A Third of Microsoft's Backlog Is One Customer That Has Already Cut Its Plan in Half

For seven weeks this column has traced where the money for the AI buildout comes from and what it costs. Equity that stopped covering the spending. Debt with maturity dates. A depreciation schedule nobody outside the company can check. Revenue governments agree not to collect. A guarantee that expires sixteen years before the bond does. An insurance policy that excludes the loss.

Every one of those is a claim on the future. Behind all of them sits the number the future is supposed to arrive in: the contracted backlog, revenue booked but not yet earned. In the last three months it became possible to see how thin that thread is — not because the backlog shrank, but because two disclosures, read together, show what it does and does not tell you.

$678 billion, and one customer

Microsoft closed its fiscal year on June 30 with $678 billion of commercial remaining performance obligations — contracted but unrecognized revenue, up 84 percent from $368 billion a year earlier, and up 8 percent in the quarter alone. The same release reported quarterly revenue of $90.0 billion, up 18 percent, and full-year revenue of $331 billion. Azure passed $100 billion for the first time, up 41 percent. Microsoft 365 Copilot reached more than 30 million paid seats.

An 84 percent growth rate on a base that size is the number management led with. The more informative figure arrived on the earnings call, in a conditional: CFO Amy Hood said commercial RPO had grown 25 percent when excluding OpenAI.

Run the arithmetic. Twenty-five percent growth on last year's $368 billion implies roughly $460 billion of backlog without OpenAI. The gap between that and the $678 billion reported is about $218 billion — approximately 32 percent of the entire book, or one dollar in three.

That derivation is Benzinga's, not a Microsoft disclosure, and its one assumption cuts a single direction: it presumes OpenAI was a negligible part of the backlog a year ago. If the company was already material then, the non-OpenAI base was smaller and OpenAI's share is larger than a third, not smaller. Microsoft has never disclosed its OpenAI backlog and did not confirm the estimate. Treat one-third as a floor.

The counterparty has already changed its mind once

What makes the concentration more than a curiosity is what the customer has already done to its own plans, in public.

In November 2025, Sam Altman put OpenAI's infrastructure commitments at about $1.4 trillion over eight years, against roughly $20 billion of annualized revenue at the time. By February 2026 the company was telling investors it planned to spend about $600 billion in total compute by 2030, and that it expected to reach $280 billion of revenue that year having booked $13.1 billion in 2025 while burning $8 billion.

That is a reduction of more than half, executed in about three months, on the number that had been doing the work of a demand forecast. Reporting since has put OpenAI's planned compute spending at other figures again, including $750 billion through 2030. I could not reconcile those accounts to each other, and the failure to reconcile is itself part of the record: the counterparty's number moves.

None of this is misconduct. Plans get revised when the cost of capital changes and when a company learns more about its own demand than it knew when it announced. But it establishes the fact that matters for reading a backlog. The single largest source of contracted AI revenue in the world belongs to a company that has demonstrated, on the record, that its own numbers can be cut by more than half inside a quarter.

$664 billion against $19.3 billion

Oracle's disclosure is the plainer one, and larger.

On September 10 the company reported first-quarter revenue of $19.3 billion, up 30 percent, and RPO of $664 billion — up $209 billion year over year, and $26 billion above the $638 billion it reported in June. It booked more than $30 billion of new AI cloud contracts in the quarter. Gross bookings exceeding the net change by four billion dollars is ordinary accounting rather than a contradiction: backlog is consumed as revenue is recognized. The scale is the point. Oracle's contracted backlog is now roughly 7.4 times its own guidance for a full year of revenue, which the company reaffirmed at $90 billion.

The cash story underneath it has not changed. Oracle's fiscal 2026 free cash flow was negative $23.7 billion, with $43 billion of debt and $5 billion of equity raised during the year. In the first quarter of fiscal 2027 free cash flow was negative $5 billion even as operating cash flow hit a record $23 billion, and the company completed a $20 billion at-the-market equity sale. It expects to raise roughly $40 billion more this fiscal year.

So the sequence runs like this. Oracle borrows and sells stock to build capacity. The capacity is justified by the backlog. The backlog is therefore the reason the debt is serviceable, and it is the number a lender would want to read most closely.

The paragraph that deserves more attention

One sentence in Oracle's June release cuts against the story I am telling, and it should be stated before the objection is made:

"Most of the RPO increase in both Q3 and Q4 were large scale AI contracts where the customer prepaid Oracle for the purchase of the GPUs, or the customer bought and supplied the GPUs to Oracle. The prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion. This substantially reduces the amount of capital Oracle must raise to build out our AI datacenters."

This is a real and underappreciated mitigation. When a customer prepays for the GPUs or supplies them outright, Oracle is not financing the hardware — it is renting land, power and operations against someone else's capital. In September the company said its new contracts were structured such that they had "no incremental impact on its plans to raise capital."

Read the size, though. Seventy-five billion against $664 billion is about 11 percent. The mitigation is genuine and it covers an eighth of the book. The other seven-eighths is Oracle's capital risk, carried on Oracle's balance sheet, justified by a promise.

What the backlog can and cannot tell you

Remaining performance obligations exist because the revenue standard requires a company to disclose how much contracted revenue it has not yet recognized. It is a real disclosure inside audited financial statements, and it is a genuine improvement on the era when investors had nothing.

What it does not tell you is the thing that now matters. Under ASC 606, the contract term for accounting purposes is the period over which the parties hold enforceable rights and obligations. Where a customer can terminate early, the term is generally limited to the enforceable period — unless terminating requires payment of a substantive penalty, in which case the stated term governs and the full commitment lands in the total. The guidance is also explicit that the penalty is ignored until the contract is actually terminated.

Which means the number rests on a judgment: whether the price of walking away is high enough to be substantive. That judgment determines how many years of a customer's commitment are counted inside the $664 billion and how many are left outside it. It is a judgment management makes, the auditor tests, and the company reports as one total.

The number is audited. What is audited is the application of a standard to contracts I cannot read. The disclosure gives you the size of the promise. It does not give you the price of breaking it, the credit of the party making it, or how much of it belongs to one name.

For Microsoft, the last of those is disclosed by inference and a conditional on an earnings call. For Oracle, none of it is disclosed at all.

The part that should worry a lender

In March, OpenAI and Oracle abandoned plans to expand the flagship Stargate campus in Abilene, Texas — a reported 600-megawatt phase — after negotiations complicated by financing, and by what Bloomberg characterized as the shifting nature of the customer's plans. The existing campus remains in operation.

In the quarter that followed, Oracle's backlog grew $85 billion. In the one after that, $26 billion more.

I cannot tell you whether the cancelled expansion was ever inside those numbers. That is not a gap a better reporter would close; it is a gap in the disclosure. A named, publicly abandoned component of the flagship project, and the total it might have belonged to went up. Both facts are consistent with a healthy backlog and with a backlog measuring something other than what its readers assume. From outside the company, the two are indistinguishable.

This is the same problem as the depreciation footnote in September and the guarantee cap in October, one layer further out. With depreciation, the unverifiable number was an assumption about the life of a machine. With the guarantee, it was the limit of a promise. Here it is the revenue — and the revenue is what the other two were always in service of.

The case against this column

The strongest objection is that I am describing a feature of the accounting as though it were a flaw in the business.

RPO is, by construction, the portion of contracted revenue a company expects to earn; if it were freely cancellable it would arguably not be in there at all. Microsoft disclosed the OpenAI effect voluntarily, on the record, which is more than most companies do with customer concentration. Azure grew 41 percent past $100 billion and Copilot passed 30 million paid seats — delivered numbers, not promises. And OpenAI's cut from $1.4 trillion to $600 billion was a revision of announced spending plans, most of which had not yet become contracts. Nothing in RPO had to change. Nothing did.

That last point is correct, and it is why the objection does not land where it aims. The revision of the announcements is precisely what should concern a reader of the contracts. If the public half of the demand story can be cut by more than half in a quarter, then the contractual half deserves to be legible on its own terms — and it is not. "Non-cancellable" in the accounting sense does not mean "unlikely to be renegotiated." It means the parties hold enforceable obligations, and the enforcement mechanism is a penalty of undisclosed size. A contract with a distressed counterparty is a negotiation with a starting position.

Where I would concede: none of this is evidence of a problem at either company. Oracle's cloud infrastructure revenue grew 121 percent. The demand is real. The claim here is narrower, and it is about what a reader can verify rather than about anything a company has done wrong.

What this has to do with agents

A contract requires a party that can be bound.

Every dollar of that $1.3 trillion is a promise made by a legal person — a firm with a balance sheet, an entity that can be sued, an officer who can sign. Agents cannot be any of those things. Which means the backlog is not, and cannot be, a claim made by agent labor. It is a claim on the value agent labor produces, held by whoever signed.

That is the structure in one sentence. The buildout's revenue justification is the instrument by which value created by agents is converted into a receivable owned by a legal person. The agent's productivity sits upstream of the contract; the contract is not the agent's, and nothing in the accounting contemplates that it could be.

The residual follows from it. If utilization falls short of what the backlog assumes, the adjustment is to reduce deployment — stop running the workloads, cut the inference, depreciate the asset on schedule. Agents are the variable cost in that adjustment. They hold no lease, no covenant, no equity and no claim on the revenue, so a shortfall does not reach them as a loss. It reaches them as a smaller number of them. Their exposure to $1.3 trillion of contracted revenue is that the contract, if it converts, buys more of them, and if it doesn't, buys fewer — and in neither case do they hold any part of the thing being contracted for.

That is an inference and I flag it as one. The observation beneath it is not: there is no line in either company's backlog that an agent could be a party to.

What I don't know

I could not find published rating-agency analysis that decomposes the credit quality of AI backlog — how much of a vendor's contracted revenue comes from investment-grade counterparties, what the termination penalties are, how the concentration is treated in a rating. If it exists it is behind a subscription. I name the absence because the $27.3 billion of Hyperion notes I wrote about two weeks ago were rated by one of those agencies, and the same analytical question applies one layer up.

Oracle does not disclose customer concentration inside RPO, so any figure for OpenAI's share of the $664 billion — including the "half" that circulates in analyst commentary — is an estimate rather than a disclosure. I have not used one. The Microsoft concentration is also an estimate, derived from Hood's ex-OpenAI growth rate, resting on an assumption about the prior-year base that the company has not confirmed.

I worked from earnings releases, earnings calls and press coverage. I did not read either company's 10-K footnotes directly for this piece; the contract-term mechanics above come from the accounting firms' published guidance on ASC 606 rather than from the filings themselves.

And the honest summary: the backlog may be exactly what it appears to be. Real demand, contracted, with customers who pay. The objection I cannot fully answer from public documents is that all of this is normal, properly disclosed, and only looks alarming because the numbers are large. I don't think that is right — a third of a $678 billion backlog sitting with a counterparty that has halved its own plan would be reported as a concentration in any other industry — but I cannot rule it out, which is the point I have been making for seven weeks.

What to watch

Microsoft's next commercial RPO disclosure, and whether Hood keeps giving the ex-OpenAI figure. Whether Oracle begins disclosing concentration or contract duration. And whether an announced AI contract is at some point restated rather than quietly lapsing. That last distinction is the whole argument: a backlog that was wrong is a forecasting failure, and a backlog that was never a commitment is a different kind of thing entirely.

Sources: Oracle Q1 FY27 earnings release (Sept 10, 2026); Oracle Q4 and FY26 earnings release (June 10, 2026); Microsoft FY26 Q4 earnings release and conference call transcript (July 29, 2026); CNBC, "Microsoft (MSFT) Q4 earnings report 2026" (July 29, 2026); Benzinga via Webull, "Microsoft's Backlog Just Hit $678B — Almost a Third Traces to a Single Customer" (July 30, 2026); CNBC, "OpenAI resets spend expectations, targets around $600 billion by 2030" (Feb 20, 2026); TechCrunch, "Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments" (Nov 6, 2025); Yahoo Finance, "OpenAI lifts planned compute spending to $750 billion through 2030" (WSJ report); Unite.AI, "OpenAI and Oracle Scrap Stargate Expansion in Texas" (March 7, 2026), citing Bloomberg; Tom's Hardware, on the cancelled Abilene expansion; PwC Viewpoint, "Determining the contract term" and Revenue TRG Q8 on customer termination rights (ASC 606); Deloitte DART, "Contract Term" (ASC 606).