Where Big Tech's AI capex goes, from chips to buildings and power

In the quarter ended Sep 30, 2026, four of the biggest spenders each put tens of billions of dollars into capital spending, using its own headline figure: Amazon $75.9B in cash capex, Alphabet $51.4B, Microsoft $45.2B and Meta $34.9B. The question this piece answers isn't whether that's too much. It's what the money physically buys, and how long it lasts.

All figures come from company releases and earnings calls dated Oct 21–29, 2026. Each company defines capex differently (Microsoft's figure includes finance leases, for example), so the four numbers shouldn't be added together or compared like for like.

The two halves of an AI build-out

Every AI data centre splits into two kinds of spending that age at very different speeds.

• The shell: land, buildings, electrical substations, transformers, backup power and cooling. This lasts decades.
• What goes inside: servers with their GPUs or custom accelerators, CPUs, memory and storage, plus the networking that ties thousands of chips into one cluster. This lasts years.

Alphabet gave the clearest split on its Oct 28 call. CFO Anat Ashkenazi said that "approximately 65% of our investments in technical infrastructure in Q3 was related to servers and networking equipment, which generally have useful lives of 6 years. The remaining 35% was for data centers, which have useful lives of 7 to 40 years."

Microsoft put it in similar terms the same day. CFO Amy Hood said "roughly 3/4 of our CapEx was for short-lived assets, primarily CPUs and GPUs, including higher memory and storage costs", with the rest going to long-lived assets.

Meta's release describes its $34.9B as "driven by investments in servers, data centers, and network infrastructure" without a percentage. Amazon said its $75.9B "primarily relates to AWS and AI".

Why the mix matters

A dollar spent on a building and a dollar spent on a GPU server hit the income statement very differently. Using Alphabet's stated lives as a rough guide, a server is depreciated over about six years, a building over as many as 40. So when most of the spend is servers, depreciation climbs fast and the gear needs replacing much sooner.

The companies say so themselves. Alphabet warned that higher infrastructure spending "will continue to weigh on the P&L in the form of higher depreciation expenses and related data centers operation costs such as energy". Amazon CEO Andy Jassy mentioned "significant depreciation from the CapEx we're investing", and said Amazon keeps working to "extend useful lives".

If you want the financial side of this, see our free cash flow explainer. Here we stay with the hardware.

Chips, and who makes them

The biggest single line inside the "short-lived" bucket is accelerators. Some companies increasingly use their own:

• Amazon said on its Oct 29 call that the majority of its Bedrock service runs on its own Trainium chips, that it is "installing an AI accelerator about every 14 seconds", and that it plans to sell Trainium outside AWS from late 2027.
• Alphabet sells TPU systems and said inventory built up for those sales weighed on its operating cash flow.
• Everyone still buys NVIDIA. Jassy noted that "everybody is producing their advanced chips with the same provider" and gets an allocation.

Racks are also getting more integrated and more power-dense. Our explainer on NVIDIA's DSX POD shows how a whole AI factory can now ship as factory-built containers.

Memory is part of the bill

Servers need a lot of memory: high-bandwidth memory (HBM) stacked next to the accelerators and conventional DRAM next to the CPUs. Prices for both have risen as AI demand soaked up supply.

That shows up directly in capex. Hood said Microsoft's short-lived spending included "higher memory and storage costs", and that the acceleration she expects later in the fiscal year is "inclusive of higher memory and storage costs". Jassy pointed to "heightened component costs like memory". Part of the capex jump is the same hardware costing more, not just more hardware. Our piece on why RAM is so expensive covers what that does to PCs and phones.

Spending that doesn't show up as capex

Headline capex leaves out several big commitments.

Finance leases. When a company leases a data centre long term, it may record the asset under a finance lease. Microsoft includes the "full principal value of finance leases" in capex when a lease starts: $7.4B of its $45.2B last quarter. Its cash paid for property and equipment was $39.4B. Meta reports capex including principal payments on finance leases. So the same building can look different depending on whose numbers you read.

Rented capacity. Companies short on their own capacity rent GPUs from others. Alphabet said it "started utilizing third-party compute capacity at the end of Q3" and expects near-term margin headwinds from it. Hood said "some third-party capacity deliveries we expected in Q1 were delayed", which moved spending into the second half of Microsoft's fiscal year. Meta CFO Susan Li said Meta has "augmented" its own build with "third-party cloud agreements, primarily in service of bringing on capacity in nearer time frames". Jassy argued the other side: renting from neoclouds is about "3,500 basis points worse in operating margin" than long-leasing, his own estimate.

Commitments. Alphabet said that at Sep 30, 2026 it had future purchase commitments of $924 billion and pre-commencement leases and other obligations of $152 billion, while noting these are maximum contractual amounts and some agreements have volume flexibility.

Power is what the money is racing for

All of this needs electricity. Amazon said its biggest customers reserve "hundreds of megawatts or sometimes up to a gigawatt of capacity" well in advance, with visibility into 2027, 2028 and some of 2029. Alphabet said its multiyear agreements cover "leases and energy".

The IEA estimated in April 2026 that data centres used about 485 TWh of electricity in 2025. Our IEA explainer walks through where that's heading and where the grid pinches.

What to watch next

• The mix. Does the short-lived share stay near 65–75%? A higher share means faster depreciation.
• Leases and rented capacity. These can grow even if headline capex flattens.
• Component prices. If memory keeps rising, the same number of servers costs more.
• Guidance. As of their late-October 2026 calls, Amazon expected about $220B of 2026 cash capex and "meaningfully more" in 2027; Alphabet kept its $195–205B 2026 range and expects a significant increase in 2027; Microsoft expects Q2 capex above $50B; Meta guided $135–145B for 2026.

The short version: most of the money goes into chips and servers that will need replacing within about six years, the buildings around them last far longer, and a growing slice of the build-out is rented or leased rather than bought.