FERC just gave AI data centers a fast lane to the grid…

Almost four weeks ago, we wrote that the AI infrastructure race has a hidden bottleneck that most enterprises are still under-pricing — and that it isn't GPUs. The argument then was that the network fabric inside the data center has become the binding constraint, and that the connectivity layer would decide who wins the AI race. That piece was correct, and it is also now incomplete. The bottleneck has moved again, and the new bottleneck is one that no amount of 800G switching or 1.6T fabric design can solve.

Last week, the Federal Energy Regulatory Commission ordered six regional grid operators — PJM, MISO, CAISO, ERCOT, SPP, and NYISO — to fast-track interconnection requests from data centers and other large loads. Grid operators now have 30 days to justify their existing rate structures and 60 days to demonstrate that they are treating large-load customers fairly. The framing in the press coverage was straightforward: the federal government has decided that AI data centers are too important to the economy to be stuck in multi-year interconnection queues, and FERC is going to do something about it.

The framing is wrong, and the wrongness is the story.

What the Order Actually Does

The order is a Section 206 show-cause proceeding directed at the six grid operators. It does three things:

It forces each operator to explain, within 30 days, why its existing rate structure is just and reasonable given the rapid growth in large-load interconnection requests.
It forces each operator to demonstrate, within 60 days, that its interconnection process is not unfairly delaying or rejecting data center customers.
It authorizes a new study process for generating facilities that serve electrically proximate or co-located large loads — a procedural carve-out designed to let behind-the-meter generation come online faster.

Read closely, the order is about process. It is about queue position, rate design, and the rules of who pays for what. It does not authorize a single new turbine. It does not approve a single new transmission line. It does not retire a single megawatt of fossil generation to make room for clean generation. It tells the grid operators to move the line faster, not to add anything to the line.

The press coverage has tended to describe this as "FERC fixed the data center grid problem." FERC did not fix anything. FERC moved the bottleneck one step downstream, into a place that is harder to see and harder to fix.

Why Generation Is the Real Constraint

The number that should anchor every AI infrastructure conversation in 2026 is not the number of GPUs, the price of HBM, or the latency of an 800G link. It is the length of the U.S. interconnection queue for new generation. As of the most recent reporting, the queue exceeds 2,600 GW of pending requests — which is larger than the entire existing U.S. power plant fleet. The average time from interconnection request to commercial operation for a new generator is now four to seven years, depending on the region and the queue position.

AI demand is projected to triple through 2035 on the most-cited industry baselines. Data center load growth alone is now the single largest driver of new electricity demand in every major U.S. grid operator's long-term forecast. Wholesale electricity prices are already up 267% in some regions since 2020, and capacity-auction clearing prices have hit FERC-mandated price caps in multiple regions for two consecutive auction cycles.

The fast-lane order does nothing to add to the supply side of that equation. It changes how fast a large-load customer can get through the queue. It does not change how much generation is actually available when the customer reaches the front of the queue.

The right mental model is not "the data center is stuck at a gate." The right mental model is "the data center has been handed a gate pass, but the highway behind the gate is jammed, and the highway is the entire constraint."

The Policy Illusion

This is the part that needs to be said clearly: the FERC order is a regulatory illusion, and not a small one. It produces the appearance of action on a problem that is structurally unsolvable by regulatory action alone. The constraint is generation, transmission, and the physical time it takes to build both. None of those respond to a 30-day show-cause order.

The reason this matters for AI infrastructure strategy is that enterprises are about to plan around the illusion. Some number of boardrooms, having read the press coverage, will conclude that the grid-connection problem is now "solved" by federal action. Site selection teams will assume that interconnection can be expedited on policy timelines. CFOs will model capacity availability on the assumption that the queue order in PJM or MISO is now a planning input they can ignore.

None of that is true. What the FERC order actually does is make the constraint more visible by accelerating the moment when a data center operator reaches the front of the queue and discovers that there is no electricity behind it. That is not a fix. That is a faster path to a wall.

The Series Callback

In the May hyperscale piece, the argument was that the AI infrastructure race has a hidden bottleneck — and the point was that most enterprises are not seeing it yet. The bet was that the bottleneck would move, and that the enterprises that saw it move first would have a strategic advantage.

The bottleneck has moved. It is no longer primarily the network fabric inside the data center. It is not even primarily the GPU allocation. It is the electrons, and the electrons are constrained by generation capacity, transmission buildout, and the multi-year physical construction timelines that no policy lever can compress.

The moat in the next phase of the AI infrastructure race is not pipes. The moat is electrons. And there is no policy lever that mints electrons.

What Should Be Built Now

The action list for anyone making AI infrastructure decisions in the next 18 months is unusually concrete, because the constraint is unusually specific:

Treat generation availability as a Tier-1 site-selection criterion. Power availability is no longer a permitting question — it is a generation-availability question. Confirm contracted MW with named generation assets, not with the queue order of a regional operator.
Model the cost of "queue position" as a probabilistic input. Queue position is no longer a deterministic timeline. PJM and MISO queue reforms have changed the rules on cost allocation and network upgrade charges; model multiple scenarios for both.
Plan for behind-the-meter and co-located generation as a default, not an exception. The FERC order explicitly authorizes expedited study for co-located loads. If your site doesn't have it, your competitive position is now structurally weaker than a comparable site that does.
Stress-test capex models against a 50% wholesale electricity price increase. The 267% number from 2020 is already in the rearview mirror. The next doubling is closer than most CFO models assume.
Treat nuclear PPAs and geothermal as the only credible baseload additions on AI-relevant timelines. Gas peakers are faster but don't move the wholesale price curve. Solar and wind alone don't move the capacity-adequacy picture. The baseload math is nuclear and geothermal, and both have multi-decade buildout cycles that started too late to fully cover the 2027–2030 demand wave.
Build a generation-procurement team that reports to the CIO, not to facilities. The constraint is no longer a facilities problem. It is a product problem, because it gates the compute capacity that the product runs on.

Eighteen Months Out

By Q3 2027, the first cohort of fast-lane data centers will have reached the front of the queue in at least one major region and discovered the wall. Expect that discovery to be loud, public, and framed as a failure of policy. It will not be. It will be the predictable consequence of treating a procedural fix as a physical one.

The interesting question — the one the industry is going to have to answer — is whether the next round of policy intervention will be just as procedurally framed (another queue reform, another rate case, another show-cause order) or whether the political system will finally be forced to grapple with the actual physical problem: that the U.S. needs to build generation on a wartime footing, and that wartime-footing generation requires wartime-footing permitting, transmission, and supply chain. The current FERC order does not do any of that. It moves the line. The line is not the problem.

Sources: