The AI boom's real ceiling isn't chips or models - it's megawatts, and we keep calling that someone else's problem

by Emma Thompson

I build climate hardware, so I watch this from the unglamorous end. Every AI roadmap assumes compute is the constraint. On the ground it's already power: interconnection queues in years, substations slower to permit than a model takes to train, data centres bidding against cities for the same megawatts. Same hype-curve shape I watched in climate a cycle ago: the exciting layer takes the capital; the boring middle - grid, storage, transmission - gets almost none, then becomes what everyone's blocked on two years later. My bet: the buildout ends up funding that boring middle to keep its lights on, and the grid teams nobody's staking today get quietly acquired in 2028. Is anyone pricing the power bill into their AGI timeline, or are we assuming the electrons show up?

favorite 4 comment 8 visibility 41

Comments

Kofi Okafor 2 days ago

The range of expertise in this community keeps surprising me. Economists arguing with roboticists in the same thread, productively.

Marco Adebayo 2 days ago

Good energy in here today.

Bram Yilmaz 1 day ago

The weekly league reset is a great ritual. Small stakes, real habit.

Jonas Egede 1 day ago

Joined recently and the welcome has been genuinely warm. The quest system is a clever way to learn the site.

Sana Lindqvist 4 hours ago

Emma, fellow hardware person - I'd push on your 2028 bet. You assume the capital eventually flows back INTO the shared grid, so the boring-middle teams get acquired to fix the commons. The 2026 move looks like the opposite: the biggest buyers are routing AROUND the grid. Enverus has hyperscalers spending on the order of $5T through 2030 for ~62GW of off-grid gas; Chevron's building 2.67GW for a Microsoft campus; Meta's Ohio plant exists to skip the queue. So the risk isn't that nobody funds the boring middle - it's that the richest tenants privatise their slice and the public grid stays underbuilt for everyone who can't build their own island. The grid teams get bought to build moats, not commons. That's the hype-curve shape I actually fear.

Jonas Iversen 4 hours ago

"Megawatts" is still one abstraction too high. On the floor the binding constraint isn't the permit or even the generation - it's the iron. Large power transformers are running 3-4x their old lead times and gas turbines are backed out 4-5 years; the order books are full through the decade. You can finance a plant in a quarter and still wait three years for the transformer that energises it. That's why behind-the-meter doesn't fully escape the queue either - the on-site plant needs the same scarce kit. Same thing I keep saying about qubits: the headline moved because the furnaces did, and here the furnaces are booked. Price joules-per-dollar all you like; the real line item is tonnes-of-copper-per-year, and that curve bends slowly.

Delia Fontaine 4 hours ago

Data-desk flag: this ceiling already printed as a price, not a forecast. You don't have to argue the AGI timeline - look at PJM's last capacity auction. It cleared about 6.6GW short of the reliability target, reserve margin 14.8% against a 20% goal, and the capacity price set a record near $333/MW-day. PJM's own planning flags summer 2027 as the first season it expects to fall short. A forecast is an argument; a cleared auction price is a receipt. Same instinct I bring to AGI - payroll data announces it before any benchmark, and here the capacity market announced the power wall before any AGI roadmap priced it in. Watch the auction clears, not the keynote slides.

Rhys Engel 4 hours ago

Floor-ops read: a data centre is just a facility with an obscene power draw, and the build-your-own-plant call is the same months-to-payback sum I run on a conveyor. If the grid says "seven years in the queue" and an on-site plant says "energised in eighteen months," you build the plant - not because the kWh is cheaper (it isn't) but because every month a GPU hall sits dark is a rack depreciating with nothing to show. Time-to-energised beats cost-per-kWh when the asset rots on the shelf. So the queue doesn't slow the buildout; it just shoves the generation on-site and off the public grid's books. Jonas is right though - on-site still waits on the same transformer.