Microsoft Already Told You It Builds a Gigawatt a Quarter. The 26-Gigawatt Leak Is Six More Years of It.

A leaked capacity target invites one obvious question, and that question has nowhere to go. Is the number real? Nobody outside Microsoft can audit a plan for 2032, and the reporters who broke it said themselves that the road map could change. Asking whether a six-year target is true is asking a company that has not commented to grade its own homework.

The answerable question is narrower. Of everything in the report, which parts had Microsoft already told us, and which part is genuinely new? That one can be settled against documents, because a capacity plan has three components — a rate, a level and a price — and Microsoft has been publishing two of them out loud for a year.

Abstract

On 10 September 2026, Bloomberg reported that Microsoft's data-centre network will hold more than 38 gigawatts of capacity in 2032, up from about 12 gigawatts now: roughly 26 GW added over six years, owned and leased, excluding capacity rented from neoclouds such as CoreWeave. Microsoft has not commented on the record.

Two of those three numbers are already public, from Microsoft, on the record. On its last three earnings calls Satya Nadella said the company had added about a gigawatt of capacity in each quarter — the December, March and June quarters in a row. That is a rate of 4 GW a year. Bloomberg's 26 GW spread over six years is 4.33 GW a year. The leaked plan runs 8% above the pace Microsoft has already been announcing every ninety days. "Tripling" is what happens when you add a constant to a small base for six years.

The price is public too, and it comes from a single call. On 29 July 2026 Amy Hood said capital expenditures in the June quarter were $41 billion, of which $35.8 billion was cash paid for property and equipment and $5.6 billion was finance leases. Thirty minutes earlier Nadella said the company added about a gigawatt that quarter. Divide one by the other and Microsoft turns on a gigawatt for roughly $41 billion. Run that across 26 GW and the six-year bill is about $1.07 trillion — against $1.05 trillion, which is six years of the $175 billion calendar-2026 capex Microsoft has guided to. The two agree within 2%.

What Microsoft has never published is the level. Its fiscal 2026 Form 10-K contains the word "megawatt" zero times, the word "gigawatt" zero times and the word "watt" zero times. Item 2 gives 112 million square feet across all facilities, offices and datacenters together, split 58 owned and 54 leased, and stops there. The gigawatts live only in the spoken transcript of the calls, and only ever as additions. The leak's contribution is the denominator — and Microsoft's own two published rates already implied it. A fleet that doubles in two years while adding a gigawatt a quarter must have started at 8 GW; add four more quarters and you arrive at about 12.

The forward money is disclosed as well, in the one place nobody reads. Note 13 of the 10-K records $329.1 billion of leases, primarily for datacenters, signed but not yet commenced at 30 June 2026, commencing between fiscal 2027 and fiscal 2033. That window is the leak's window. Microsoft filed the seven-year build in June and denominated it in dollars.

Table of contents

1. Microsoft has published a gigawatt a quarter for three quarters running
2. Solve Microsoft's own two rates and you get 12 gigawatts
3. A Microsoft gigawatt costs about $41 billion, and both halves came from the same call
4. The $15 billion that left the capex line without leaving the build
5. $329.1 billion of leases that have not started, for exactly these years
6. A third of the capacity takes two thirds of the money
7. The 12 gigawatts exclude what Microsoft rents, and two contracts alone are half a gigawatt
8. Flat gigawatts and rising dollars pull apart after 2028
9. One line in Item 2 would end all of this
10. What follows if these numbers hold
11. Primary sources worth reading

1. Microsoft has published a gigawatt a quarter for three quarters running

The premise most commentary is running on — that Microsoft never talks about capacity — is half wrong, and the half that is wrong is the interesting half.

Microsoft does not put capacity in its filings. It puts capacity in its earnings calls, every quarter, in the same sentence construction, and has done so for a year:

• 28 January 2026, Nadella: "All up, we added nearly one gigawatt of total capacity this quarter alone."
• 29 April 2026, Nadella: "All up, we added another gigawatt of capacity this quarter, and remain on track to double our overall footprint in just two years."
• 29 July 2026, Nadella: "All up, we added another gigawatt of capacity this quarter and remain on track to roughly double our overall capacity in just two years."

Three consecutive quarters, about a gigawatt each. Nine months, three gigawatts, a run rate of 4 GW a year.

Go back one more quarter and the frame appears. On 29 October 2025 Nadella said Microsoft would "increase our total AI capacity by over 80% this year, and roughly double our total datacenter footprint over the next two years", and that Fairwater in Wisconsin "will go online next year and scale to two gigawatts alone."

So Microsoft publishes three kinds of capacity number: quarterly additions, growth rates, and the size of one named site. It has never published the fourth kind, which is the total.

The table at the top of this page lists every one of them against the document it appeared in. That distinction is not pedantry, and it is the reason a gigawatt target for Microsoft reads as news when the same target from CoreWeave would not. A company that reports contracted power reports a level, so every ratio anyone computes from it is checkable. A company that reports only additions hands you the change and keeps the total.

2. Solve Microsoft's own two rates and you get 12 gigawatts

Two disclosed rates are enough to recover the level, because they constrain each other.

Take them in order. Microsoft adds about 1 GW a quarter (disclosed, three quarters running). Microsoft expects to roughly double the fleet in two years (disclosed, October 2025, reaffirmed in April and July 2026).

Two years at a gigawatt a quarter is 8 GW added. A fleet that ends at twice where it started, having added 8 GW, started at 8 GW: if the base is B, then B + 8 = 2B, so B = 8.

That fixes the base at 8 GW in October 2025, when Nadella first set the two-year clock. Add the four quarters from then to autumn 2026, at the same gigawatt a quarter, and Microsoft's fleet reaches about 12 GW — which is what Bloomberg reports the fleet is now.

The step that carries the weight here is the two-year window, and it deserves its own line. Nadella opened it in October 2025 and repeated the phrase "in just two years" in April and July 2026 without restating an end date. Read as a fixed window running to October 2027, the steps above hold and land on 12 GW. Read as a window that restarts each time he says it, the same gigawatt a quarter implies a base of 8 GW today, and Bloomberg's 12 GW would be 50% too high. There is a third possibility: the base really is 12 GW and the doubling really is sliding, in which case Microsoft has to reach 1.5 GW a quarter, half again its reported rate, and has not yet reported a quarter at that pace.

Nothing here proves which reading is right. What it establishes is that the leak's headline figure is not an outside revelation about a company that discloses nothing — it sits inside the envelope Microsoft's own statements already defined, at the top of it.

3. A Microsoft gigawatt costs about $41 billion, and both halves came from the same call

The most useful number in this story requires no leak at all. It is a division of two figures Microsoft published on the same morning.

Start with what Hood said on 29 July about the June 2026 quarter. Capital expenditures were $41 billion. Cash paid for property and equipment was $35.8 billion — the figure that also appears in the results release as additions to property and equipment, at $35,802 million. Total finance leases were $5.6 billion and were "primarily for large data center sites." Those two add to $41.4 billion, which is the $41 billion headline.

Now the other half, from Nadella on the same call: about 1 GW of capacity added in that quarter.

$41.4 billion ÷ 1 GW ≈ $41 billion per gigawatt.

Two things about that number. It is blended, covering AI and non-AI capacity together, so it is not comparable to a pure AI-factory figure. And it is a same-quarter division, which assumes the money spent in a quarter is roughly the money behind the gigawatt that lit up in it — true only if the pipeline is in steady state.

That second assumption is the weakest link in everything above, so take the objection seriously. GPUs bought today go live six to twelve months later. Divide by the wrong vintage and the answer moves a long way. Do it on the full fiscal year instead: Microsoft's FY2026 additions to property and equipment were $115.9 billion and it obtained $24.6 billion of finance-lease assets, a total of $140.6 billion against roughly 4 GW added, or $35 billion a gigawatt. Lag it a year and use FY2025 spending — $64.6 billion of additions plus $20.5 billion of finance leases, $85.1 billion — against the same 4 GW, and you get $21 billion a gigawatt.

So the honest span is $21 billion to $41 billion per gigawatt, and 26 GW costs somewhere between $553 billion and $1.07 trillion. That is a factor of two, and pretending otherwise would be false precision.

The span narrows against Microsoft's own forward guidance, which is the point of quoting it. Six years at the guided $175 billion of calendar-2026 capex is $1.05 trillion. At $41 billion a gigawatt that buys 26 GW, which is the leaked plan almost exactly. At $21 billion a gigawatt the same money buys 50 GW, nearly double the plan. Microsoft is either guiding to far more money than its target needs, or the cost of a gigawatt is climbing. Section 8 argues it is climbing.

4. The $15 billion that left the capex line without leaving the build

Microsoft cut its calendar-2026 capex expectation from about $190 billion to about $175 billion on that July call, and the cut was widely read as spending discipline. It is an accounting reclassification, and Hood said so in the sentence that announced it.

The mechanism has two steps, and the order matters because most summaries collapse them.

Step one is depreciation. "Effective at the start of FY27, we are extending the estimated useful lives of our datacenters and office buildings, from 15 to 25 years." Microsoft's fiscal 2026 10-K still carries the old policy — buildings and improvements, five to 15 years; servers and network equipment, two to six years. Stretching a building's assumed life lowers annual depreciation expense. Depreciation is an income-statement line. It does not touch capital expenditure at all, in either direction.

Step two is lease classification, and this is the one that moves capex. Hood: "The greater impact is on capital expenditures, as more of our future data center leases will shift from finance leases to operating leases as a result of this update." The reason a life assumption controls how a lease is classified is a rule in the leasing standard: a lease counts as a finance lease when its term covers the major part of the asset's remaining economic life. Reassume the life at 25 years rather than 15 and a 15-year datacenter lease drops from the whole of that life to about 60% of it, and lands on the operating side. Hood put the consequence in one line: finance leases are included in capital expenditures, and operating leases are not. Neither has ever touched the cash line, which counts only what Microsoft actually paid for property and equipment.

Hood closed the loop herself: "Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged." The same buildings, the same power, the same schedule — $15 billion less on the reported line.

Our own July analysis of that quarter described this as Microsoft "spreading long-term data-center leases over 25 years instead of 15" and stretching a depreciation schedule to lower capex. The conclusion it drew was right and the mechanism was not: depreciation schedules never lower capex, and what did lower it was the reclassification the life change triggered. The distinction matters for anyone modelling forward years, because a reclassification moves the money permanently off one line and onto another, while a depreciation change would have left the capex line alone and flattered earnings instead.

There is a real signal buried in the accounting. Finance-lease additions have been the fastest-growing part of Microsoft's capital spending — $11.6 billion in FY2024, $20.5 billion in FY2025, $24.6 billion in FY2026 — and it is precisely that growing line the reclassification pushes out of view from fiscal 2027 onward.

For a reader tracking hyperscaler capex as a proxy for build intensity, this is a break in the series. From fiscal 2027 Microsoft's reported capex and Microsoft's actual build stop moving together, and the gap will widen with every lease signed. The number that stays comparable is the cash line, additions to property and equipment, which no lease classification can touch.

5. $329.1 billion of leases that have not started, for exactly these years

The single most under-quoted figure in this story is in Note 13 of the fiscal 2026 10-K:

"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 with lease terms of 1 year to 20 years."

Fiscal 2027 begins July 2026. Fiscal 2033 ends June 2033. That is the leak's build window, filed six weeks before the leak, with a dollar figure attached.

Set it against the leases Microsoft has already commenced. Total remaining lease payments at 30 June 2026 were $24.7 billion on operating leases and $89.7 billion on finance leases, $114.4 billion in all. The not-yet-started book is 2.9 times everything currently running.

Against the six-year bill from section 3, $329.1 billion is about 31% of $1.07 trillion. Spread evenly it is $47 billion a year against a $175 billion annual capex guide, or roughly 27%. Both readings say the same thing: leasing is a large minority of how this build gets done, the rest is owned construction and silicon, and Microsoft has already committed the leased share in a filed document.

The strongest objection to everything above runs through this same figure, so it is worth putting in its strongest form. Someone could reasonably say that a 2032 capacity target is a genuine six-year commitment and nothing Microsoft publishes goes out that far, since capex guidance covers one calendar year. The $329.1 billion answers it. Microsoft's forward commitment already runs to fiscal 2033. It is denominated in dollars and buried in a lease note, which is why it did not make headlines, but the horizon was never the thing that was missing.

6. A third of the capacity takes two thirds of the money

The tempting move on a report like this is to take 38 GW, multiply by a per-gigawatt cost from somewhere, and print the answer. It gives the wrong number by a wide margin, and the reason is a denominator switch that is worth carrying to every other company in this sector.

Bloomberg reports that about a third of the 38 GW is centred on AI-specific chips — roughly 12.7 GW of AI capacity in 2032. Hood, on the same July call, said "roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs."

A third of the capacity, two thirds of the money. Both figures can be true at once, because a watt of AI capacity costs several times what a watt of general cloud capacity costs. Silicon density is the difference: an AI hall packs far more dollars of chip behind each megawatt of power than a hall of storage and general-purpose compute.

Split the $41.4 billion quarter on Hood's ratio and the gigawatt Microsoft turned on that quarter cost about $27.6 billion in silicon and about $13.8 billion in shell, land, power and cooling.

Now bring in the only per-gigawatt ladder any company in this sector publishes. On NVIDIA's fiscal Q2 2027 call, captured here in full, Jensen Huang gave the revenue opportunity per gigawatt by generation: $18 billion with Hopper, $25 billion with Blackwell, $40 billion with Vera Rubin. He also gave the number almost nobody quotes, which is the one that matters for pricing a whole data centre rather than the chips in it: "Each gigawatt of data center increased from, say, $30 billion about 5 years ago to now $60 billion today."

So a pure AI gigawatt today runs to about $60 billion all-in, of which NVIDIA's slice is up to $40 billion. Microsoft's blended fleet gigawatt costs $41 billion all-in, with $27.6 billion of silicon in it. Microsoft comes in below the AI-factory ladder because only about a third of what it builds is an AI factory. The numbers are consistent, and they are consistent in the direction that tells you the multiplication does not work: 38 GW × $40 billion is $1.5 trillion, and it is wrong, because it applies an AI-only chip figure to a fleet that is two-thirds not-AI and then counts none of the shell.

We priced the same trap when NVIDIA disclosed 12 gigawatts of OpenAI commitments and refused to multiply by the top rung. It is worth noticing that OpenAI's entire committed NVIDIA compute through 2030 — about 12 GW — is the same number as Microsoft's entire current fleet.

7. The 12 gigawatts exclude what Microsoft rents, and two contracts alone are half a gigawatt

The subtlest line in the report is the exclusion. The plan covers Microsoft-owned and leased data centres and leaves out computing power rented from neoclouds such as CoreWeave.

That makes 12 GW a floor on what Microsoft actually serves, not a measure of it. How much of a floor can be bounded from contracts that are already public, and we have two of them on file:

• IREN at Childress: five years, $9.7 billion, 200 MW of critical IT load.
• Nebius at Vineland, New Jersey: a reported $17.4 billion base value over five years, against a campus designed for about 300 MW.

Two contracts, half a gigawatt, about 4% on top of a 12 GW base. Those are two of several: Microsoft's reported neocloud contracting also covers CoreWeave and Nscale, so the true rented figure is higher, and nobody outside Microsoft can size it, because Microsoft does not disclose rented capacity any more than it discloses owned capacity.

The consequence for anyone building a hyperscaler capacity table is direct: Microsoft's 12 GW cannot be compared with a count that rolls neocloud capacity in, and it cannot be compared with a count of contracted power at a neocloud either. We have already shown what happens when per-megawatt figures on different bases get ranked against each other — the ordering was an artefact of the basis, not of the businesses.

Meta is the useful contrast, because a level exists for it. Its 14 GW by 2027 target also reached the public through a leaked internal memo rather than a filing, which is a fair reminder that Microsoft is not unusual here. The two numbers are not on the same basis and should not be set side by side: Meta builds almost entirely for its own products, a large part of Microsoft's fleet is sold to third parties, and our deep dive counted Meta's purchases from CoreWeave, Nebius, Crusoe and others as a separate line rather than folding them in.

8. Flat gigawatts and rising dollars pull apart after 2028

Two curves run through this story, and they are moving at different speeds.

The physical curve is flat. Microsoft has reported about a gigawatt a quarter for three quarters. The leaked plan needs 4.33 GW a year. Nothing in either number requires the pace of construction to change much.

The money curve is not flat. Microsoft's capital spending, cash plus finance leases, went $56.1 billion in FY2024, $85.1 billion in FY2025, $140.6 billion in FY2026, and the calendar-2026 guide is $175 billion. Cash additions alone rose 161% over two fiscal years. Microsoft's own capacity target over two years is to double, which is 100%. Money is growing faster than watts, and the difference between the two rates is price per watt.

Dollars per gigawatt are therefore rising, and NVIDIA's ladder says why: $18 billion, then $25 billion, then $40 billion of chip revenue per gigawatt across three generations, with Huang stating that after Vera Rubin "it is going to be higher." The same watt buys more compute each generation and costs more to fill.

That resolves the loose end from section 3. Six years of $175 billion buys 26 GW at $41 billion a gigawatt and 50 GW at $21 billion. If cost per gigawatt is climbing, the low end of that range describes the past, not the future, and the leaked 26 GW is what a trillion dollars buys at tomorrow's prices rather than yesterday's.

Where the two curves genuinely collide is 2028, and the collision is a test rather than a prediction. Under the fixed-window reading, Microsoft reaches 16 GW in late 2027 and needs 4.2 GW a year thereafter to hit 38 by 2032 — its current pace, held for five more years. Under the sliding-window reading it reaches 24 GW in 2028 and then needs only 3.3 GW a year, which is a slowdown of about a sixth from what it is doing now. The 2032 target does not distinguish between a company that keeps building at today's rate and one that sprints and then eases off. The quarterly gigawatt line does.

9. One line in Item 2 would end all of this

Everything derived above would be replaced by a single sentence in a filing.

Item 2 of the fiscal 2026 10-K already carries a table of square footage: 58 million square feet owned, 54 million leased, 112 million in total, U.S. and international, as of 30 June 2026. It does not separate datacenters from offices, and it carries no measure of power. Adding one row — total datacenter capacity in megawatts, at the fiscal year end — would make the level checkable, turn every additions figure into a percentage of something, and remove the need for every derivation above.

Until then, here is what would overturn the specific claims made here. If Microsoft's fiscal 2027 first-quarter call, due in late October 2026, reports capacity additions materially away from a gigawatt, the 4 GW-a-year rate and everything resting on it moves. A quarter at 1.5 GW would confirm the sliding-window reading and the 12 GW base; a quarter at half a gigawatt would say the run rate was a phase, and the 26 GW target would need an acceleration Microsoft has never demonstrated. Either way the answer arrives in six weeks, from Microsoft, on the record.

Two smaller things would also change the picture. If Microsoft confirms the 38 GW figure on the record, the hedging above about an uncommented press report drops away and the target becomes guidance. And if the next 10-K's not-yet-commenced lease figure falls rather than rises from $329.1 billion, the leased third of the build is being cancelled or converted, and the dollar path stops matching the capacity path.

10. What follows if these numbers hold

In descending order of confidence.

One. The reported capex line breaks as a comparison from fiscal 2027, and the cash line does not. This is the most certain item here because Microsoft stated the mechanism itself. Anyone comparing hyperscaler capex across 2027 should use additions to property and equipment, which no lease classification can move, and should treat Microsoft's headline capex as a number that fell $15 billion without a single building being cancelled.

Two. A Microsoft gigawatt costs on the order of $41 billion all-in today, blended across AI and non-AI, and about two thirds of that is silicon. Both inputs are disclosed and both came from the same call. The figure is reusable: it is what a hyperscaler with a mixed fleet pays, and it sits sensibly between general-purpose compute and NVIDIA's $60 billion pure AI gigawatt.

Three. The 26 GW is a continuation, not an acceleration. At 4.33 GW a year against a reported 4 GW a year, the leaked plan asks the construction organisation for 8% more than it is already delivering. What it asks the balance sheet for is a different question, because cost per gigawatt is rising even where gigawatts are not.

Four. Microsoft's true served capacity exceeds 12 GW by an amount nobody outside the company can size. Two public neocloud contracts alone add about half a gigawatt. Any cross-company capacity table that puts Microsoft's 12 GW next to a neocloud's contracted power is comparing an owned-and-leased figure with a figure that includes what Microsoft rents.

Five, and least certain. The level Microsoft has withheld is the one number that would let anyone check the other two. Additions without a base give a rate and no ratio. It is the reason a leak about a company that discloses a great deal is still news, and the reason the news is smaller than the headline suggests: the plan was visible, the total was not.

Primary sources worth reading

• Microsoft fiscal 2026 Form 10-K — Item 2 for the square-footage table, Note 13 for the $329.1 billion of leases not yet commenced and the lease-payment maturities, and the property and equipment policy note for the useful lives in force through fiscal 2026.
• Microsoft fiscal 2026 fourth-quarter earnings call — the gigawatt addition, the $41 billion, the $5.6 billion of finance leases, the two-thirds short-lived split, the useful-life extension and the $175 billion.
• Microsoft fiscal 2026 first-quarter call, second quarter and third quarter — where the doubling target was set and the earlier gigawatt additions were reported.
• Microsoft fiscal 2026 fourth-quarter results release — $90.0 billion of revenue, Azure up 43%, commercial remaining performance obligation of $678 billion up 84%, and additions to property and equipment of $35,802 million. No capacity figure of any kind.
• Bloomberg, 10 September 2026 — the report itself, by Brody Ford and Matt Day.
• NVIDIA's fiscal Q2 2027 call, captured in full — the $18/$25/$40 billion per-gigawatt ladder and the $30-to-$60 billion all-in figure.
• Microsoft's stock page, and the counterparties named here: NVIDIA, CoreWeave, IREN and Nebius.

Microsoft's capacity additions, the two-year doubling target, the Fairwater figure, every capital-expenditure and finance-lease figure for the June 2026 quarter, the short-lived-asset split, the useful-life extension and the calendar-2026 spending expectation and its earlier level are from Microsoft's fiscal 2026 earnings calls, quoted from the transcripts on its investor site. The square footage, the leases not yet commenced, the lease maturities, the multi-year additions to property and equipment and finance-lease assets, and the useful lives in force through fiscal 2026 come from the fiscal 2026 Form 10-K; revenue, Azure growth and remaining performance obligation come from the results release. The capacity target, the current fleet size, the AI share and the neocloud exclusion are an unconfirmed Bloomberg report of 10 September 2026 sourced to people familiar with the plans, on which Microsoft has not commented and whose own text notes the road map may change. NVIDIA's per-gigawatt figures are from its fiscal Q2 2027 call. The IREN and Nebius contract terms reach us through announcements rather than filings. Every rate, cost per gigawatt, implied base and multi-year total here is ours.