The interesting number isn’t the $1.1 trillion. It’s what happens when you put the two figures next to each other. Google, Amazon, Microsoft and Meta have spent $1.1 trillion on capex since the AI boom began in 2023. They plan to spend $745 billion of it this year. So one calendar year is going to come close to matching the three that preceded it, combined.
That’s the whole story, really. Everything else is detail.
The detail still matters though. Amazon guided to $200 billion and hasn’t moved off it since February. Microsoft is tracking around $190 billion for the fiscal year. Alphabet raised to the $175 to $185 billion range. Meta started at $115 to $135 billion and has been creeping toward $145. Add it up and you get roughly 77% growth over the $410 billion the four of them spent in 2025, which itself was a record that nobody thought would be beaten so quickly.
Capital intensity is now at levels that would have been laughed out of an investment committee three years ago. Meta is spending something like 54% of revenue on capex. Microsoft is in the high forties. Alphabet is close behind. Amazon looks restrained at about 25% only because retail revenue is enormous and dilutes the ratio.
Here’s where it gets uncomfortable. Until recently this was all funded out of operating cash flow, which was the entire reason the market tolerated it. That’s over. Alphabet’s long-term debt roughly doubled in the first half of this year to around $98 billion, the company priced an $84.75 billion equity raise in June, and it went free cash flow negative in the second quarter for the first time anyone can remember. Amazon’s long-term debt jumped 81% in a single quarter to $119 billion. Analysts have Microsoft’s free cash flow going negative in its fourth quarter, which if it happens would be the first time since at least 2001.
Incremental debt has gone from something like 9% of hyperscaler capex two years ago to roughly a third of it now. The money is no longer coming from the business. It’s coming from the bond market, and increasingly from equity holders.
The other thing nobody talks about enough: a meaningful slice of this year’s increase buys no additional compute at all. Microsoft has said roughly $25 billion of its capex number is component price inflation. That’s memory, mostly. DRAM and HBM pricing has gone vertical because HBM production cannibalizes conventional DRAM wafer capacity, and the hyperscalers are the ones absorbing it. So when you see capex up 77%, delivered compute is not up 77%. Some of that dollar growth is just a transfer to Micron, SK Hynix and Samsung. Good for them. Less good for the people writing the checks and modelling the return.
Then there’s the depreciation problem, which is the one I keep coming back to. Roughly two-thirds of Microsoft’s capex goes into short-lived assets, GPUs and CPUs on five to six year schedules. Data center shells and substations depreciate over fifteen years or more, but the silicon does not. So $745 billion spent this year turns into well over $100 billion of annual depreciation expense flowing through income statements starting in 2027, and it keeps flowing whether or not the AI revenue shows up on schedule. Every hyperscaler has already extended useful life assumptions once. There isn’t much room to do it again without the auditors and the short sellers noticing at the same time.
The market has clearly worked some of this out. Alphabet raised capex guidance alongside second quarter results and the stock dropped 7% the next day, dragging the other three down with it. Two years ago a capex raise was read as a demand signal and the stock went up. Same announcement, opposite reaction. That’s a regime change, and it’s the single most important thing to have happened in this trade all year.
None of which means any of them will slow down. That’s the trap. Each of these companies has concluded, correctly from where they sit, that being short on compute is a far worse outcome than overspending. If Meta underbuilds and Google doesn’t, Meta loses the decade. So they all build simultaneously, and in doing so they take on the same risk at the same time in the same direction. Four companies, one bet, no diversification anywhere in the stack. If enterprise adoption disappoints even modestly, all four re-rate together, and so does everything they buy from.
For anyone positioned in the supply chain, the read is that 2026 revenue is already locked in. Orders are placed, lead times on purpose-built AI data centers are running eighteen to twenty four months, power contracts are signed. The debate isn’t about this year. It’s about whether the 2027 number that everyone is pencilling in at over a trillion actually gets funded, and by whom.
They can afford this year. Next year is the question.
