$329.1bn of Leases That Haven't Started (September 2026): Microsoft's 38-Gigawatt Plan Was in the Footnotes Six Weeks Before It Leaked

Bloomberg reported on 10 September that Microsoft intends to more than triple its datacentre footprint to over 38 gigawatts by 2032, from about 12 today. The number is unconfirmed and sourced to people familiar with the plans. But the company's own annual report, filed six weeks earlier, already disclosed $329.1 billion of datacentre leases signed and not yet commenced, scheduled to begin between fiscal 2027 and fiscal 2033. The leak and the footnote describe the same build. Only one of them is audited.

The leak and the footnote describe the same build

Bloomberg's report, relayed by Reuters on 10 September 2026, put Microsoft on a path to more than 38 gigawatts of datacentre capacity by 2032 against about 12 gigawatts today. Roughly 2GW of the current estate is given over to AI-specific chips; on the reported roadmap that becomes about 13GW of the eventual 38GW. The figure covers owned and leased facilities and excludes capacity rented from neocloud providers. Microsoft did not respond to a request for comment, and the people cited cautioned that the roadmap could change, because datacentres take years to build and customer demand can move.

Treat that as a report, because that is what it is. Then open the annual report Microsoft filed on 29 July 2026 and read the last sentence of its lease note.

As of June 30, 2026, we had additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion, with some arrangements subject to certain contractual conditions being met. These leases will commence between fiscal year 2027 and fiscal year 2033.

Fiscal 2027 to fiscal 2033 is the same window as 2032. The two numbers are not independent corroboration of one another so much as one commitment viewed through two units — dollars of contracted rent in the filing, gigawatts of delivered capacity in the report.

The number that grew 3.6x in twelve months

The disclosure is a single line, and the year-on-year movement is the point.

*Fiscal 2024 disclosed operating and finance amounts separately, $8.6bn and $108.4bn; 2025 and 2026 disclose a single combined figure. Source: Microsoft fiscal 2024, 2025 and 2026 annual reports.

Two rows deserve to be read against each other. Operating cash flow grew 34% in fiscal 2026 to $182.9bn, an exceptional result by any standard. Free cash flow, on the simplest definition, fell for the third consecutive year. Everything the business generated, and more, went into the ground. On the fourth-quarter call Hood said she expects Microsoft to be free-cash-flow positive in fiscal 2027, which is a statement about the sign rather than the size.

Two smaller disclosures in the same filing describe the pace. Purchases of property and equipment still sitting in accounts payable at year-end were $26.7bn, against $6.9bn a year earlier — a measure of how much work was done but unpaid on 30 June. And $34.6bn was committed for construction of new buildings and improvements, primarily datacentres, on top of everything already leased.

Why gigawatts became the honest unit

There is a specific reason to prefer a physical unit here, and it is not aesthetic. The accounting boundary around capital expenditure is moving in two directions at once, and both of them shrink the reported number without shrinking the build.

Step three is where the definition bends. A finance lease puts the asset on the books and splits the cost into amortisation of the right-of-use asset and interest — in fiscal 2026, $5.4bn and $2.5bn respectively. An operating lease produces a single rent charge, $7.0bn for the year, and sits outside the capital spending measure altogether. Hood told analysts on 29 July that more future datacentre leases would be classified as operating rather than finance. The capacity is identical. The line it lands on is not.

Alongside that, Hood said Microsoft would extend the estimated useful life of office and datacentre buildings to 25 years from 15. The fiscal 2026 filing still carries the old policy — buildings and improvements at five to 15 years, servers and network equipment at two to six, internal-use software at three — because the change applies from fiscal 2027.

The arithmetic is worth doing once, as a scale check rather than a forecast. Buildings and improvements stood at $182.7bn at cost on 30 June 2026. Straight-line over 15 years is about $12.2bn a year; over 25 years, about $7.3bn. The difference is near $4.9bn annually, roughly 3% of fiscal 2026 operating income of $155.2bn and about 14% of the year's depreciation charge. The realised effect will be smaller and messier, because the change applies prospectively to assets already part-depreciated. But the direction is unambiguous, and it arrives in the same year the asset base is set to grow fastest.

None of this is unusual or improper. Buildings genuinely do last longer than servers, and Microsoft is far from the first company to revisit an estimate as its estate changes shape. The consequence for a reader is narrower and more useful. If the definition of the dollar figure is in motion, year-on-year comparisons of capex carry less information than they used to — and a gigawatt is a gigawatt in 2026 and in 2032.

What 26 gigawatts of new load actually asks for

The reported plan implies adding roughly 26GW over six years, an average of more than 4GW a year. Some arithmetic to anchor the scale.

Run flat out for a year, 38GW would consume about 333 terawatt-hours. Total retail electricity sales to all end-use customers in the United States were 4,058 billion kilowatt-hours in 2025, according to the Energy Information Administration, with net utility-scale generation of 4,429 billion kWh. So the ceiling case is around 8% of everything sold in the US last year — for one company, across a global footprint, at an unrealistic 100% utilisation. The honest reading is not that Microsoft will consume 8% of American electricity. It is that a single corporate plan is now large enough that the comparison is worth making at all.

That load has to be connected before it can be filled, and the constraint stack behind it — generation, interconnection queues, and the transformer and turbine supply chains — is the part that does not respond to a purchase order. We looked at the equipment end of that problem in the piece on the bulk power system order and transformer lead times. It is also why capacity and capital have stopped being interchangeable words in this sector.

Where this touches an instrument

Microsoft is among the largest constituents of both the S&P 500 and the Nasdaq-100, so US500/ES and NAS100/NQ carry the exposure directly, diluted by index weight. The second-order channel runs through suppliers — the power equipment, construction and semiconductor names whose order books this build populates — which is why Reuters noted Nvidia shares ticking higher after the Bloomberg report while Microsoft's own move stayed under a percent.

The market's reaction was muted, and the thesis of this piece explains why. The information was not new. It was filed on 29 July, in a footnote, denominated in dollars instead of watts.

The wider structural point is that AI infrastructure commitments increasingly sit in places a balance sheet does not show: leases before commencement here, residual value guarantees elsewhere. We took apart one of the other structures in the piece on Nvidia's guarantee disclosures, and the pattern of capital outrunning cash across the same cohort in Oracle's capex-against-cash-flow gap.

In currencies, the honest answer is close to nothing. What moves the dollar is rates, growth, positioning, risk appetite and commodities — the five factors scored here across the eight majors. There is a slower aggregate channel worth naming, though. If hyperscaler capital spending is financed increasingly with debt rather than operating cash flow, that adds corporate issuance competing for the same pool of savings, and the long end of the curve is a genuine dollar input. That is a story about the cohort's funding mix over years, not about one lease footnote. How we keep those channels separate is set out on the about page.

What would change the picture

Four checkable things, in order of how much they would tell you.

The fiscal 2027 first-quarter filing. Hood guided to roughly $50bn of capital spending in the September quarter. That document is also the first to apply the 25-year building life, so the accounting policy note will confirm whether the change landed as described, and the depreciation line will show its early effect.

The next annual disclosure of leases not yet commenced. A single number, once a year, and the best available readout of whether the 38GW path is being contracted or slowed. Watch the commencement window as well as the dollars — if the range extends past fiscal 2033, the build is being stretched rather than shrunk.

The split between finance and operating leases. If the operating lease liability starts growing at the rate the finance lease liability has been growing, the reclassification Hood described is real, and the headline capex figure will understate the build by more each year.

Free cash flow. It has fallen for three years while operating cash flow grew 54%. Hood expects it to stay positive in fiscal 2027. That is the line where the question of whether this is funded from earnings or from the balance sheet gets answered, one quarter at a time.

None of that is a view on the company or its shares. It is the list of documents that turn an anonymous gigawatt figure into something checkable — and the useful observation is that most of the checking material was published before the number leaked.

Educational macro context only — not investment advice.