The Risks of Microsoft’s Relatively Restrained Capex

Microsoft came out of this summer’s earnings season looking like a prudent spender. Unlike rival cloud computing firms Amazon and Alphabet, which both burned cash thanks to their heavy AI investments, Microsoft not only generated $19.6 billion in free cash flow in the June quarter but forecast it will continue to generate cash for at least the next year. The result helped swing investment sentiment to the positive side, lifting the stock 29%.

What hasn’t received enough attention is how Microsoft has managed to stay cash flow positive. One longstanding reason is that Microsoft leases data center capacity from other companies including neocloud firms—such as CoreWeave—much more heavily than either Amazon or Google. That reduces its near-term capital expenditures but could give it less control over costs in the future. 

The other reason, as Microsoft disclosed in its June-quarter earnings report, is that the company made an accounting change that shifted some spending out of the capex bucket, allowing Microsoft to reduce its 2026 calendar year capex projection by $15 billion, or 8%. 

“I think there are a bunch of machines that trade on the headlines and saw a lower capex number and thought that was good,” said Charles Fitzgerald, a Seattle-based angel investor and former Microsoft executive. “The spend is the same. It just shows up on a different line item, as opex rather than capex. It’s not like they canceled a bunch of data centers and canceled orders for Nvidia [servers].”

The accounting change highlights the discretion companies have to massage the numbers they report. As Microsoft Chief Financial officer Amy Hood revealed on an earnings call, effective from July 1, the start of the company’s 2027 fiscal year, it extended its estimate of the “useful life” of its data centers and office buildings from 15 years to 25 years. 

As a result, Hood said, “more of our future data center leases will shift from finance leases to operating leases,” which means the lease payments move out of capex. Finance leases treat a data center, for example, as though Microsoft had purchased it directly, whereas operating leases treat data centers as rentals. That gets reported as operating expenses rather than capex. Operating expenses reduce reported profits, whereas capex reduces free cash flow, a metric investors have been paying close attention to in light of high capex costs.

The increase in data centers’ estimated useful life also means the depreciation associated with each new data center will be spread over a longer period of time, resulting in a smaller hit to earnings, although Hood said the impact would be “minimal” this fiscal year. The implication was that the depreciation impact would be greater in future years.

Microsoft still plans to ramp up its capex significantly. In its first quarter of fiscal 2027, which started July 1, it projects capex rising 43% to over $50 billion. Hood said she expected capex to grow in Microsoft’s fiscal 2027, which ends in June, as a response to increased demand for its cloud computing capacity.

Outside Data Centers

What may be a bigger factor in keeping a lid on capex is Microsoft’s leasing of capacity from other firms, particularly neoclouds such as CoreWeave and Nebius. As Microsoft says in securities filings, “in addition to datacenters we own or operate, we rely on third-party providers, including colocation facilities, leased datacenters, and cloud infrastructure providers, to support portions of our operations.”

Amazon Web Services, in contrast, operates mostly in its own data centers, although it leases some capacity from other firms, as in a deal it announced in November with Cipher Mining.

The same is true of Google, although it has done some deals with outside compute providers, such as its recently announced $30 billion deal to rent capacity from SpaceX. More such deals are likely. On the company’s earnings call last month, Chief Financial Officer Anat Ashkenazi said Google planned to “expand the use of third-party capacity” in the third quarter “as a bridging strategy while we build up more internal capacity.” She noted that given the cost of that capacity, it would “put some pressure on operating margins for Cloud.”

Microsoft, though, has a longer history of such deals and contracts with more partners. It has committed to spending at least $60 billion with various neocloud data center providers, including Nscale, Nebius, Iren and Lambda, to rent data centers full of AI chips and related hardware, many of which are five-year contracts, Bloomberg first reported. 

Since outside data center leases are typically classified as operating leases, “there’s no doubt that at one point a ton of their infrastructure was showing up as opex, not capex,” said Fitzgerald.

Microsoft’s heavier use of outside firms is partly a result of its initially fast pace to secure computing capacity for AI after the release of ChatGPT. Building data centers takes time, whereas leasing capacity from a provider that has already built them and secured the power connections and cooling equipment to operate them may have enabled Microsoft to move faster at a time of surging demand for that capacity—both for its own models through OpenAI, and for its customers. 

“I think Microsoft got off the blocks faster than everybody else because they were the first to see the explosion in OpenAI demand because they were providing all of the infrastructure capacity,” Fitzgerald said. “It took Amazon much longer to start that build-out.”

While it’s unclear what percentage of Microsoft’s data centers are owned versus operated by third parties, the downside to this strategy is that in the long run it leaves Microsoft exposed to price increases by its cloud computing partners.

To be sure, Microsoft struck five-year deals with these outside cloud firms, which might reduce its vulnerability to neocloud price hikes. And as D.A. Davidson analyst Gil Luria points out, it’s possible that in the next three to five years, Microsoft will have built out enough of its own data centers that it “won’t need neocloud capacity.” Because it rents out this capacity to Azure customers, most notably OpenAI, Microsoft is charging a markup to those customers above the price it pays to neoclouds to rent the space and equipment, Luria added.

On the other hand, as Fitzgerald noted, “everybody is scrambling for any capacity that they can find, and as prices go up, if you’re a third party who has actually got power and can build it out and turn it on, people who need the capacity will pay what they have to pay.”