The duration mismatch at the heart of the AI build

The question, and the one the filings actually answer

Ask who wins the AI datacenter era and you will usually get an answer about market share. The filings point somewhere else. Every one of these companies is making the same trade — signing long-dated obligations to buy capacity, against revenue contracted for a much shorter term — and what separates them is not who has the best chips or the biggest cloud. It is how far apart those two durations are, and what happens to each balance sheet if the gap does not close.

Two of them say so themselves, in their own risk factors, which is the most useful thing in this entire subject.

Microsoft states both sides of the gap

Microsoft's FY2026 accounts show what a software company looks like after it has bought an industrial base. Additions to property and equipment were $115.9 billion in fiscal 2026 against $64.6 billion and $44.5 billion in the two prior years ^[1]. Net property and equipment rose to $313.1 billion from $205.0 billion in a single year, and the depreciation that follows it went to $34.3 billion from $22.0 billion ^[1].

The commitments behind that are larger than the balance sheet shows. Microsoft reports $743.8 billion of total contractual obligations at June 30 2026, of which $194.1 billion sits in purchase commitments that the filing's own footnote says "primarily relate to datacenters" ^[1]. Separately, it discloses $329.1 billion of leases "primarily for datacenters" that had not yet commenced, scheduled to begin between fiscal 2027 and fiscal 2033 on terms of one to twenty years ^[1].

Now the other side. Commercial remaining performance obligation — the contracted revenue — was $678 billion, up 84%, "with a weighted average duration of approximately 2.3 years" ^[1]. Microsoft expects to recognise roughly 30% of it within twelve months ^[1].

That is the mismatch in one company's own numbers: a lease and construction book running to 2033 and beyond, funded against a revenue book whose average life is a little over two years. The filing does not dress it up. Its risk factors say the investments "are in advance of fully developed revenue streams" and that the associated revenue "may not be realized in the expected timeframes or at expected levels" ^[1].

For now the cash engine covers it. Operating cash flow was $182.9 billion in fiscal 2026 ^[1], comfortably more than the capex. The number to watch is not the capex line but the distance between those two.

Oracle is the same trade at higher gearing

Oracle's version is more dramatic in both directions. Its remaining performance obligations were "$638 billion and $138 billion as of May 31, 2026 and 2025, respectively" ^[2] — a book that grew more than fourfold in a year, against fiscal 2026 revenue of $67.4 billion ^[2].

A backlog of that size relative to that revenue base is either the best position in this industry or the most levered, and Oracle's own risk disclosure explains why the distinction is hard to make from outside. To grow its infrastructure business, the company says, it enters "long-term lease commitments with third-party data center providers and other significant commitments with suppliers of chips and other data center infrastructure" ^[2]. And then, plainly:

if we overestimate customer demand or any of our key customers are unable to pay or otherwise perform under their contracts with us, we could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue ^[2].

The sentence that follows is the one that matters most, because it is the general case rather than a worst case: "the terms, renewal options and pricing adjustments in our long-term data center leases typically do not align with the duration and pricing of customer contracts" ^[2].

Oracle is describing the duration mismatch as a standing feature of the business, not a tail event. What makes it sharper at Oracle than at Microsoft is the base it is carried on: Oracle's underlying franchise is durable — the filing notes that "substantially all customers opt to purchase software support contracts when they purchase an Oracle software license" ^[2] — but the cash generation supporting the build is a fraction of Microsoft's, at $32.0 billion of operating cash flow in fiscal 2026 against total liabilities of $218.7 billion ^[2].

Amazon shows what the build costs a company that pays cash

Amazon is the incumbent with the largest installed base, and its 2025 accounts show the price of staying there. Purchases of property and equipment were $131.8 billion in 2025, against $83.0 billion in 2024 and $52.7 billion in 2023 ^[3]. Operating cash flow rose to $139.5 billion ^[3].

Put those together as Amazon itself does, and the result is the single most arresting figure in this subject: free cash flow fell to $11.2 billion in 2025 from $38.2 billion in 2024 ^[3]. AWS revenue grew 20% to $128.7 billion over the same period ^[3].

That is not a company in trouble. It is a company converting nearly all of its cash generation into plant, in a year its cloud segment grew a fifth. Amazon is running the same trade as the other two; it is simply paying for more of it up front, which is why its mismatch shows up in the cash flow statement rather than in an off-balance-sheet lease schedule.

Alphabet is running it most conservatively

Alphabet's capex went to $91.4 billion in 2025 from $52.5 billion in 2024 and $32.3 billion in 2023 ^[4], on revenue of $402.8 billion, up 15% ^[4], and operating cash flow of $164.7 billion ^[4]. Total liabilities were $180.0 billion against $125.2 billion a year earlier ^[4].

Of the four, that is the smallest gap between what is being spent and what is being generated, on the largest equity base. Alphabet is also the only one of the four that designs and deploys its own training accelerator at scale, which changes what a marginal dollar of capex buys — though the size of that advantage is not something the filings quantify, and this article will not estimate it.

The neoclouds are the same trade, much further out

The pure-play AI clouds make the structure visible because they have nothing else on the balance sheet to hide it.

CoreWeave's 2025 accounts show revenue of $5.13 billion, up from $1.92 billion in 2024, against total liabilities of $45.97 billion, up from $16.53 billion ^[5]. Revenue grew about 2.7 times; liabilities grew about 2.8 times. Those two ratios are arithmetic on the filed figures rather than anything the filing states, but the shape is the point: this is a business whose obligations scale with its contracts, because the contracts are what the obligations are raised against.

What those obligations look like in practice is on the other side of the same trade. Applied Digital has entered "two approximately 15-year lease agreements" with CoreWeave to deliver 250 megawatts of critical IT load at its Ellendale, North Dakota campus, from which it "anticipates generating approximately $7 billion in total revenue" over the term, with CoreWeave holding an option on a further 150 MW ^[7].

Fifteen years, one counterparty. That is the neocloud model stated exactly: long-dated physical commitments underwritten by customer contracts, where the credit of the customer is the asset. It is the same duration mismatch the hyperscalers carry, with two differences that both cut the same way — the terms are longer, and there is no software franchise generating cash beside it.

The constraint is power, and it is contracted in decades

The reason none of this resolves quickly is that the binding input is no longer chips. It is power, and power is procured on timescales that make a two-year revenue book look very short.

Constellation signed a twenty-year power purchase agreement with Microsoft to restart Three Mile Island Unit 1 as the Crane Clean Energy Center, adding approximately 835 megawatts of carbon-free generation to the grid ^[6]. The plant was expected online in 2028 ^[6] — four years after the agreement was announced.

Twenty years, and four years before the first electron. That is the clock the whole industry is actually on, and it explains why the lease books look the way they do: you cannot buy this capacity in the spot market when demand shows up.

So who wins?

The measurements above are the article's; what follows is judgment, and the filings do not state it.

On the record as it stands, the ranking is by how survivable each company's mismatch is rather than by how large its ambition is. Alphabet and Microsoft look best positioned — both fund the build from operating cash flow with room to spare, and both have a franchise that keeps paying if AI demand grows more slowly than the lease schedules assume. Amazon is close behind, with the largest installed base and the most cash already committed; its free cash flow compression is a choice it can reverse faster than a lease can be unwound.

Oracle is the highest-variance position in the group. A backlog that grew from $138 billion to $638 billion in one year is either the best-timed commercial move in enterprise software or a concentration of counterparty risk, and the company's own risk factor declines to tell you which. The neoclouds are the same bet with the safety net removed — they win disproportionately if demand holds, and they have no second business to retreat into if it does not.

But the honest answer to the question is that the biggest winner may not be on this list. Every one of these five buys from the same suppliers and contracts for the same power, and the firms selling into all of them — the accelerator vendors, the interconnect makers, the independent generators — collect from each of these balance sheets regardless of which cloud eventually holds the customer.

What would change this view

• Microsoft's commercial RPO duration extending materially beyond 2.3 years, which would close the gap this article is built on from the revenue side rather than the cost side.
• Oracle converting a visible share of the $638 billion backlog into recognised revenue on the schedule its leases assume, with the customer concentration disclosed.
• Amazon's free cash flow recovering while capex stays elevated, which would show the build reaching a plateau rather than still climbing.
• Any of the neoclouds losing or repricing a large contract — the event their entire structure is underwritten against.