Microsoft, Meta, Amazon pledge $710B

Cloud spending used to sound abstract. Now it looks like a physical buildout on the scale of an industrial boom. Microsoft, Amazon, Meta, and Alphabet are all telling investors the same thing in 2025 — they are going to spend astonishing sums on data centers, servers, networking, and power to keep up with AI demand. Add the company guidance together and you get well over $300 billion for this year alone, not $710 billion. The bigger number floating around online seems to mix years, leases, and forward commitments into one pile. ### What are these companies actually spending? The cleanest way to frame it is company by company. Microsoft said in January it is on track to invest about $80 billion in fiscal 2025 on AI-enabled data centers. Amazon said in February that 2025 capex should be about $100 billion, with the vast majority tied to AI for AWS. Alphabet said in April it still expected about $75 billion in 2025 capex, then raised that to about $85 billion in its July 2025 earnings call. Meta started 2025 at $60 billion to $65 billion, then lifted guidance in April to $64 billion to $72 billion. ### So where does the $710 billion come from? Basically, it does not line up with the official 2025 guidance from those four companies. If you use the lower 2025 figures — Microsoft $80 billion, Amazon $100 billion, Alphabet $75 billion, Meta roughly $69 billion at the midpoint — you get about $324 billion. If you swap in Alphabet’s later $85 billion update, you get about $334 billion. That is still huge. But it is nowhere near $710 billion unless you start adding future-year estimates, finance leases, or other companies. (blogs.microsoft.com) ### Why are they spending this much? Because AI has turned cloud into a capacity race. Training frontier models eats enormous amounts of compute, but inference is the part that keeps the meter running every day. Once millions of users start querying models, generating images, or running agents, the bottleneck shifts from “can the model work?” to “can the infrastructure serve it cheaply and fast enough?” That is why these budgets are going into campuses, racks, networking fabric, and custom silicon — not just Nvidia GPUs. (blogs.microsoft.com) ### Why does power keep coming up? Because a data center is really an electricity conversion machine. Chips turn power into tokens. The catch is that AI clusters need dense, reliable power in places where utilities often cannot add capacity quickly. That is why the constraint is moving outward — from chips, to servers, to networking, to substations, transmission, and cooling. Meta has been talking openly about power and grid innovation, and Microsoft has made “community-first” infrastructure a public theme as it expands data centers. (aws.amazon.com) ### Does this mean margins get worse? In the short run, often yes. Heavy capex turns into depreciation, financing costs, and higher operating costs later. But these companies think the payoff is worth it. AWS, Azure, Google Cloud, and Meta all want to own the base layer that everyone else rents. If AI really becomes a default computing layer, the winners are the firms that built enough capacity before everyone else needed it. (about.fb.com) ### Why not just wait for cheaper chips? Because waiting can be more expensive than overspending. If demand outruns supply, customers go elsewhere, model launches slip, and developers build on a rival cloud. That makes this feel less like discretionary spending and more like laying railroad track before the freight arrives. ### What’s the bottom line? The important correction here is simple — the public 2025 numbers support a story of roughly $330 billion in hyperscaler capex, not $710 billion. (aws.amazon.com) But the broader point still lands. AI is no longer just a software story. It is an infrastructure story — and increasingly an energy story.