Microsoft Needs About 1GW a Quarter Until 2032. Transformer Lead Times Run to Three Years.
Bloomberg reports Microsoft is targeting more than 38GW of data center capacity by 2032, up from around 12GW now. That is roughly 26GW of new build in six years, or about a gigawatt a quarter sustained. The thing standing in the way is not contractor capacity. It is a transformer market running three-year lead times, constrained by a workforce of a few thousand people who wind copper coils by hand.
Table Of Content
On 10 September, Bloomberg reported that Microsoft plans to take its global data center footprint past 38GW by 2032, citing people familiar with the plans. The company has not confirmed the figure in a filing, an earnings call or a press release, and declined to comment. Treat it as credible reporting rather than guidance.
Even so, the shape of it matches what Microsoft has said out loud. Satya Nadella told investors on 29 July that the company remains on track to roughly double total capacity within two years. It added 1GW in the quarter ending 30 June, bringing 31 data centers online in that quarter alone and 88 across the full fiscal year. The reported 2032 target needs an average of about 4.3GW a year. Microsoft’s own near-term guidance implies a faster rate than that.
For anyone who builds, supplies or sells into this work, the interesting question is not whether the target is real. It is which physical constraint gives way first, and whether it is one you can do anything about.
What is actually committed, and what is reporting
Bloomberg’s number against Microsoft’s own disclosures
The confirmed material is worth separating from the reported material, because the two get blended in most coverage. Confirmed: capex of $145.3bn in 2026, after $115.9bn in 2025 and $55.7bn in 2024, with roughly $175bn indicated for the following year. Also confirmed: $329.1bn of leases that had not yet commenced as of 30 June 2026, up from $196.6bn the previous quarter. That second number is the clearest signal of intent in the whole set.
Reported rather than confirmed: the 38GW figure itself, the 2032 date, and the mix. Bloomberg puts roughly 2GW of current capacity on AI-specific chips, rising to about a third of the eventual footprint. Microsoft also changed the assumed useful life of its data centers from 15 to 25 years effective fiscal 2027, which moved more leases into operating rather than finance classification and pulled headline calendar 2026 capex guidance from around $190bn down to around $175bn without changing a single construction plan.
What 26GW looks like as a build programme
Pecos is 2GW and takes five to seven years
Microsoft is developing a campus at Pecos in Texas that it described in June as one of the largest single capacity additions in its history. It is about 2GW, and it is expected to take five to seven years to build out. Against the roughly 26GW of additional capacity the reported target implies, Pecos is less than a tenth of it.
That comparison is the useful one. Sites vary enormously in size and design so it does not translate neatly into a campus count, but it makes the scale legible. You cannot reach 38GW by building a dozen flagship campuses. It requires a continuous, parallel, multi-continent programme running at a pace nobody in the industry has sustained, combining owned build with leased capacity from CoreWeave, Nscale, Lambda, Iren and Nebius.
One small piece of good news for contractors and their supply chains: Microsoft stopped requiring NDAs on its data center projects back in March. More of this programme is visible than used to be the case.
The constraint sits upstream of the site
A record backlog, and no new capacity until 2030
Siemens Energy’s Grid Technologies unit reported orders up 27.6 percent to €5.37bn in the quarter ending 30 June, a book-to-bill ratio of 1.48, and a record backlog of €51bn. Large power transformers drove most of that. A book-to-bill of 1.48 means the unit booked half again as much work as it shipped. The queue is getting longer, not shorter.
The company does not expect to bring its planned 50 percent increase in transformer and gas-insulated switchgear capacity online until 2030. Every project energizing between now and then competes for the existing lines. In the US, standard power transformers have been averaging around 128 weeks and generator step-up units around 144 weeks, with specialty units quoted as far out as four years. Five years ago comparable units shipped in 50 to 80 weeks.
Microsoft’s programme dates are not being set by planning consent, contractor availability or capital. They are being set by whoever holds a slot in a transformer factory, and those slots were allocated a long time ago.
The bottleneck inside the bottleneck
Roughly 15,000 people in the US wind these coils
Most coverage of the transformer shortage stops at grain-oriented electrical steel, the magnetic core material that has roughly doubled in price since early 2020 and is produced domestically at a single mill in Butler, Pennsylvania. That is a real constraint and a well-documented one.
The harder constraint is human. The US transformer workforce runs to roughly 15,000 people, and the estimate is that it would need to roughly triple to meet demand. At the center of it are coil winders, whose craft the trade press describes as “almost an art form”. That workforce is aging out and the vocational pipeline that used to refill it disappeared a generation ago.
Capital does not fix this quickly. You can fund a new line, pour the slab and install the presses, and still wait years for people who can wind a 400kV coil to a tolerance that passes impulse testing. It is the same category of problem the construction industry has with high-voltage commissioning engineers, which is to say a labor problem wearing a supply chain costume.
What contractors and ConTech should take from this
Procurement moves in front of design
If long-lead electrical equipment sets the programme, then the sequence most delivery teams were trained on is backwards. Ordering a transformer after design freeze and permit puts your energization date three years out from a decision you could have made eighteen months earlier. The firms winning this work are placing orders during concept design, carrying the spec risk, and treating equipment slots as the scarce asset rather than crew availability.
There are three practical consequences. Owner-furnished equipment becomes more common, because the client wants control of the slot. Bid validity windows keep shrinking, because nobody can hold a price against a market moving like this. And substation and high-voltage EPC capability becomes the differentiator on these bids rather than shell-and-core delivery, which is precisely why acquisitions in behind-the-meter power architecture are now trading at software-like multiples.
For ConTech founders, the demand signal is not where the press releases point. It sits in equipment logistics, electrical prefabrication, commissioning management and anything that compresses the gap between a connection agreement and a live site. That is where the schedule is being lost.
Join the construction briefing people read to stay one step ahead
News, deals, and deep takes on the forces reshaping construction.
Join 3000+ Readers
The target against the supply chain
| Input | What the programme needs | What the market is delivering |
|---|---|---|
| Capacity growth | About 4.3GW a year to 2032, faster in the near term | 1GW added in the quarter to 30 June 2026 |
| Large power transformers | Orders placed at concept design | Around 128 weeks, generator step-up units around 144 |
| OEM factory capacity | Additional lines available now | Siemens Energy’s 50 percent expansion lands in 2030 |
| Transformer labor | A workforce roughly three times current size | About 15,000 people in the US, aging out |
| Core steel | Multiple domestic sources | One US mill, price roughly doubled since 2020 |
| Capital | Sustained multi-year commitment | $329.1bn of leases not yet commenced, not the binding constraint |
Frequently asked questions
No. The figure comes from Bloomberg reporting on 10 September citing people familiar with the plans. Microsoft declined to comment and has not stated it in a filing, earnings call or press release. What the company has confirmed is guidance to roughly double capacity within two years. (Source)
It is a mix. Microsoft combines self-owned and operated facilities with leased capacity from third parties including CoreWeave, Nscale, Lambda, Iren and Nebius. The $329.1bn of leases not yet commenced as of 30 June gives a sense of how much sits on the leased side of that ledger. (Source)
An accounting change rather than a change of plan. Extending assumed data center useful life from 15 to 25 years effective fiscal 2027 pushes more leases into operating rather than finance classification, and operating leases do not show up in reported capital expenditure. Headline calendar 2026 guidance moved from roughly $190bn to roughly $175bn as a result. (Source)
It means Siemens Energy’s grid business booked roughly one and a half times as much work as it shipped in the quarter. A ratio above 1 sustained over several quarters is what produces a lengthening queue, which is exactly what the €51bn backlog represents. It is an order book, not revenue. (Source)
Not on the timeline this programme needs. Siemens Energy’s planned 50 percent capacity increase arrives in 2030. Beyond that, the binding constraint is reported to be skilled winding labor rather than capital or steel, and a workforce of roughly 15,000 in the US would need to roughly triple. Factories are faster to build than the people who staff them. (Source)
Move long-lead electrical procurement in front of design freeze, price escalation into bids rather than absorbing it, and build or buy high-voltage and substation capability. Equipment slots, not crews, are the scarce input on these programmes, and clients increasingly want to own that risk through owner-furnished equipment. (Source)
Related Articles
Power Is the New Lumber: The Supply Chain Bottleneck No One Saw Coming
Big Tech Promised $650 Billion in Data Centers. Most of It Isn’t Being Built
Vertiv Is Paying $1.45bn for a Company That Doesn’t Build Data Centers
Data Center Dynamics: Microsoft targets 38GW of data center capacity in 2032
Cloud Computing News: Microsoft reportedly targets 38GW of data centre capacity by 2032
MGrid: Siemens Energy’s grid order backlog hits a record €51 billion
Turbomachinery Magazine: Siemens Energy posts record backlog
Manufacturing Mag: The grid-scale transformer shortage and the winding labor bottleneck
