Analysis
Microsoft is planning to more than triple its data-center footprint to roughly 38 gigawatts of capacity by 2032, Bloomberg reported Thursday, citing people familiar with the company's internal planning. The company currently operates about 12 gigawatts of capacity, of which only around 2 gigawatts is dedicated to AI-specific chips -- a share Microsoft expects to grow to roughly a third of the much larger 38-gigawatt base it's targeting six years from now.
The plan is a direct response to a shortage Microsoft has already had to ration around: severe hardware constraints have forced the company to turn away some cloud and AI business, restrict certain subscriptions and absorb service disruptions over the past year, according to the same report. That's an unusual admission given the company's own spending pace:
- FY2026 capex -- $145 billion, already a company record.
- Q1 FY2027 guide -- roughly $50 billion in a single quarter.
- 2026 calendar-year pace -- on track for about $175 billion.
“That's an unusual admission given the company's own spending pace: - FY2026 capex -- $145 billion, already a company record.”
Even that historic level of spending hasn't kept pace with internal demand for AI compute.
The scale problem
A 38-gigawatt footprint would draw more electricity at its peak than the entire state of New York, multiple outlets reported following Bloomberg's account, underscoring just how far outside historical utility planning this generation of AI infrastructure has moved. Microsoft isn't alone in this scale of ambition -- Pulse has tracked Amazon, Google and Meta's own multi-gigawatt commitments throughout 2026 as all four hyperscalers compete for the same scarce inputs: turbines, transformers, grid interconnection queues and, increasingly, their own dedicated power generation.
That race for power is exactly why 2026 has produced a wave of hyperscaler-backed nuclear financing deals -- Google's federal loan to restart Iowa's Duane Arnold plant, Microsoft's own Three Mile Island restart, and Amazon's Talen Energy campus all represent the same underlying constraint the 38-gigawatt target now quantifies at Microsoft's own scale. It's also the direct cause of the political backlash on the other side of the ledger: Massachusetts, New York and Oklahoma have each moved this year to force data-center developers to bring their own clean power or pay into ratepayer-protection funds, rather than let hyperscaler demand quietly raise electricity bills for everyone else.
What the target doesn't guarantee
Microsoft has not publicly confirmed the 38-gigawatt figure, and a stated internal planning target is meaningfully different from permitted, financed, energized capacity actually coming online by 2032. Every hyperscaler in this buildout cycle faces the same execution gap between an ambition announced in a boardroom and megawatts actually flowing through a substation -- grid interconnection queues in many U.S. regions already stretch years, turbine and transformer lead times remain multi-year bottlenecks industry-wide, and local opposition to new server farms has intensified enough that several state governors have moved to halt new construction outright. The risk is that an announced target and delivered gigawatts are not the same thing, and none of that makes the target implausible, but it does mean the real test of this plan won't be Thursday's report -- it will be Microsoft's actual gigawatts-online figure in each of the next several fiscal years, a number the company already discloses and that will make clear whether 2032 is a realistic target or an aspirational one.
For AI labs and infrastructure investors, the practical signal is that compute scarcity isn't closing anytime soon: if the company spending the most on AI infrastructure still can't meet its own internal demand today, pricing and access to frontier compute will likely stay a seller's market well past this specific 2032 target date.
