Hyperscalers' AI capex: what it means for the prices enterprises pay

In brief

Q3 2026 filings show about $237 billion of quarterly capex across five companies. How depreciation and utilization shape what enterprises pay for AI.

Steve Elliott

Steve Elliott · · 7 min read


On this page
  1. How much did the hyperscalers spend in Q3 2026?
  2. What do the companies say about demand and margins?
  3. Why does depreciation drive the price of AI capacity?
  4. Are useful-life choices changing the numbers?
  5. Which price moves have providers already made?
  6. What does this mean for the prices enterprises pay next?
  7. Sources

Amazon, Microsoft, Alphabet, Meta and Oracle reported roughly $237 billion of capital spending in their latest quarters, on each company's own measure. Amazon, Microsoft and Alphabet each tie most of theirs to AI or to technical infrastructure. In our view, that spending is already reaching enterprise bills: scarce AI capacity costs more, AI model discounts are expiring, and providers are making commitments harder to unwind.

The reason is the cost structure. Servers and GPUs are typically depreciated over five to six years whether or not anyone uses them, so a provider's cost per hour sold depends on utilization. This post reads the Q3 2026 filings for what they say about that cost base, and what it implies for price direction. For how to negotiate a commitment, see negotiating cloud commitments when capacity is tight.

How much did the hyperscalers spend in Q3 2026?

Each company defines AI capex differently, so the figures below are each company's own measure, not a like-for-like total.

CompanyCapex in the quarterFull-year outlookDepreciation in the period, vs a year earlierServer useful life
Amazon (Q3, to Sept 30)$77.3B purchases of property and equipment, vs $35.1B a year earlierAbout $220B cash capex in 2026Quarterly D&A (combined line) $20.9B, up 25%Five to six years; a subset cut from six to five in 2025
Microsoft (fiscal Q1 FY27, to Sept 30)$45.2B including $7.4B of finance leases; $39.4B cashAbout $170B for calendar 2026Quarterly depreciation $11.1B, up 56%Servers and network equipment two to six years
Alphabet (Q3, to Sept 30)$51.4B$195B to $205B in 2026, and "increase significantly" in 2027Nine-month depreciation $22.0B, up 45%Servers and network equipment six years
Meta (Q3, to Sept 30)$34.9B including finance lease principal$135B to $145B in 2026Quarterly D&A $7.2B, up 44%Most servers and network assets 5.5 years
Oracle (fiscal Q1 FY27, to Aug 31)$28.5BNo capex figure in the releaseQuarterly depreciation $3.2B, up 134%Not broken out

Two details matter for reading the trend. Microsoft says its longer data-center life, covered below, will shift more future data-center leases from finance to operating leases. Finance leases count as capex and operating leases don't, so Microsoft's reported capex will capture less of its build than before. And depreciation lags capex. Oracle spent $28.5 billion in its quarter against $3.2 billion of depreciation; most of the cost of what it is building has not reached its income statement yet.

What do the companies say about demand and margins?

Revenue growth, backlog and supply constraints are covered in negotiating cloud commitments when capacity is tight. For cost structure, the margin and cash signals matter more.

  • Microsoft. CFO Amy Hood said "customer demand continues to exceed available supply." Microsoft Cloud gross margin fell to 66% and is guided to about 65% next quarter, which Microsoft attributes to the mix shift to Azure. Hood also said revenue per token in the quarter was higher than Microsoft expected, and that token pricing could add variability.
  • Alphabet. CFO Anat Ashkenazi said Google expects near-term margin headwinds as it uses third-party compute capacity.
  • Amazon. Trailing-twelve-month free cash flow turned to an outflow of $42.4 billion as purchases of property and equipment rose, an increase Amazon says primarily reflects AI investment.
  • Oracle. Free cash flow was negative $5.4 billion in the quarter, which Oracle attributes to investment in its Cloud Infrastructure business.

Meta doesn't sell cloud capacity to enterprises. In our view its spending still matters to buyers, because it competes for the same GPUs, memory and power.

Why does depreciation drive the price of AI capacity?

Because most of the money goes into short-lived assets. Microsoft said roughly three quarters of its quarterly capex was for short-lived assets, mainly CPUs and GPUs. Alphabet said the vast majority of its $51.4 billion went to technical infrastructure, and about 65% of that technical-infrastructure investment went to servers and networking equipment with six-year lives, with 35% for data centers with lives of 7 to 40 years.

By our calculation, Alphabet's Q3 server and network spend was up to about $33.4 billion, which adds up to about $5.6 billion a year of straight-line depreciation for six years. Each later quarter at that pace stacks another layer on top.

Here is how useful life and utilization change the cost floor for one machine. The server price is illustrative, not a quote:

Illustrative 8-GPU server, $400,000Depreciation per yearPer GPU-hour at 100% usePer GPU-hour at 60% use
Six-year life$66,667$0.95$1.59
Five-year life$80,000$1.14$1.90
Three-year life$133,333$1.90$3.17

These figures cover depreciation only, before power, buildings, network, staff and cost of capital. For scale, AWS publishes a Capacity Block rate of $5.970 per H100 accelerator-hour; the gap between a figure like that and depreciation is what has to pay for everything else, including margin. A fleet at 60% utilization must recover two-thirds more per hour sold than one running flat out. And if the hardware's economic life turns out shorter than its accounting life, the cost per hour rises sharply.

Are useful-life choices changing the numbers?

Yes, in both directions, and the amounts are material. Useful life is a management estimate, and the filings show companies revising it.

  • Meta extended most servers and network assets to 5.5 years from January 1, 2025. On assets in service at the end of 2024, that cut 2025 depreciation by $2.92 billion.
  • Amazon shortened a subset of servers and networking equipment from six years to five from the same date, citing the pace of technology development, particularly in AI and machine learning. That added $1.4 billion of 2025 depreciation, mostly in AWS.
  • Microsoft extended data centers and office buildings from 15 to 25 years from the start of fiscal 2027, which it expects to bring a minimal benefit to fiscal 2027 operating income.

On the illustrative table above, the gap between a five-year and a six-year life is about 19 cents per GPU-hour at full use and 32 cents at 60%. These revisions don't change cash spent. They change when the cost shows up, and so the cost base a provider prices against. In our view, a buyer should read Amazon's shortening and Meta's extension as two different judgments about how long AI hardware stays competitive. This describes what the filings disclose, not how any company should account; confirm accounting questions with your advisers.

Which price moves have providers already made?

Three are published. In our view, all three point to providers charging more for scarce AI capacity and locking in usage. Terms as of November 2026:

  • AWS EC2 Capacity Blocks. AWS says reservation prices are updated regularly based on supply and demand, with the next update scheduled for January 2027. Its own pricing example, dated March 2025, uses $31.464 per instance-hour for a p5.48xlarge (8 H100 GPUs) in US East (Ohio). The published rate today is $47.757, about 52% higher. These are two data points, not a series, and AWS can move the rate either way at the January 2027 update. Blocks are charged up front.
  • Google Gemini models. Introductory pricing ends on schedule. Promotional rates, given as 50% credits on net spend, end November 30, 2026 for Gemini 4 Argon and Gemini 4 Carbon, and December 31, 2026 for Gemini 3.8 Flash and others; Cyber Permissive variants are excluded. For Gemini 3.8 Flash (global, prompts up to 200K tokens), introductory pricing of $0.75 and $3.75 per million input and output tokens becomes standard pricing of $1.50 and $7.50 from January 1, 2027. A separate 50% credit on Gemini 3.8 Flash provisioned throughput also ends December 31.
  • Azure reservations. From February 1, 2027, new reservations for services that savings plans cover can't be exchanged, which Microsoft says gives customers a clear, predictable framework for choosing commitments.

None of these is a list-price rise on general compute. In our view, they show where the pressure is first released: in capacity that is scarce, in promotions that end, and in commitment terms that shift utilization risk to the buyer.

What does this mean for the prices enterprises pay next?

What follows is our view, not a forecast from any of the companies.

  1. Scarce AI capacity stays expensive while demand outruns supply. Neither Microsoft nor Alphabet gave an end date; Microsoft said some delayed capacity moves into the first half of calendar 2027. Expect reserved GPU capacity and frontier-model tokens to hold or rise, while older GPUs and general compute see less pressure.
  2. Promotions are a price signal. When a provider launches a model at half price for a set period, plan on the full rate from the end date. Date-stamp every AI unit price in a business case.
  3. Commitments will carry more of the utilization risk. Providers carrying fixed depreciation want predictable demand. Upfront payment, fewer exchanges and capacity reservations all move idle-capacity risk to the customer. Before buying a committed use discount or a capacity reservation, check break-even utilization, as in our post on AI capacity reservations.
  4. The turn will show in the filings first. We'd watch three signals: "demand exceeds supply" dropping out of earnings calls, depreciation growing faster than cloud revenue, and the Capacity Block price update in January 2027. If utilization falls, we'd expect price cuts on capacity and harder pushes for long commitments.

For planning, we suggest modeling AI and cloud spend with an explicit price scenario, not a single rate, as described in forecasting volatile cloud spend. A price assumption only helps if it sits next to the workload, owner and funding decision it affects, which is the case for keeping them in one operating model.

Sources

Frequently asked questions

Will hyperscaler AI capex make cloud prices go up?

For scarce AI capacity it already has, in places. AWS's published Capacity Block rate for an H100 instance in US East (Ohio) is about 52% above the rate in its own March 2025 example, and Google's introductory Gemini pricing ends on schedule in late 2026. General compute list prices have moved less. In our view, the direction depends on whether utilization of the new capacity stays high.

How much did hyperscalers spend on capex in Q3 2026?

On each company's own measure, Amazon reported $77.3 billion of property and equipment purchases, Alphabet $51.4 billion, Microsoft $45.2 billion including finance leases, Meta $34.9 billion and Oracle $28.5 billion in its quarter to August 31. That is roughly $237 billion, though definitions differ, so the total is approximate rather than like-for-like.

Why do server useful lives matter to enterprise buyers?

Useful life sets how fast a provider expenses its hardware, and so the cost base it prices against. On an illustrative $400,000 GPU server, moving from a six-year to a five-year life raises depreciation by about 19 cents per GPU-hour at full use. Amazon shortened some server lives in 2025, while Meta extended its own, showing the judgment cuts both ways.

Does a longer server useful life make cloud cheaper for buyers?

Not directly. A longer life spreads the same cash cost over more years, so reported depreciation per year and per hour falls. Meta's move to 5.5 years cut its 2025 depreciation by $2.92 billion without changing what it spent. Whether a provider passes any of that into prices is its choice. In our view, utilization and hardware turnover matter more to buyers.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.