Microsoft Plans to Expand Data Center Capacity to 38 GW by 2032
Sources told Bloomberg that Microsoft plans to increase its global data center capacity from the current approximately 12 GW to over 38 GW by 2032, more than tripling its capacity, which exceeds the peak electricity consumption of New York State. The goal is to address the computing power gap that has forced the company to turn down some artificial intelligence and cloud business requests.
The plan includes both owned and leased data centers, excluding computing power rented from "new cloud" providers like CoreWeave. Of the existing approximately 12 GW, about 2 GW is dedicated to artificial intelligence chips; by 2032, AI-specific capacity is expected to account for about one-third of the total 38 GW, with the remainder still serving general cloud and CPU loads. Cloud infrastructure executive Alistair Speirs stated that next-generation AI tools are increasingly reliant on CPUs in addition to GPUs.
The shortage has reached the order level. The company had previously restricted new cloud subscriptions in certain key areas of the U.S. and Europe, potentially directing business to competitors. Satya Nadella stated that they are ramping up capacity as quickly as possible and increasing output by improving the utilization of existing hardware. In early 2025, Microsoft temporarily halted some development projects due to concerns about overbuilding, which was later viewed internally as one of the reasons for the current bottleneck, with some shelved sites being taken over by Amazon, Google, Meta, and Oracle.
Capital expenditures are already at a high level. In the most recent fiscal year, capital expenditures were approximately $145 billion, with analysts expecting continued increases in subsequent years; the company has also launched about 1 GW in a single quarter and stated that it will double its overall footprint within two years. Ongoing and planned projects include the Fairwater campus in Wisconsin and a 2 GW-level park in Pecos, Texas, with self-supplied energy arrangements. The 38 GW corresponds to a net increase of about 26 GW, with a timeline extending to 2032, while power supply and grid connection remain constraints.
The expansion will serve workloads for Azure, Microsoft 365, and GitHub Copilot, rather than just a single training cluster. The plan aims to parallel the deployment of general computing and AI cabinets, avoiding a repeat of the structural gap where "only GPUs are piled up, and CPUs and storage cannot keep up." The leasing of data centers and self-built facilities will run concurrently, with external rentals from CoreWeave not counted in this 38 GW metric; the reported capacity does not equate to the cloud quotas available for external sale.
In market mechanics, this is a rush driven by supply shortages. Buyers include cloud and model customers who have been denied capacity, as well as Microsoft itself, which must secure electricity; sellers include power, land, diesel backup, GPU/CPU, and contractors. Funds flow from Microsoft's capital expenditures to grid access, chips, and contractors, with areas capable of fulfilling orders retaining rent and electricity costs, while areas unable to accommodate orders pass subscriptions to AWS, Google Cloud, and Oracle. Beneficiaries are power and chip suppliers capable of delivering in GW; those under pressure include peak grid capacity, sites that have been paused and are now re-entering the queue, and new customers still subject to capacity quotas.
On a supplementary level, third-party tracking of Microsoft’s AI-related data centers does not completely overlap with the company’s internal 12 GW metric; the planning figures come from informed sources and are not formal guidance. External rented computing power from CoreWeave remains a buffer layer and is not included in the 2032 owned plus leased target.
Source: Public Information
ABAB AI Insight
This is not the first time Microsoft has turned cloud computing into an electricity issue. In the 2010s, Azure relied on leased data centers to catch up with AWS, and in 2023-2024, it rushed to secure land and electricity, only to hit the brakes in early 2025 due to concerns about oversupply, resulting in competitors filling the sites. Now, with the 2032 target of 38 GW, Microsoft is rewriting the "brake" into a "catch-up". Amy Hood's concerns about overbuilding back then and Nadella's current push for rapid deployment are two seasons on the same capital expenditure sheet: cutting projects when demand proves insufficient, and writing GW back into plans when orders are turned down.
The capital path is "self-built plus leased, with external rentals counted separately". The annual capital expenditure of around $145 billion first buys short-cycle chips, then long-cycle power and facilities; campuses like Fairwater and Pecos include energy arrangements in their project packages to reduce single-point reliance on utilities. Not counting CoreWeave in the 38 GW metric keeps the flexible layer of externally purchasable capacity off the books, with only cabinets that can be controlled internally committed on the books. The motivation is not to predict the model form in 2032, but to reclaim the cloud subscriptions that were previously denied.
In contrast to Amazon's long-term regional stockpiling, Google's use of self-developed chips to reduce output per watt, and Oracle's use of Abilene clusters for training: Microsoft this time seems to be compensating for the three years of pause with a slope leading to 2032. The industry phase has shifted from "who first announces a $10 billion capital expenditure" to "who can turn GW into sellable regional quotas". Control lies in power and grid connection, not in the model names in press releases.
Structural changes belong to the reconstruction of the industrial chain. The mechanism is: AI demand has transformed cloud computing from a software gross profit business into a delivery business of power and civil engineering; when capacity is insufficient, subscriptions are rationed, and customers divert, thus pricing power shifts from software agreements to sites that can connect to the grid before 2032. The 38 GW is benchmarked against New York State's peak electricity, indicating that the next round of competition will measure units in terms of the grid, not virtual machine lists.
ABAB News · Cognitive Law
- The upper limit of the cloud is first determined by electricity, then by software.
- A site paused for a year will become three years of market share for competitors.
- Capital expenditure buys quotas, not model stories.