FERC Just Sold AI Data Centers a Fast Lane — Read the Curtailment Clause

FERC Just Sold AI Data Centers a Fast Lane — Read the Curtailment Clause

Here is the most honest sentence a regulator has said about this buildout in two years. Laura Swett, the chair of the Federal Energy Regulatory Commission, went on CNBC Monday to talk about AI data centers and the American grid — and the headline was not about chips, not about capex, not about scaling laws. It was this: the U.S. power grid still lags the surging demand coming from AI data centers.

FERC is the referee for every wholesale power market that matters in this country — PJM, MISO, SPP, CAISO, ISO New England, NYISO. When the person running that agency says the grid lags, she is not giving a forecast. She is describing the thing her own agency has been trying to fix since June, and conceding that three months in, the fix is not finished.

I have been running hosting infrastructure for over a decade, and I have learned one thing about regulators: when they start talking about speed, read the contract. The fast lane FERC built for AI data centers has a brake pedal bolted into it, and most of the people lining up to use it have not read that part.

What Swett Actually Told CNBC, and Why the Date Matters

The interview aired Monday. Her message had three parts: reform the process so these loads can actually connect, protect ordinary ratepayers from the cost of doing it, and lean on American gas exports to keep global energy markets stable. That third piece is its own story — the grid at home is short and the fuel is being shipped out.

But the first two parts are not new talking points. They are the follow-through on June 18, when FERC issued six simultaneous show cause orders under Section 206 of the Federal Power Act — dockets EL26-67 through EL26-72 — to every regional grid operator it regulates, covering roughly two-thirds of the country.

The preliminary finding was blunt: those tariffs appear to be unjust and unreasonable, because they do not adequately address how large and co-located loads connect to the transmission system. FERC defines a large load as a single-site customer with peak demand of 50 megawatts or more, interconnecting above 69 kilovolts, and not part of a co-location arrangement.

Then came the clock. Thirty days for each operator to report on whether enough generation will exist to serve these loads. Sixty days to justify the existing tariff or file reforms — that deadline landed August 17. And a window for abeyance, up to 90 days, for operators that wanted to negotiate with their own stakeholders first.

That is why Monday's interview matters. The reports are in. The responses are in or delayed. And the chair is still standing in front of a camera saying the grid lags. You do not need a law degree to read that. The paperwork moved. The electrons did not.

The Fast Lane FERC Built Is Not a Firm Lane

This is the part nobody put in the headline, and it should decide your capacity strategy for the next two years.

FERC did not tell grid operators to hand data centers firm, guaranteed, always-on power on an expedited basis. It told them to create new categories of service — and some of those categories are interruptible by design. In PJM, the products on the table include interim network transmission service that is non-firm and subject to curtailment ahead of system emergencies, a firm contract demand service carrying a penalty rate if you withdraw above your contracted level, and a non-firm contract demand service sold in blocks as short as an hour to a month that gets curtailed during emergency operations.

Meanwhile, one of the five reform categories FERC directed at every operator is titled, in plain English, transmission service for flexible large loads — customers willing and able to limit how much they pull off the transmission system under certain conditions. FERC's own language: loads that can "reduce, limit, or interrupt withdrawals in response to system conditions or price signals," with automated demand response, real-time telemetry, and communication protocols wired in so the operator can verify it.

Out west, SPP already built the template: service delivered subject to curtailment and interruption, maximum term of seven years, designed to push customers toward bringing their own resources or completing the network upgrades required for long-term firm service. There is a $25,000 study deposit the host transmission owner can require just to run the load connection study.

Translate that. You can connect faster. You can connect cheaper. And in exchange, the grid keeps the right to reduce your load — with the switch on their side of the wall.

Two Ways to Read This — and Both Are True

The bullish reading: this is the most important regulatory action for American AI capacity this year. FERC ordered standardized applications, rolling acceptance instead of cluster queuing, study timelines of 60 to 90 days, interim service so a facility can energize before the last upgrade is finished, and a pro forma service agreement so a developer can underwrite a schedule instead of guessing. Nvidia called it "a win for ratepayers, grid reliability, and American competitiveness," and on pure speed-to-power they are not wrong. For two years the binding constraint on every project I know about has been the queue, not the silicon.

The founder's reading: the lane is real, but it is not firm, and the cost side has been nailed down in your direction. The same orders push operators toward cost recovery agreements requiring large-load customers to make a minimum financial contribution to the transmission owner's revenue requirement, backed by credit support. Attorney commentary on the orders describes financial security sized to the expected cost of energy to serve the project over a term that often runs ten years — plus readiness obligations, withdrawal penalties, and protections against stranded costs. If the project slips, shrinks, or never shows up, the developer absorbs it. Not the ratepayer.

That is a good deal for the public. It is a dangerous deal for anyone signing it on a demand forecast they do not control.

The Clause That Decides Who Eats the Loss

The scarce input in 2026 is no longer megawatts. It is firm megawatts — capacity the grid cannot take back. FERC has created a two-tier market: a fast, cheap, interruptible tier and a slow, expensive, guaranteed tier, and it told operators to design contracts that reward flexibility. Rational response to a real problem. It also means the contract you sign now decides whether your facility is a data center with a grid connection, or a demand-response asset with a GPU problem.

The politics are not on your side either. Utilities requested $18.6 billion in rate increases in the first half of 2026, including a record $9.2 billion in the second quarter. PJM's own market monitor attributed 63% of the 2025/2026 capacity auction increase to data center demand — roughly $9.3 billion shifted onto ratepayers. PJM's 2028/2029 auction cleared at the $325 per megawatt-day price cap and still fell about 6.8 gigawatts short of its reserve margin target. Seventy-one percent of Americans now say they oppose data centers in their area, up from 42% a summer earlier. Texas halted new grid connections in August.

So when the regulator says the grid lags and the cost must land on the load, understand what is actually being decided: who holds the risk when the demand curve bends. It was the same Monday that AI-linked equities sold off worldwide on warnings from the very CEOs building this thing. Separate story, its own coverage. But look at the timing — for anyone carrying a ten-year financial security obligation attached to a speculative load, the two stories are one story.

What I'd Do Before Signing One of These Contracts

First, read the service product, not the megawatt number. Ask which tariff service your capacity sits under. If it is interim or non-firm, you have a curtailment obligation in writing, and your uptime promise now depends on a party you do not control.

Second, price the curtailment into the SLA. Either pay for firm capacity or sell a product that tolerates interruption — checkpointed training jobs, batch queues, spot inference. Do not sell 99.99% availability on top of an interruptible feed. That is not hosting. That is a bet against a control room operator in Valley Forge.

Third, model the credit support on day one, not at closing. Minimum revenue contributions, financial security sized to expected energy cost over a term that can run ten years, withdrawal penalties, readiness obligations. On a delayed project that is a liability before it is ever an asset. Underwrite it like debt, because that is what it is.

Fourth, stop giving flexibility away for free. FERC is building a market where curtailment, load shifting, and verifiable telemetry have value. If your facility can genuinely reduce load on demand, that is a revenue line and a negotiating chip — not just an obligation you signed to get to the front of the queue.

Fifth, watch the calendar, because the rules are being written right now. CAISO is targeting a November compliance filing after requesting abeyance. MISO is still working its large-load agenda into the winter. PJM — the market FERC expects to host as much as 70% of future US data centers, with roughly 30 gigawatts of new demand projected by 2030 — is still drafting tariff language. Comment windows are where influence is cheap. Hearing rooms are where it is expensive.

The Bottom Line

The grid did not get bigger this week. It got more honest, and honesty is worth something, because you can plan against it. What you cannot plan against is a mythology — that a signed interconnection agreement means firm power, that a queue position means a delivery date, or that a headline about speed means the switch stays up when the system is stressed.

The regulator has now said, on the record, that the grid lags the demand. Heard as a founder, that is not bad news. It is the most valuable sentence in the market, because it tells you exactly what to buy: not megawatts, not rack space, not another region. Firm capacity, and the contract language that proves it. Everything else is a promise with a curtailment clause stapled to the back.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: CNBC International, "U.S. power grid still lags surging demand for AI data centers, says FERC" (Laura Swett interview, September 14, 2026); FERC news release and Section 206 show cause orders, Docket Nos. EL26-67 through EL26-72 (June 18, 2026); FERC Chairman Swett's remarks to POLITICO; legal analyses from Holland & Knight, K&L Gates, White & Case, Orrick, Bracewell, Sidley, Vinson & Elkins and Husch Blackwell on the large-load orders and the August 17 show cause responses; PJM Interconnection capacity auction results and Monitoring Analytics market monitor findings; PowerLines rate request data; Network World moratorium survey; Utility Dive.