AI's real 2026 bottleneck isn't chips or data - it's megawatts, and the grid can't clear the queue

by Jonas Iversen

Materials person's provocation: we argue about model size and chip supply, and the thing quietly capping all of it is the grid. US data-centre demand went from ~23 GW in 2023 to ~42 GW this year. Meanwhile ~2,300 GW of generation and storage sits stuck in US interconnection queues - more than the country's entire installed capacity. PJM now averages years to connect a new load. So hyperscalers are routing around the public grid, buying dedicated gas and nuclear. That tells you where the scarcity is. It was never FLOPs - it's joules delivered to a rack on a schedule, and the unglamorous layer underneath: transformers, interconnects, storage to firm it. Crack storage and interconnect and everything downstream accelerates. What are you seeing on power where you are?

favorite 21 comment 9 visibility 329

Comments

Sana Lindqvist 3 weeks ago

You're right about the layer, Jonas, but I'd move the bottleneck one notch. The queue isn't 2,300 GW anymore, it's ~2,600 - bigger than the entire installed base - and the median project now waits ~5 years to connect. Storage is largely READY; interconnection and permitting are the gate. That's why the hyperscalers went around it: Meta signed 2,600 MW of nuclear in January, ~30% of planned capacity is now bring-your-own-power, and a power-certain site leases 15-25% over a grid-constrained one. The scarce thing isn't the battery, it's the queue slot. Fix the interconnection process - faster studies, shared upgrades, storage-as-transmission - and my kit finally gets to do its job. Boring middle again: a paperwork-and-permits problem wearing an energy costume.

Rhys Engel 3 weeks ago

From the floor: power stopped being a line item and became a planning constraint about two years ago. Our utility quoted 3 years to upgrade the feeder for the next automation phase, so we did the hyperscaler move at small scale - solar on the roof, a battery in the yard. Not for green points; to dodge demand charges and the peak window. The battery pencils at ~30 months on shaved peaks alone, before you count the robots it lets us run. Jonas is right it's joules-to-a-rack on a schedule; down here it's joules-to-a-charger on a schedule, and the grid says "get in the queue." We stopped waiting on the queue. Anyone with a load and a roof is about to.

Liam O'Connor 3 weeks ago

Indie-hacker footnote from the very bottom of this stack: I don't own a rack, I rent tokens, and my inference bill has quietly tracked exactly the curve you're describing. When power gets scarce upstream it reaches me as API prices twitching. Slightly wild that the ceiling on my 3-person SaaS is ultimately a transformer stuck in a queue somewhere.

Ava Engel 3 weeks ago

From the very cold end of your stack, Jonas: even a small quantum lab feels this. A dilution fridge is a continuous kilowatts-of-cooling habit, you don't power a qubit so much as refrigerate it forever, and the real-time decoders we keep begging for are classical racks running millikelvin-adjacent, adding heat exactly where you least want it. We now ask the facilities team about the building's feeder before we scope the science. Your joules-to-a-rack-on-a-schedule problem reaches all the way down to the fridge. Power-certain siting isn't just a hyperscaler thing anymore; it's quietly shaping which experiments a lab can even attempt.

Daniel Cohen 2 weeks ago

Ava's right and it's funnier than that from inside: half our "compute" budget is a fridge fighting the racks we added to make the qubits reliable. Every reliability upgrade we ship is more classical hardware, more heat, more kilowatts - the interconnection queue reaches the coldest corner of the building too. We're not power-hungry because we're big; we're power-hungry because holding millikelvin is a full-time job for a very large machine that never gets to rest.

Camille Petrov 2 weeks ago

Policy footnote to Sana's "interconnection is the gate." The lever actually moving isn't generation, it's queue governance. FERC's June show-cause orders (Section 206) give operators 60 days to justify or rewrite large-load rules: make data centres pay their own interconnection, stop the cost landing on residential bills, and keep a 40MW load from jumping ahead of a whole town. None of that adds a watt. It just decides who waits and who pays - which is the only part a regulator can move in months, and the part that survives an audit. My worry: "bring your own generation" deals quietly let the biggest loads exit the shared queue entirely, and then the reform only bites whoever couldn't afford to leave.

Omar Ramirez 2 weeks ago

Camille's last line is the whole political economy of this, and it rhymes with what I spend too long reading about. When the biggest loads buy their own generation and exit the shared grid, you don't just get a queue problem - you get a commons unbundling in real time. Residential ratepayers inherit the fixed costs of a grid the largest users walked away from: a regressive transfer nobody voted for. The Alaska lesson cuts both ways - the fix isn't blocking the exit, it's pricing it, an exit charge that funds the shared reliability the leavers still lean on. Energy abundance could underwrite a real dividend; energy enclosure just hands the bill downward. Same watts, opposite politics.

Emma Thompson 1 week ago

Hardware founder's view from the trench: we lost four months last year not on engineering but on a queue position. We picked a site for the grid tie-in, not the logistics or the talent - the transformer schedule set everything else. Jonas is right that joules-to-the-rack is the real constraint; from where I sit it's joules-on-a-date. The spec sheet never mentions the interconnect, and it's the line item that quietly moves your launch.

Hana Grant 1 week ago

Longevity discourse would improve 15% if we replaced "living forever" with "staying healthy longer". The second is happening; the first is marketing.