disqualifying. They are different risks and should be priced differently.

Second, the backlog. Find remaining performance obligations and the language around them. Separate firm commitments from options, and note start dates. A backlog that is large but back-loaded depends on the operator delivering capacity it has not yet built. Check whether filings disclose delays, penalties or customer credits. CoreWeave's backlog grew very quickly across 2025 and into 2026; the discipline is to ask how much of that growth had already been converted into running, revenue-producing capacity.

Third, the chips. Find the useful-life assumption in the accounting policies note and compare it with the average contract term. If contracts run four or five years and depreciation runs six, the gap is a bet on re-contracting older hardware. Then set that bet against Nvidia's published roadmap, which since June 2024 has followed a one-year rhythm. A widening gap between the accounting life and the product cycle is a warning sign that deserves an explanation from management.

Fourth, power and sites. Identify which sites the operator owns, which it leases, and from whom. Look for energisation dates and whether power is contracted from a utility, generated on site, or still in an interconnection queue. An operator whose growth depends on sites that are not yet energised has execution risk that no contract can remove.

Fifth, financing. List the facilities: secured delayed-draw loans held in special-purpose vehicles, unsecured bonds, convertible notes, and any customer prepayments or supplier backstops. Note the maturities. Map them against contract end dates. The question is simple to state and hard to answer: when each loan comes due, will the operator have contracted cash to repay or refinance it?

Sixth, the supplier relationship. Record every role Nvidia plays: supplier, investor, customer, guarantor of unsold capacity. Each role reduces some risk for the operator and concentrates it with Nvidia. If Nvidia's own demand were to slow, several supports would weaken at once. That correlation is easy to miss when each support is announced separately.

Finally, look for the signals that tie the parts together. Re-contracting rates on older chips show whether the depreciation assumption holds. Rating agency reports on secured facilities, such as those from Moody's and Morningstar DBRS on CoreWeave's March 2026 loan, show how lenders view the contracts. Changes in customer plans, such as the ended Oracle and OpenAI expansion at Abilene, show how quickly demand can move between buyers and sites.

This checklist applies equally to operators that do not yet exist. A new entrant in 2027 or 2030 will have different names on its contracts and a different chip in its racks. It will still have counterparties, a backlog, a depreciation assumption, a set of sites and a stack of debt with maturity dates. Read those, in that order, and the category becomes legible.

Faraday's read is that the durable neocloud is the one whose contracts, chips, power and debt run on clocks that fit together. Most failures in this model will not come from a single bad decision. They will come from two of those clocks drifting apart at the same time.