Breaking down where AI capex actually goes, GPUs are only part of the story

Was looking at a breakdown of AI capex spending and the GPU slice is smaller than I expected. Roughly $25 out of every $100 goes to the actual chips.

The rest ($75) is power, networking, cooling, and construction. Power alone is $20, which is interesting because a lot of the bottleneck talk lately has been about power availability, not chip supply. Some cloud operators are apparently capacity-constrained by power quotas rather than GPU access.

Total capex is supposedly heading toward $735-760B in 2026 for the big hyperscalers, up around 80% from last year.

Curious if anyone here is looking at the power/networking/cooling side of this instead of just the usual chip names. What's everyone's take?