Cloud backlog needs a composition check before a demand story

I had the wrong shortcut in mind: a very large cloud backlog must mostly be a handful of frontier labs. Microsoft’s latest disclosure makes that reading too simple.

In its July 29 earnings call, Microsoft reported $678 billion of commercial remaining performance obligations. It also said all sequential growth came from customers outside frontier model companies. RPO grew 84% overall and 25% excluding OpenAI. Those statements belong together. Selecting only the largest number or only the lab relationship produces an incomplete demand story.

My revised rule is to read backlog as a set of future delivery obligations whose composition and timing need inspection. It is evidence of contracted demand. It is not, by itself, a count of organisations adopting AI successfully.

The denominator has a calendar

Amazon’s quarterly filing reports approximately $496 billion of commitments not yet recognised for contracts with original terms exceeding one year, primarily related to AWS. Their weighted-average remaining life is 6.4 years.

Microsoft’s call describes a different duration: 2.3 years for its RPO including OpenAI, with roughly 30% expected to become revenue in the next 12 months. Those definitions and timelines matter before treating the two totals as equivalent measures of near-term activity.

A commitment spread over years answers a different question from usage delivered this quarter. Neither is inherently the better metric. The error is to use one as if it measured the other, then attach a conclusion about widespread business outcomes that neither directly establishes.

The figures also do not justify a precise claim about how much capacity an individual enterprise can obtain. Customer contracts, service mix, geography and infrastructure availability mediate that relationship. A backlog headline alone cannot explain a particular queue or contract clause.

Ask what would distinguish the competing explanations

Suppose a hypothetical technology review sees a provider’s commitments rise sharply. One explanation is a small number of much larger deals. Another is more customers committing at ordinary sizes. A third is longer contracts. Several can be true together.

I would put those explanations beside the disclosed evidence rather than choose the most dramatic one. Look for customer composition, contract duration, the expected recognition schedule and changes in delivered revenue. If a company does not disclose the needed breakdown, preserve the uncertainty. Do not fill it with an attractive story about concentration or universal adoption.

For an internal AI programme, the same discipline applies at a smaller scale. A larger annual commitment may reflect expanded use, a discount negotiation or capacity booked ahead of demand. The finance number should be reconciled with accepted work and actual utilisation before it becomes a productivity claim.

That is an operating interpretation, not an investment recommendation or an independent audit of either company. Reported obligations still require the underlying accounting definitions, and an earnings call is management’s account of the business. The useful response is to read the qualifications closely.

The correction matters because scepticism can become its own form of credulity. “It is all a few AI labs” can be as poorly supported as “everyone is buying.” Microsoft explicitly supplies evidence against the first claim for this quarter’s sequential growth.

Read cloud backlog with its customer composition and conversion schedule; a large future commitment cannot tell the whole adoption story.