Editorial Article
The $700 Billion Is Not a Building Count
Three of the four numbers changed this year, and only one of them was about capacity
At their July 2026 earnings calls, the four largest hyperscale operators guided to somewhere between USD 720 billion and 745 billion of combined capital expenditure for the calendar year — Amazon around USD 220 billion, Alphabet USD 195–205 billion, Microsoft about USD 175 billion including finance leases, Meta USD 130–145 billion. Bloomberg put the April figure at as much as USD 725 billion, “primarily on AI data center equipment.”
Read as a demand signal for anyone who builds, cools or powers data centres, that is the largest number in the industrial economy. Read as a measurement, it is unstable in a way that matters more than its size. Three of those four guidance figures moved during 2026, and only one of the moves was about building more.
Amazon went up, from about USD 200 billion to about USD 220 billion, and Andy Jassy gave the reason on 30 July: the higher cost of memory. Alphabet went up because, in CFO Anat Ashkenazi’s words, capacity is being delivered faster to meet demand. Microsoft went down — from roughly USD 190 billion in April to about USD 175 billion — because it extended the assumed useful lives of its data centres and office buildings from 15 years to 25, which pushed a set of data centre leases out of finance leases and into operating leases. Microsoft said its underlying plans were unchanged.
So the same year contains a capex increase caused by a memory price, a capex increase caused by schedule, and a capex decrease caused by an accounting assumption. Only the middle one is a statement about how much building is happening.
What the dollars actually buy
This is not a criticism of the companies; it is a warning to anyone who sizes a market from their guidance. Capital expenditure at this scale is a contents index before it is a construction index. The marginal dollar has moved from concrete, steel and switchgear toward accelerators, high-bandwidth memory and the networking that binds them, because that is where the performance constraint sits.
Our own forecasts show the physical layer growing at a completely different rate. The global data centre cooling system market — air and liquid together — is USD 18.7 billion in 2025 rising to USD 39.6 billion in 2033 at 9.8%, with precision air conditioning still the largest single segment. Liquid cooling specifically is USD 4.3 billion to USD 10.2 billion at 11.4%.
Set 9.8% against the growth in that capex line and you are looking at two different businesses. Double-digit growth in thermal equipment is a good business. It is not the business that a compound annual growth rate of thirty-odd percent implies, and the difference is not a forecast error — it is what happens when the majority of the incremental dollar goes into silicon rather than into the building that houses it.
The definitional problem, stated precisely
The four numbers are not comparable with each other, and the companies say so. Amazon quotes cash capex. Microsoft and Meta include finance leases — Meta counts principal payments on finance leases. Alphabet reports a capex line without stating how it treats leases.
That is not pedantry, because lease classification is exactly the lever Microsoft used. Extending a building’s assumed life from 15 years to 25 is a substantive economic judgement — it asserts that the shell will still be earning in 2051 — but its immediate effect on the reported number is to move assets between two lines. A supplier benchmarking “capex per megawatt” across the big four is therefore comparing cash spend, finance-lease-inclusive spend and an undisclosed lease treatment, and then drawing conclusions from the residual.
The practical rule: treat hyperscaler capex guidance as a statement of intent and pricing power, and treat megawatts commissioned as the demand signal. The first tells you how much the buyers are willing to spend; the second tells you how much equipment they need.
Why the megawatt count is the harder constraint
If capital is not the binding constraint — and at these balances it isn’t — then what binds is the things capital cannot accelerate: an interconnection agreement, a transformer with a multi-year lead time, a gas turbine allocation, a water permit, a skilled electrical contractor, and a rack of accelerators that someone else has also ordered.
Those constraints do not appear in a capex line at all. They appear as a delay between the announcement and the revenue, which is why the same companies can raise capex guidance and still report that capacity is arriving late. And they explain the shape of the equipment market we forecast: high-single-digit growth in cooling and power equipment is not evidence of weak demand, it is evidence that the demand is rationed by the physical chain rather than funded by the balance sheet.
That reframing changes who is investable. A supplier whose product is on the critical path of a megawatt — switchgear, busway, transformers, chillers, CDUs, on-site generation — is selling into a queue and can hold price. A supplier whose product scales with the dollar value of the building’s contents, rather than with the megawatt, is selling into a procurement process that is actively trying to commoditise it, because the buyers at this scale design their own hardware and dual-source everything.
What would change this reading
- Capex guidance falls while megawatts commissioned rises. That would prove the capex number is a price index and the physical build is intact — the cleanest confirmation of the thesis.
- Memory and component prices normalise. If Amazon’s USD 20 billion of memory inflation reverses, capex could fall without a single project being cancelled. Watch the capex line and the megawatt line diverge in opposite directions.
- A useful-life reversal. The 15-to-25-year assumption is a bet on the economic life of an AI-era shell. If a major operator shortens it again — because a generation of accelerators proves to have a shorter useful life than the building — depreciation rises, and the incentive to own rather than lease changes with it.
The number is a commitment device
Seven hundred billion dollars is not a forecast. It is a public statement by four companies that they intend to keep buying, made to an investor base they need to hold through a depreciation cycle. It should be read the way you read a central bank’s forward guidance: as an instrument with real effects and imprecise content.
What is measurable is downstream and much smaller. In 2025 the world spent USD 18.7 billion on cooling systems for every data centre on earth, and USD 4.3 billion of that on liquid cooling. Those are the figures a supplier can build a factory against. The seven hundred billion tells you the customer is serious; it does not tell you how many buildings you will be filling in 2028, and it cannot, because it was never built to measure that.
Frequently Asked Questions
How much are the hyperscalers spending on data centres in 2026?
At their July 2026 earnings calls, Amazon, Alphabet, Microsoft and Meta guided to a combined roughly USD 720–745 billion of capital expenditure for calendar 2026. The figures are not directly comparable: Amazon quotes cash capex, Microsoft and Meta include finance leases, and Alphabet does not state its lease treatment.
Why does the data centre cooling market grow slower than hyperscaler capex?
Because the marginal dollar is going into accelerators, memory and networking rather than into the shell. Amazon raised guidance from about USD 200 billion to about USD 220 billion explicitly because memory cost more, and Microsoft said about USD 25 billion of its April figure reflected higher component prices. The physical equipment market grows at high-single-digit rates while the capex line grows far faster.
What is the right demand signal for data centre suppliers?
Use megawatts commissioned, not dollars guided. Capital is not the binding constraint for these buyers — interconnection, transformers, turbines, permits and accelerator allocation are. Those constraints appear as delay between announcement and revenue, not in a capex line, so they explain why guidance can rise while capacity is delivered late.
Why did Microsoft's capex guidance fall during 2026?
Microsoft's guidance moved from roughly USD 190 billion in April to about USD 175 billion after it extended the assumed useful lives of data centres and office buildings from 15 to 25 years, which moved a set of data centre leases from finance leases into operating leases. The company said underlying plans were otherwise unchanged, so the reported reduction was a classification effect rather than a cancellation.