The Constraint Has Moved
One month in, the same pattern keeps showing up: the grid constraint hasn’t disappeared—it has shifted
A month ago I launched The AI Grid Report. The premise: capital is pouring into AI infrastructure on the assumption that power, transmission, and interconnection are solvable inputs. They are not. They are the constraint — and the constraint is not priced.
Over five issues plus bonus content, I have walked through what that looks like in practice. The grid that cannot absorb what capital wants to build. An oil shock that rewires the electrification story through the wires rather than the headlines. A methodology for scoring where the grid breaks. A transformer bottleneck that puts America’s AI timeline in China’s hands. A single-name wipeout that showed what happens when the macro is right and the execution inverts the order of operations. And a Texas paradox where everything connects — and that is the problem.
What follows is a guided tour of the first month: the sharpest argument from each issue, condensed. If any of it lands, the full archive is available at theaigridreport.com
ISSUE 01 · AI Needs Power — The Grid Can’t Keep Up
The launch issue opened with a number that should have reframed the AI infrastructure debate: in PJM, the largest U.S. power market: 67.4 percent of projects that enter the interconnection queue eventually withdraw, and in CAISO, the figure is 75.6 percent. Two out of three projects in the mid-Atlantic, and three out of four in California, never reach commercial operation. The system is not failing; it is filtering—and it is filtering on patience and balance sheet depth, not on project quality.
The scale underneath those withdrawal rates is what gives them weight. Roughly 2,300 gigawatts of generating capacity currently sits in U.S. interconnection queues—nearly double the 1,200 GW installed across the entire country, spanning over 25,000 projects with median wait times of three to seven years. A hyperscale AI training campus needs 500 megawatts to 1 gigawatt of continuous power on a two-to-three-year construction timeline, while a transmission line takes seven to twelve years. That gap is not a temporary mismatch; it is how the two systems operate.
The issue ended with the preliminary view of the AI Grid Constraint Index—a heatmap scoring nine U.S. grid regions across five constraint dimensions, with PJM and CAISO at the top for overall constraint, and ERCOT showing a zero percent withdrawal rate and a story we would have to come back to. The point of the index is not to describe the grid, but to measure where it breaks.That distinction matters, because most models still treat the system as if it processes projects rather than eliminates them.
ISSUE 02 · The Grid Can’t Keep Up With Oil’s Shock
The Strait of Hormuz disruption gave us a natural experiment in how electrification demand interacts with infrastructure that cannot respond at the same speed: Brent crude moved above 100 dollars, and a chokepoint responsible for roughly 20 percent of global oil flows was open only to vessels with Iranian escort. The first-order effect was the obvious one—substitution toward electric alternatives accelerates when fuel prices spike. The second-order effect was the more interesting one: WTI and Brent inverted, because the replacement for lost Middle East supply was not prompt North Sea production but U.S. exportable crude, making WTI the marginal scalable barrel pulled from both directions.
The point of the issue was not the oil price, but that electrification does not remove constraints—it relocates them from oil to electricity. Capital is concentrating in precisely the regions least able to absorb it: PJM accounts for over 40 percent of committed U.S. data center capital while sitting firmly in the constrained zone; CAISO carries 14 billion dollars in committed capital against the worst interconnection profile in the country; and ERCOT, still buildable, is accelerating toward the same wall.
The analytical frame shift is the part that matters. The consensus view treats the energy transition as a demand story, where capital follows visible growth; the constraint view treats it as an infrastructure reality, where capital selection depends on where the system can support that demand. The mismatch between the two—demand compounding in months, infrastructure responding in years—is the systematic mispricing, and it is also the edge.
BONUS ISSUE· Behind the Score
This was the methodology issue—the one that explains how the Constraint Index is built, because if the scores are going to influence capital allocation, anyone using them should know exactly how they are calculated. The index combines interconnection queue data with federal energy data to produce five constraint scores per region: queue depth, wait time, withdrawal rate, demand pressure, and capacity margin. Each score isolates a different failure mode of the grid, and together they form a constraint profile rather than a snapshot.
The counterintuitive finding is that a large interconnection queue is often the clearest signal that a system is constrained, not that it is growing. Queue size measures competition for access –not conversion. In several U.S. power markets, the volume of generation waiting in queues is multiple times larger than current peak demand—SPP at 6.0x, CAISO at 5.5x, and PJM at 4.0x. The biggest queues are frequently the worst places to build, because most of that pipeline will never connect, with inflows consistently exceeding completions and the backlog compounding rather than stabilizing.
The thresholds are calibrated so that current U.S. conditions span the full one-to-five range, with room for scores to move as constraints tighten or ease. A score of four is severe because the system is approaching failure, and a score of five is critical because the constraint is already binding. The purpose of the scale is not classification but early detection—tracking how close each region is to tipping into the next level of constraint before the market reprices.
ISSUE 03· America Can’t Build AI Without China’s Permission
This is the issue that surprised me most to write. The AI buildout does not have a chip problem or a capital problem—it has a transformer problem. Thirty to fifty percent of U.S. data centers planned for 2026 will be delayed or canceled, according to Sightline Climate, and of the 12 GW expected online this year, only 5 GW is under construction. Power transformer lead times have stretched to 128 weeks, with generator step-up transformers at 144 weeks, while the data center deployment cycle runs under 78 weeks. The math does not work.
The U.S. imports roughly 80 percent of its power transformer supply. Chinese transformer imports to the U.S. rose from 1,500 units in 2022 to over 8,000 through October 2025—a five-fold increase in three years—and then surged another 182 percent in the first two months of 2026. A decade of reshoring has not moved the needle. The country racing to dominate artificial intelligence cannot manufacture the electrical hardware to plug it in, while the policy response—tariffs on steel and copper—raises the cost of the equipment it cannot produce in sufficient volume.
The deeper story is demand convergence. Six sectors are now competing for the same transformers and switchgear simultaneously: AI data centers, clean energy deployments, aging infrastructure replacement, extreme weather rebuilds, manufacturing reshoring, and electrification. Total demand is approaching 455 GW-equivalent against a supply chain sized for a fraction of that. The transformer bottleneck is not one industry’s problem waiting to be solved; it is the collision point where five other structural demand drivers arrive at the same factory floor—and one country makes most of what comes off it.
BONUS ISSUE· Fermi — When the Model Breaks at the Company Level
Fermi was supposed to be the vertically integrated answer to the hyperscaler power problem: a private 11-to-17 gigawatt campus in the Texas Panhandle: gas and solar by 2026, nuclear by 2032, eighteen million square feet of data center shell, Pantex next door, a 99-year Texas Tech lease, a Korean EPC partner, Siemens turbine orders, and Executive Order branding—everything except the one thing that had to come first: a signed anchor tenant. Three SEC filings in three days told the rest of the story, with the co-founder CEO out on April 17, CFO Miles Everson resigning “without Good Reason” on the 19th, and being parked on the board via family-trust designation, then Fermi2.0 unveiled on April 20 with Marius Haas as Executive Chairman and Jeffrey Stein—a restructuring-situations director—on the board. Shares fell ~19% on the day and ~75% from the October IPO.
The macro thesis was correct: interconnection queues do outrun GPU generations, and hyperscalers will pay a premium for grid-independent gigawatt-scale power. The single-name execution inverted the order of operations the model required, as Fermi drew equipment-level debt against an unbuilt customer base, priced four AP1000 reactors with a 2031-to-2038 delivery window as near-term cash flow, and upsized nameplate from 11 GW to 17 GW without a single binding offtake agreement to carry one. The incoming chairman is a BayPine and ex-Dell transactions operator, and the new director’s last five board seats include Rite Aid, Westmoreland Coal, and Dynegy—this is not a build-the-grid dream team, but a sell-it-or-restructure-it team.
The lesson generalizes: offtake is the gating input, not GW nameplate. A one-gigawatt position with a signed hyperscaler contract is worth materially more than a seventeen-gigawatt position without one, and powered land with a short path to energization is the scarce asset. The integration premium Fermi tried to capture end-to-end was never the premium the market was offering, and selling into that scarcity is often better than attempting to own it. The bundle was the bug.
ISSUE 04 · Everything Connects in Texas. That’s the Problem
ERCOT has a zero percent withdrawal rate, and every project that enters the interconnection queue eventually gets through, with no mass cancellations, no multi-year attrition, and timelines shorter than any other major U.S. grid—on paper, the most efficient interconnection process in the country. The paradox is that ERCOT also carries the highest demand pressure ratio in the country, at 5.4x, with nearly six times more capacity attempting to connect than the system currently supports. A queue that clears and a grid under the most stress cannot both be true —unless the constraint has moved somewhere else.
It has: ERCOT trades friction for speed, while PJM filters projects at the queue and ERCOT lets them through, absorbing the consequences on the operating grid. Connection is not delivery; when the transmission path to load centers is saturated, connected projects hit curtailment, congestion, and negative nodal prices, meaning a wind farm can clear the queue, connect to the grid, and still have nowhere for its power to go. Speed of process does not eliminate the constraint—it relocates it from the application stage to the electrons that can’t reach the customer willing to pay for them. ERCOT’s CEO confirmed the volume problem in mid-April, noting that the grid is on track to nearly triple in size by 2030, and that the ISO is shifting from serial to batch interconnection studies to keep up.
For AI infrastructure, the tradeoff is clear enough: lower power costs, faster interconnection, and favorable regulatory treatment, but no guaranteed reserve during stress events, exposure to an energy-only market, and reliance on self-sufficiency. The fastest system is not necessarily the safest. Capital is still pricing connection as completion, which it is not—connection is cheap, delivery is uncertain, and cost is variable. The constraint has not disappeared in Texas; it has moved from the queue to the grid itself.
ISSUE 05 · Withdrawal Rates
The launch issue opened with the headline number — 67.4 percent withdrawal in PJM, 75.6 percent in CAISO. This issue went underneath it, across all seven tracked queues, and translated the statistic into the language capital allocators actually use. Across PJM, CAISO, NYISO, ISONE, MISO, SPP, and ERCOT, 14,856 projects have withdrawn — approximately 2,093 gigawatts of proposed generation, more than twice the entire installed capacity of the United States, walked away. No storm. No policy reversal. No single cause.Projects simply did not make it out of the queue. The headline numbers suggest abundance; the withdrawal numbers tell a different story; most models still treat the headline as real.
The spread is the analytical anchor. ERCOT clears at 0 percent; NYISO loses 81.2 percent. That spread is not random and it is not a supply problem — it is a grade on process design. Study timelines, fee structures, cost allocation rules, and speculation filters separate the systems that convert queue entries into operating generation from the systems that don’t. Regions with longer studies and ambiguous cost allocation lose three out of four projects. Cluster-study mechanics make withdrawal contagious — when one project leaves a cluster, costs reallocate to the survivors and some of those become uneconomic and withdraw too. Withdrawal begets withdrawal. And the filter is not neutral: in MISO 95 percent of withdrawn projects are clean, in SPP 89 percent, in CAISO 80 percent. Gas plants are more likely to survive their study windows. The queue is a clean-energy tax that nobody wrote.
The framing this suggests is private equity. Treat each queue entry as a deal moving through a PE shop, and the US interconnection system has a 67 percent deal-break rate after DD in its largest market. No PE fund could survive those metrics. LPs would stop re-upping, stated deployment capacity would be discounted, the strategy would be wound down. Headcount would be reduced. The capacity illusion follows directly: PJM’s ~254 GW active queue, haircut by the 67 percent rate, is closer to ~83 GW of expected installed capacity. CAISO’s ~89 GW becomes ~22 GW. The queue is three to four times larger than the actual pipeline, and a PPA signed for a project in that queue is not a power contract — it is an option on a power contract.
The Through-Line
Five issues, one argument: Small change, big impact.
Capital is modeling AI infrastructure as if the grid is a neutral input that will scale to meet demand, but the grid is not neutral, and it will not scale on that timeline. Interconnection queues filter rather than process; transformers arrive on multi-year lead times into a demand curve that accelerates in months, and fast systems relocate constraints rather than eliminate them. Single-name execution fails when sequencing is inverted. The mismatch between where capital is flowing and where the grid can deliver is the systematic mispricing—and it is the edge. The open question is how long the market can continue pricing the headline before it is forced to price the outcome.
Go deeper on each issue
This piece pulls together the first five issues plus bonus content. Each one stands on its own if you want to go deeper.
– AI doesn’t scale where power isn’t available—and most timelines assume that constraint doesn’t exist→
– Energy demand is no longer tracking GDP, which means traditional signals are starting to break down→
– The U.S. can design the system, but still depends on external supply chains to actually build it→
– What happens when the macro is right but execution fails (and capital learns the hard way) →
– The bottleneck didn’t disappear—in ERCOT, it moved from interconnection to delivery, where it’s harder to see and harder to price →
– How to actually measure where the grid breaks (the Constraint Index methodology) →
– The queue looks like capacity, but most of it never becomes supply—and the withdrawal rate is where that illusion breaks →
What Comes Next
That is the first month, and the Constraint Index will continue to update weekly. The regional deep dives continue—PJM next, followed by the withdrawal rate as a cross-market metric, and then the transmission layer underneath all of it. The full archive, the weekly index, and the paid deep dives are available at theaigridreport.com.