MMOOZZGG.com

AI investment

Microsoft AI investment: infrastructure, stakes and deals

6 min read5 sources

Microsoft AI investment in fiscal year 2026: $115.9 billion on equipment and construction, against $64.6 billion a year earlier. Below: what this money buys, which stakes the company holds in other AI companies and which deals it made in 2026. There are no share price forecasts here: our guide records decisions and figures from company reports and large media.

On this page · 5 sections

How much it invests

Microsoft’s fiscal year runs from July to June. In fiscal 2026, from July 2025 to June 2026, the company spent $115.9 billion on equipment and construction, according to its press release on Q4 results.

A year earlier the sum was $64.6 billion. By our count, spending almost doubled in a year.

The pace did not slow by the end of the year: of the $115.9 billion, one quarter, April-June 2026, took $35.8 billion.

Timing matters. Microsoft builds and buys capacity before clients have paid for the demand, so it risks its own money in advance. The quarterly report shows how fast this bet grows.

Forbes estimates Microsoft’s capital expenditure for calendar 2026 at about $190 billion. This is the outlet’s estimate for the calendar year, and you cannot compare it directly with the reported fiscal year figure: the periods differ.

By the same Forbes estimate, Amazon, Microsoft, Alphabet and Meta together will spend $700–725 billion in 2026, against $410 billion in 2025. How this money splits among the four companies is on the page about capex of the largest companies.

In what: infrastructure and models

Most of the money goes into computing capacity: data centers, chips and the equipment for them. A separate line is the company’s own chips, which reduce dependence on outside suppliers.

On 26 January 2026, Microsoft introduced its own AI accelerator, Maia 200. It is built to run models that are already trained, not to train them, and TSMC makes it on a 3 nm process.

According to the company, Maia 200 gives 30% better performance per dollar than Microsoft’s other newest hardware. The US Central data center region in Iowa got the accelerator first.

The choice of task for the chip shows the priorities. The company expects the main spending to go not to training models but to the daily work of AI for clients, and it builds hardware for this load.

The financial logic is simple. Models answer clients every day, and the price of each answer decides whether the cloud makes money on AI. Its own chip lowers this price and gives the company an argument in talks with chip makers.

The second line is renting other companies’ capacity. In April 2026, Microsoft agreed to rent another 30,000 Nvidia Vera Rubin chips from Nscale in Narvik, Norway.

The company secured access to frontier models through its stake in OpenAI. The next section covers this stake and its terms.

Stakes in other companies

Microsoft’s main AI stake is OpenAI. After the OpenAI restructuring on 28 October 2025, Microsoft owns about 27% of OpenAI Group PBC.

The stake was valued at about $135 billion. This is a valuation at the date of the restructuring, not a price at which the stake can be sold.

In return, OpenAI committed to buy an additional $250 billion of Azure cloud services. By our count, the commitment is almost twice the value of the stake itself.

The structure shows how a large company ties an investment to orders: money and capacity go one way, and cloud payments come back. A separate page of our guide explains in more detail how the giants invest in startups.

For a CFO, the lesson is about the deal structure: the stake and the cloud contract were signed together. So the value of the stake for Microsoft also depends on OpenAI’s future orders, not only on the company’s valuation.

We cover OpenAI’s rounds and valuation in 2026 on the company’s page.

Our sources name no other large Microsoft stakes in AI companies for 2026, so our guide has none. New deals will appear here after a check against company reports or large media.

Deals of 2026

Microsoft’s AI moves in 2026 that we checked against primary sources, by date:

1. 26 January: launch of its own Maia 200 accelerator, with US Central in Iowa as the first region.

2. 15 April: an agreement to rent another 30,000 Nvidia Vera Rubin chips from Nscale in Narvik. The site was earlier prepared for OpenAI’s Stargate project.

3. April-June: $35.8 billion on equipment and construction in one quarter, according to the company report.

Microsoft had already made commitments of $6.2 billion for the Narvik site. The deal shows how fast plans change: capacity prepared for one project moves to another tenant.

For Microsoft, it is a way to rent a ready site rather than build from scratch. For the market, it is a signal: even the largest buyers of capacity revise their plans within a year.

All checked AI market deals of 2026, with amounts and sources, are in our deal tracker.

What a Russian company can learn

Microsoft’s scale does not carry over directly to a Russian company: about $190 billion of capex in 2026, by the Forbes estimate. What carries over are the decisions behind this sum.

1. Count the cost of running a model, not only of buying it. Microsoft made a separate chip for trained models because the cost of daily answers grows with every client. In your project, set the price of one request and multiply it by the number of requests per month.

2. Rent until the load is confirmed. A company with reported spending of $115.9 billion a year still rents capacity from Nscale. Your own server makes sense when the load is clear and stable.

3. Tie an investment to orders. The OpenAI stake came with a commitment to buy $250 billion of cloud services. If you invest in a contractor or partner, write counter commitments into the contract, not only the stake.

4. Report by quarter. Microsoft shows its spending every quarter, and you can see how much went out in April-June. Your board needs the same rhythm: plan, actual and effect for each AI project once a quarter.

5. Keep a path to replacement. Its own accelerator next to Nvidia chips reduces dependence on one supplier. For your company, the same means a model and a server that you can replace without changing your processes.

Where to start in your company is shown by the AI Solutions Map: the processes where AI pays back first, and the order of steps.

Sources and fact-checking

Checked by the editors on 14 October 2026.

  1. Microsoft, press release on Q4 fiscal 2026 results, 29 July 2026 (microsoft.com)
  2. Microsoft, official blog, Scott Guthrie, 26 January 2026 (blogs.microsoft.com)
  3. Microsoft, official blog, 28 October 2025 (blogs.microsoft.com)
  4. Bloomberg News, Brody Ford, Kurt Wagner, Mark Bergen, 15 April 2026 (news.bgov.com)
  5. Forbes, Jason Kirsch, 7 October 2026 (forbes.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us