Microsoft targets 38GW of data centre capacity by 2032, Bloomberg reports
Tripling a hyperscaler's footprint in five years turns on gigawatts, grid connections and lease deals as much as on chips.
Microsoft is planning to more than triple its data centre capacity within five years, targeting 38GW in 2032, according to a Bloomberg report cited by Data Center Dynamics on 11 September. The company currently operates around 12GW. Bloomberg attributed the figure to people familiar with the plans, and Microsoft declined to comment on its building programme.
DCD reports the growth will come from a mix of data centres Microsoft owns and operates and capacity leased from third parties, including so-called neoclouds. The company has signed large contracts with CoreWeave, Nscale, Lambda, Iren and Nebius, and DCD put that spending at an estimated $60 billion as of November 2025.
Capital expenditure has climbed steeply alongside. DCD lists total capex of $55.7 billion in 2024, rising to $115.9 billion the following year and $145.3 billion in 2026, with an estimated $175 billion for 2027. In recent quarters the company has been bringing roughly 1GW of capacity online per quarter; across its 2026 fiscal year it opened 88 data centres, 31 of them in the final quarter. DCD calculates that hitting 38GW by 2032 requires nudging that pace to a little over 1GW per quarter on average.
Among the recent openings are two "Fairwater" sites, large AI-dedicated facilities that each house hundreds of thousands of GPUs. Microsoft said in March that it would no longer use non-disclosure agreements around its projects, and it is known to be developing in several locations worldwide. DCD said it had asked Microsoft for comment.
The report also notes growing public scrutiny. In the United States, states are increasingly requiring data centre projects to fund upgrades to local electrical and water infrastructure, and some have enacted pauses on new developments, adding complexity to hyperscaler build-out plans.
Why it matters
Capacity is now measured in gigawatts, which makes power procurement and grid interconnection the gating factors rather than servers. The reliance on leased neocloud capacity shows even the largest operators cannot build fast enough alone. Rising state-level conditions signal that community acceptance is becoming part of the engineering plan. The numbers set a benchmark that every other operator's roadmap will be read against.
Capacity is now measured in gigawatts, which makes power procurement and grid interconnection the gating factors rather than servers.
- 2024$55.7bn
- 2025$115.9bn
- 2026$145.3bn
- 2027 (est.)$175bn
Figures: Figures as reported by Data Center Dynamics; 2027 is an estimate
What you can learn from this
- Gigawatts are the unit because power, not floor space, bounds a data centre. A facility's capacity is quoted as the electrical load it can deliver to IT equipment, since every rack of accelerators draws a fixed amount and cooling scales with it. A 1GW campus needs a grid connection on the scale of a small city, so the timeline is dominated by utility agreements, substations and transmission rather than construction. Cooling and electrical losses mean the grid supply must exceed the IT load, so the quoted figure understates the connection required. That is why announcements name megawatts and gigawatts instead of square metres.
- Owned versus leased capacity is a trade-off between control and speed. Building your own site gives design control and long-run cost advantages but ties up capital for years before the first server runs. Leasing from a neocloud or colocation provider converts that into an operating expense and shortens time-to-capacity, at the price of margin and dependence on another operator's roadmap. Hyperscalers blend both because the demand curve is steeper than any single construction pipeline.
- Capital expenditure on this scale is a multi-year commitment with long payback. Data centre spend is sunk into land, power infrastructure and hardware that depreciates over a fixed schedule, so the return depends on utilisation years after the cheque clears. Rapid capex growth therefore embeds an assumption that demand will keep filling the racks, and the spend also lags demand, since money committed now buys capacity that arrives one to three years later. Reading the capex trajectory alongside utilisation is how analysts judge whether a build-out is measured or speculative.
- AI-dedicated sites differ from general-purpose cloud halls. Facilities built around hundreds of thousands of GPUs run at far higher rack densities, need liquid cooling and are laid out so accelerators can talk to each other with very low latency. That specialisation makes them less flexible to repurpose, which raises the stakes on forecasting. It also explains why such sites are announced as distinct projects rather than as expansions of existing regions.
- Local infrastructure costs are becoming an explicit line item. When a project needs new substations or water capacity, someone pays, and states are increasingly saying that should be the developer rather than ratepayers. Requirements like these change the site selection calculus, favouring locations with spare grid headroom. Pauses on new projects work the same way from the other direction, removing candidates from the map entirely.
We teach this
How to use this in practice
- Convert a capacity figure into something tangible. Take 38GW, divide by a plausible per-rack power draw for AI hardware, and estimate how many racks that implies; then do the same for 12GW and compare. Then multiply the rack count by a typical accelerator count per rack for a rough sense of device volume, and write your assumptions down explicitly so the sums can be challenged. Done looks like a short worked calculation with a sentence on what the difference means for grid demand.
- Trace one region's supply chain on a map. Pick a cloud region you use, find its published availability zones, and sketch where the power is likely to come from and which operators host the sites. Mark which parts are owned by the provider and which are leased where that information is public, and if the provider publishes sustainability or regional reports, use them to fill gaps and note where you had to guess. Done looks like a one-page diagram with a source noted beside each element.
- Read one utility or planning document end to end. Find a public interconnection request, planning application or state legislative summary related to a data centre and note the timeline, the conditions imposed and who bears the cost of upgrades. Compare those timelines with the quarterly build rate in this story to see where the bottleneck sits. Done looks like five bullet points summarising the constraints and how long each step took.
- Check how your own workloads depend on region capacity. List the cloud regions your systems deploy to and look up whether each has published capacity constraints or new-customer limits for accelerator instances. Include the instance types that matter to you, since constraints often apply to specific accelerator families rather than whole regions. Where a region is constrained, write down the fallback. Done looks like a table of regions, constraint status and a named alternative for each.
Sources
- Microsoft targets 38GW of data center capacity in 2032 - report — Data Center Dynamics
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report