Where Does $100 of AI Capex Go? Power Is the Line Item That Gates Everything Else

A recent capex breakdown from BNP Paribas Equity Research maps out where every $100 of AI infrastructure spending actually goes — semiconductors, networking, power, cooling, facilities. The semiconductor share gets most of the headlines, at roughly half of total spend. The number that should get more attention sits further down the chart: power takes roughly $20 of every $100 — more than networking and cooling combined.

THE NUMBER MOST PEOPLE UNDERRATE

Within that power allocation, grid connection and independent generation alone account for roughly half — the single largest sub-category, ahead of power distribution equipment. Cooling, by comparison, sits at around $7.50 of every $100, and networking equipment at around $15. Power isn't a secondary line item behind the glamour of chips — it's the second-largest category in the entire capex stack.

WHY THIS LINE ITEM IS DIFFERENT FROM THE OTHERS

Most of a capex breakdown scales the way you'd expect: order more chips, get more chips, roughly on the manufacturer's quoted timeline. Power doesn't behave that way. Turbines, transformers and substation equipment carry multi-year lead times that don't compress just because a buyer is willing to pay more or order more units — the constraint is manufacturing capacity and grid interconnection queues, not budget.

That asymmetry is what makes power the gating factor for the other $80. A completed data hall full of racked GPUs with no interconnection secured isn't a delayed asset — it's a very expensive building with nothing running in it. Every other line item in the breakdown assumes power arrives on schedule; power is the one line item nothing else can compensate for if it doesn't.

WHERE SMRs FIT — AND WHERE THEY DON'T CHANGE THE PICTURE

Part of the current interest in small modular reactors (SMRs) comes down to a scale-matching problem: a conventional nuclear plant runs at roughly 1 GW, far more than a single hyperscale campus needs, which means selling surplus back into a grid the buyer now depends on anyway. SMRs, landing in the 50–300 MW range, map much more closely to a single campus's load — which is part of why recent deal activity has skewed toward SMR agreements rather than new full-size reactors.

That said, SMRs don't remove the equipment bottleneck — they're still built around turbines, transformers, switchgear and grid interconnection work, all subject to the same multi-year lead times covered elsewhere on this site. A different generation source doesn't change the procurement math on the equipment that connects it to load.

WHAT THIS MEANS BEYOND DATA CENTERS

The same $20-of-$100 dynamic explains why the equipment categories behind power — transformers, turbines, generators, switchgear — are now contested between data center developers, industrial buyers, BESS projects and the EPC contractors building grid infrastructure. It's not that power is one line item among several in the AI buildout. It's the line item that determines whether the other 80% of that $100 gets to do anything at all.