Microsoft wants to triple its data centers | Future Intelligence #26
Microsoft wants to expand its data center network from about 12 to more than 38 gigawatts by 2032; the real limit of AI today is electricity capacity, not model size. Nvidia is considering up to $10 billion in Anthropic's IPO at a valuation around $2 trillion, Google is sending €13 billion to data centers in Finland, and Qualcomm secured its AWS contract with a $4 billion warrant. Meta also launched the Muse agent at $20 and $100 per month—the first serious attempt to charge for software that itself pays, buys, and books.
Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically every Sunday morning by email—in full length, including specific scenarios and market implications. Permanent access to the report is available with Bulios Black membership.
Key points
Microsoft wants to expand its data center network to several times today's capacity by 2032
Why available electricity capacity is becoming a harder limit for AI development than the chips themselves
Meta launched the Muse agent, which is supposed to pay and book on its own, and the surprising part is mainly its price tag
Qualcomm will supply custom chips for Amazon's data centers; the deal includes an unusual warrant
Which AI infrastructure supplier had the wildest week and what's happening around Nvidia's license with Groq
Microsoft wants 38 GW by 2032: capacity is the new AI limit
Bloomberg on Thursday, September 10, described a plan that gives the AI infrastructure boom a concrete unit. Microsoft, according to people familiar with the intentions, plans to expand its data center network from today's roughly 12 gigawatts to more than 38 GW by 2032. The 26 GW increase covers both owned and leased sites and excludes capacity rented from neo-clouds like CoreWeave.
The reason is not ambition but shortage. The company has been turning down some cloud and AI orders due to missing capacity and limiting some subscriptions. The company that sells servers is itself waiting for servers.
Gigawatt as the new currency of the industry
A gigawatt is a unit of power, not computation—but in the AI era it is the hardest constraint because chips can be ordered faster than a grid connection and a transformer. Of today's 12 GW, only about 2 GW is on AI-specific silicon; by 2032, about a third of the 38 GW, or approximately 12.7 GW, is expected to be AI-specific.
The ratio is more important than the sum: total footprint grows just over threefold, the AI portion more than sixfold. The portfolio transformation from classical cloud to accelerated computing is thus running faster than the headline number suggests.
For comparison: in 2025, Microsoft commissioned around 2 GW of new capacity and operated over 400 data centers.
The cost is visible in the financials. Capital expenditures including finance leases reached an adjusted $175 billion for fiscal year 2026, with the fourth quarter alone at $41 billion (+69% year-over-year). CFO Amy Hood in the spring spoke of roughly $190 billion for calendar year 2026, 61% more than in 2025, including a $25 billion impact from more expensive components.
The structure of those expenses determines the margin. Roughly two-thirds of quarterly capex went to short-lived assets, primarily GPUs and CPUs, which are depreciated over years, not decades. Gross margin of 67.6% in fiscal Q3 was the lowest since 2022 precisely due to rising infrastructure depreciation.
Where construction gets stuck
That a plan on paper is not yet capacity on the grid is shown by Project Braid—a joint venture between Alphabet and Blackstone launched in May 2026 with $5 billion from Blackstone, intended to lease Google TPUs from 2027. The 500 MW target is hitting delays at key locations.
The details are more instructive than the delay itself. Project managers estimate the chance of meeting deadlines at 50% compared with 90% three years ago; one site waited for electrical transformers with lead times approaching a year, and Google took over the Cheyenne, Wyoming project from developer Crusoe, including the permitting process.
From there a direct line leads to energy and heavy equipment. The four largest hyperscalers had contracted over 133 GW of power purchase agreements as of Q1 2026, according to data cited in Fervo Energy's registration document.
Caterpillar's backlog for power generation equipment reached $72 billion—a segment that in revenue is outpacing the traditional construction machinery cycle.
Europe, debt, and the bill for expansion
The European branch of the same trend has a Finnish address: Google announced on September 9 an investment of at least €13 billion ($15.1 billion) in data centers and related infrastructure in 2027 and 2028—three new sites plus expansion of the Hamina campus, supplemented by locations in Kajaani, Muhos, and Vaala. It is its largest single investment in Europe to date.
The choice of Finland is not marketing: the cold climate reduces cooling consumption, and the low-carbon grid addresses both cost and emissions reporting. According to the company, the construction phase is expected to contribute an average of €3.6 billion annually to Finland's GDP and support over 37,000 jobs.
Financing is shifting to the bond market. Amazon issued its first-ever sterling bond, raising £4.25 billion ($5.76 billion) in four tranches—£1.25 billion in three-year notes and £1 billion each in six-, twelve-, and nineteen-year paper with yields from about 5.2% to 6.7%. Capacity is no longer paid solely from operating cash flow.
For $MSFT stock, 38 GW is an ambiguous number: it confirms that demand exceeds supply, but it also locks in the depreciation curve for six years ahead.
After July's results the stock gained 8%—the market is pricing the order book higher than the margin pressure, which holds only until cloud revenue growth lags depreciation growth.
The distribution of the bet is broader than $GOOGL and $AMZN with this year's capex plans around $175–185 billion and $200 billion, respectively.
Delays like Project Braid shift value to supply chain bottlenecks—transformers, turbines, and power packages, where $CAT's order backlog is measured in tens of billions and where it will be decided how many of the announced gigawatts actually flow.
Meta launched Muse: an agent that pays and books for you
Meta this week introduced Muse—an AI agent that, according to the company's description, can autonomously send emails, make payments, sell a car, or book a trip on the user's behalf. The novelty is not just the ability to act but the price tag: paid tiers at $20 and $100 per month.