What the queue is, and why it matters

An interconnection queue is the waiting list for new power plants trying to connect to the grid.

But it is more than a backlog.

Every project in the queue has already committed capital, entered a multi-year approval process, and accepted the possibility of transmission-upgrade costs before it can deliver power. In practice, the queue is a map of what developers believe is worth building — and where they believe future electricity demand will justify the wait.

That matters because the queue increasingly reveals a very different picture of America’s future grid than most public policy debates do.

Across the country, developers are attempting to build massive amounts of:

• solar

• battery storage

• gas generation

• transmission-connected AI infrastructure

while coal has almost completely disappeared from new development plans.

The U.S. power system is fragmented across regional grid operators. Seven major regions — including PJM, ERCOT, MISO, CAISO, SPP, NYISO, and ISO New England — publish detailed project-level queue data. Together, along with the non-ISO Southeast and Western utility regions, they represent nearly all US electricity generation.

The scale is enormous.

Roughly 1,400 GW of generation and 890 GW of storage — 2,290 GW combined — sit in active US interconnection queues as of end-2024, per Lawrence Berkeley National Laboratory (LBNL — a US Department of Energy national lab whose Electricity Markets and Policy group publishes the annual Queued Up interconnection report). Solar represents the largest share of proposed capacity, followed by battery storage. Gas remains meaningful but regionally concentrated. Wind is increasingly uneven by geography. Coal is almost absent.

The queue does not guarantee these projects will ultimately be built.

But it does reveal something extremely important: what the market believes the future grid will need.

What the composition implies

Coal is not underrepresented in the queue; it’s almost absent.

The existing coal fleet still operates. Some plants continue clearing in PJM’s capacity auctions, while others remain online under reliability agreements that delay retirement. They’re reliable; they’re necessary, but they’re politically toxic. Developers are no longer attempting meaningful new coal capacity in any major U.S. interconnection queue.

What replaces it is where the regions diverge.

ERCOT is building solar and storage in roughly equal proportions. PJM shows a broader mix of gas, solar, storage, and a small but unusual cluster of nuclear projects. MISO’s newer queue cycles tilt more heavily toward gas than its earlier renewable-heavy cycles. CAISO, NYISO, and ISO New England show almost no new gas development at all.

The same country.

The same physics.

The same Inflation Reduction Act.

Different bets.

Why the queue matters

The queue is not a forecast of what will ultimately be built.

Withdrawal rates remain high across most regions, and projects reaching commercial operation often spend years moving through the interconnection process. The queue reflects what developers are attempting — not what will necessarily survive.

But that still makes it one of the most important forward indicators in the power market.

Every project in the queue has already passed an economic screen. Capital has already been committed. Developers have already decided the opportunity is worth years of waiting, study costs, and potential transmission-upgrade exposure.

That is a very different signal from:

• a press release

• a policy target

• a utility planning document

• a political speech

The queue is capital already moving toward the grid.

What this issue covers

This week’s premium section will explain:

• why some regions are still building dispatchable supply while others are not,

• why storage exploded across the queues, and

• what the interconnection data may already be revealing about the future shape of the US grid.

Continue Reading

Read more

Most wealthy investors think about taxes as something minimized quietly in the background.

That framing understates what’s happening. The transferable tax credit market reached $42 billion in 2025, up 27% from $32 billion in 2024. When tax equity and preferred equity are included, total tax credit monetization hit $63 billion. The market grew through OBBBA (Trump Administration signature tax legislation passed in 2025) reconciliation, not despite it.

But there is a more strategic question in this conversation:

If you are going to write the check anyway, what exactly do you want your capital funding?

A ballroom? Or a power plant?

That question sounds rhetorical until you look at how transferable tax credits are working inside the real economy.

What’s changing is not the desire to reduce taxes. That has always existed.

What’s changing is the realization that the tax code is the policy tool that influences which physical systems are built.

The §6418 transferability provision in the 2022 IRA opened a market that didn’t exist before. The transferability mechanism survived the 2025 OBBBA reconciliation. Eligibility rules tightened — meaning the pool of qualifying projects narrowed — but the buyer-side market remained intact. Some credits were eliminated; some had their eligibility windows shortened; some of the eligibility criteria, especially around foreign content, were tightened. The window that exists today is real, but it is not the same window that existed in 2024.

A transferable tax credit is usually described as a financial instrument. Technically, that is true. Economically, however, it is beginning to function more like a directional signal for capital allocation into energy, manufacturing, grid infrastructure, and industrial capacity.

The U.S. is entering a period where physical infrastructure constraints are dominating economic outcomes in a way the financial markets are just beginning to price.

The constraints appear in transformer shortages, interconnection delays, transmission bottlenecks, power availability, natural gas infrastructure, and generation lead times that can’t be resolved on software timelines.

Data center expansion timelines are accelerating faster than transmission and grid delivery capacity. That reality is changing how capital evaluates long-duration opportunity.

Capital Is Becoming More Intentional

Family offices, private investors, and long-duration pools of capital are no longer thinking purely in terms of quarterly returns.

Many are asking a different question:

“What systems do we want our capital to support?”

The U.S has entered a period where enormous amounts of physical infrastructure need to be rebuilt, expanded, or modernized simultaneously.

The AI buildout alone is exposing how underbuilt the grid really is.

Data centers require massive injections of power. Power at that scale depends on transmission, transformers, substations, permitting, land access, and infrastructure investment cycles measured in years rather than quarters.

The Grid Is Quietly Repricing Everything

This is particularly visible inside energy infrastructure. Unlike the software cycles that dominated the last decade, the outcomes here are physical and measurable — whether a transmission corridor is built, a generation project reaches commercial operation, or grid capacity expands has direct consequences for industrial growth, electricity pricing, and regional competitiveness.

Stocks like Eaton, Vertiv, and Schneider have traditionally been seen as downstream of the AI buildout. Not anymore — they are the gating layer, and the market has noticed. All three have broken out of long-term ranges to the upside, which confirms the thesis but does not, on its own, signal current entry points. The structural story and the entry timing are separate questions.

The market spent the last decade believing software was the bottleneck.

The bottleneck is becoming physical infrastructure. The result is that electricity itself is beginning to reprice.

It means the real opportunity is not simply “AI.”

It’s the underlying physical systems required to support it.

Transferable tax credits are directing capital toward those systems.

If electricity becomes constrained by transmission and deliverability rather than generation alone, how should infrastructure itself be valued relative to the applications sitting on top of it?

Family Offices May Be Uniquely Positioned

Large institutions and public markets often struggle with long-duration infrastructure cycles because the timelines rarely align cleanly with quarterly expectations. Family offices, however, can operate with substantially longer investment horizons.

The structure fits family offices unusually well. The infrastructure cycle is long-duration, illiquid, politically shaped, and operationally complex — conditions many quarterly-managed pools of capital struggle to underwrite.

That dynamic makes the current environment particularly interesting because the underlying infrastructure cycle will not resolve quickly.

Grid expansion doesn’t occur on software timelines.

Transmission projects can require seven to ten years to complete, transformer manufacturing capacity can’t expand overnight, and AI-driven power demand is rising inexorably.

That mismatch creates long-duration investment opportunities tied to hard assets, constrained systems, and real economic necessity.

The Psychology Is Changing Too

Historically, taxes were viewed primarily as something minimized quietly in the background through accounting strategies, portfolio structuring, and conventional tax planning.

Some investors are now beginning to ask a different question:

“Is there a more focused, intentional way to allocate tax dollars while still complying with tax policy?”

Transferable tax credits stop feeling like accounting exercises and begin functioning more like participation in a national industrial buildout tied to energy infrastructure, manufacturing capacity, grid resilience, and domestic economic expansion.

Investment structures carry clearer accountability than the metrics that have dominated philanthropic and ESG-scoring frameworks. A transferable tax credit settles against a placed-in-service date, a real project, and a real tax liability. An ESG score reflects a methodology the rating agency can revise after the fact.

The opportunity is no longer viewed exclusively through the lens of tax reduction or financial optimization, but through participation in the construction of systems that are strategically important to the economy itself.

That change in mindset may ultimately become one of the defining characteristics of this cycle.

Why This Matters Beyond Taxes

The market still tends to describe transferable tax credits as niche financial instruments.

What is emerging is a system in which tax policy, infrastructure constraints, industrial strategy, energy demand, and capital allocation are becoming intertwined.

The immediate question is whether this cycle produces attractive returns. The credits themselves are purchased at a discount — that’s the market incentive— and that discount is tax-free—that’s the policy tool. So, in practical terms: a $5 million tax liability becomes $4.5–4.6 million in credits directly funding a real infrastructure project with a placed-in-service date. The project exists before the capital arrives. The accountability is built in from the start.

That discount alone may not move the needle for family office allocators—though it has created a powerful enough incentive to have established a $42 billion market in two years, mostly for corporations— but pair it with the idea that taxes can be redirected toward causes and infrastructure projects family offices genuinely want to support, and the equation starts to look very different. The larger question is whether it changes how sophisticated investors think about participation in the construction of physical systems. “My taxes fund national parks” is no longer an aspirational joke at a cocktail party. “My taxes fund the energy buildout (and not a ballroom)” is a legitimate claim.

If infrastructure becomes the limiting factor behind AI, manufacturing, and industrial growth, what happens to the value of the systems that make all three possible?

The AI Grid Report

AI Grid Report is an independent research publication focused on the intersection of energy infrastructure, AI-driven electricity demand, grid constraints, industrial policy, and long-duration capital allocation. Recent editions have covered ERCOT, PJM, MISO, interconnection bottlenecks, transmission constraints, and the physical infrastructure dynamics increasingly shaping economic outcomes across the US energy system.

If these themes interest you, you can subscribe to AI Grid Report here →

Subscribe now

About Dakota Ridge Capital

Dakota Ridge Capital works with family offices, private investors, and strategic capital navigating the energy and infrastructure buildout — including direct access to transferable tax credit opportunities tied to real projects already in development.

If you are evaluating how your capital allocation intersects with the infrastructure cycle, the conversation is worth having before the next reconciliation window closes.

Request a private briefing → https://www.dakotaridgecapital.com/contact-us

Where We Are in the Walk

Six issues in, the AI Grid Report (TAGR) has covered the methodology, the macro tariff and Gulf threads, and the first two regional deep dives.

→ Issue 04 — ERCOT. The Lone Star grid clears its interconnection queue at near-zero withdrawal. Connect-and-manage interconnection, no centralized capacity market, one regulator. Texas doesn’t have PJM’s auction problem; it doesn’t have auctions.

→ Issue 06 — PJM. The capacity auction cleared at the ceiling. Twice. PJM’s queue doesn’t clear; ERCOT’s does. Same demand. Same generation. Different process.

This week, we’re looking at MISO (Midcontinent System Operator)— the largest physical footprint on the US grid by territory, stretching from Manitoba to northern Louisiana.

MISO’s headline number is not an auction price.

It is a retirement schedule.

The Number

~25 GW of MISO operating generation is on a planned retirement schedule. That is roughly 37% of MISO’s average demand.

No other major ISO is close.

SPP is next at roughly half the exposure. PJM sits in the mid-teens. ERCOT — despite being the fastest-growing load center in the country — remains below 3%.

MISO has its own retirement problem.

What’s Leaving

The 25 GW is concentrated.

Subbituminous coal accounts for roughly 14 GW. Natural gas adds another ~8 GW.

Nuclear is notably absent from the near-term retirement schedule. Xcel’s Monticello plant — the one obvious candidate, previously scheduled to retire in 2030 — received NRC (Nuclear Regulatory Commission) approval in February 2026 to operate through 2050. Prairie Island is in line for similar treatment. MISO’s nuclear fleet has been materially de-risked at the same time the dispatchable coal and gas fleet is leaving.

The issue is not simply the volume of retirements.

It is the type of generation leaving the system.

Much of what’s retiring is dispatchable, fuel-secure capacity that can operate through extended weather events, peak demand periods, and low renewable output conditions (i.e., no sun, no wind).

That distinction becomes critical once replacement timing starts slipping.

What We Are Diving Into This Week

ERCOT cleared its queue. PJM cleared at the ceiling. MISO has not had to clear either.

What MISO has is firm capacity leaving the system on a posted schedule, while much of the replacement remains trapped in the queue rather than connected to the grid.

In this week’s premium article, we’ll show which plants are retiring, which utilities own them, what the replacement queue actually contains, and why replacing coal megawatts with renewable megawatts is not a one-for-one exchange.

We’ll also cover the news thread that may have quietly turned MISO’s retirement schedule from a posted plan into an active planning problem — and why the fourth federal emergency order on a single Michigan coal unit may be the precedent the rest of the 2028 fleet now has to underwrite against.

Keep Reading ->

Read more

PJM cleared at the cap. Again.

The system keeps finding the ceiling. It hasn’t found the solution.

On December 17, 2025, PJM’s capacity auction cleared at $333.44 per megawatt-day.

That is the cap.

It is also the second consecutive PJM auction to clear at the cap. The previous auction, held in July 2025, cleared at $329.17. Both auctions hit the price ceiling that FERC set in late 2024 to keep residential bills from a sudden jump.

Together, those two auctions secured 134,479 MW of capacity for $16.4 billion of annual payments — at a procured reserve margin of 14.8%, the lowest in PJM’s history.

That number is now the most-watched price in US electricity.

It is also a price the system has never actually been allowed to discover.

It is also, increasingly, the price that determines whether AI infrastructure is built on schedule.

The Cap Was Extended — Again

The cap was supposed to be a temporary instrument. It is now in its fourth consecutive year.

On April 28, FERC approved PJM’s request to extend the price collar for two more delivery years — 2028/29 and 2029/30. The extension keeps the cap at roughly $325/MW-day and the floor at $175/MW-day, the same parameters that capped the last two auctions. Without it, PJM had estimated the next auction’s price cap would have defaulted to about $550/MW-day.

When PJM holds its next Base Residual Auction this summer, the cap will still be in place. The auction after that — in December — will be capped too.

PJM’s Base Residual Auction (BRA) procures capacity three years ahead of delivery. The December 2025 auction cleared capacity for the June 2027–May 2028 delivery year. The summer auction clears for June 2028–May 2029. The December auction clears for June 2029–May 2030. All three are now capped.

The forward gap exists so generators can build new supply for a known revenue stream. Except the forward gap is now shorter than the build cycle for new gas generation, which is the structural problem the rest of this issue describes — and which is why FERC extended the cap. New supply cannot arrive in time, so the cap stays on to protect ratepayers from the price the auction would otherwise produce.

The market still does not know what an uncapped clear looks like. It is now four years since it last did. The two capped auctions were already records.

What would the market actually clear at without the cap?
$400? $500? Higher?
There is no consensus — only suppressed price discovery.

The auction is clearing at the ceiling. The price hasn’t been allowed to move.

Same Demand. Same Generation. Different Process.

On the supply side, firm capacity is exiting faster than it is being replaced (or, at least, trying to exit). On the demand side, ~5,250 MW of new load arrives every year inside PJM’s forecast — almost all of it is data centers.

The clearing price is the cost of that gap — the dollar value, expressed per MW per day, of every megawatt of firm capacity PJM needs but doesn’t have. When supply is short, the price rises until enough supply (or enough demand response) is willing to commit. When the gap is structural — firm supply exiting faster than replacement — the price keeps rising until something stops it. This time, FERC stopped it with a cap.

In ERCOT, the same supply technologies, the same data center developers, and the same hyperscale tenants run through a queue that clears fast.

In PJM, two-thirds of projects in the queue never reach completion.

The queue is full. Most of it never meets its intended load.

That is the entire difference between a $333 / MW-day capacity clear and a market—ERCOT— that doesn’t pay for capacity at all.

If PJM and ERCOT are running the same technologies, why does one system convert and the other stall?
Is it structure, governance, or execution?

In This Week’s Edition

This week breaks down:

• what is leaving PJM (~5,500 MW of fossil capacity scheduled to retire — Brandon Shores, H.A. Wagner, Conemaugh, Keystone) — and why scheduled retirements are not real retirements

• what is replacing it (2.1 GW in 2025, almost entirely solar)

• how Northern Virginia became the place where the December auction price was decided

• what FirstEnergy’s Energize365 program — a $36B transmission build-out by PJM’s largest transmission owner — tells us about the supercycle hiding inside the auction print

• what ERCOT does differently — and which parts of it PJM can copy

The full Constraint Index shows which PJM zones are most exposed to the next auction, and which signals to watch going into the summer auction.

Continue reading:

Read more

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. A Texas paradox where everything connects — and that is the problem. And, most recently, a review of the sorry tale of projects failing to make it through the interconnection queue, and wasting billions in sunk costs.

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 can’t 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 doesn’t 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? 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. And that’s a better place to be.

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 voted for.

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→ https://www.theaigridreport.com/p/ai-needs-powerthe-grid-cant-keep

– Energy demand is no longer tracking GDP, which means traditional signals are starting to break down→ https://www.theaigridreport.com/p/the-grid-cant-keep-up-with-oils-shock

– The U.S. can design the system, but still depends on external supply chains to actually build it→ https://www.theaigridreport.com/p/america-cant-build-ai-without-chinas

– What happens when the macro is right but execution fails (and capital learns the hard way) → https://www.theaigridreport.com/p/fermi-inc-frmi-how-a-good-macro-call

– The bottleneck didn’t disappear—in ERCOT, it moved from interconnection to delivery, where it’s harder to see and harder to price → https://www.theaigridreport.com/p/everything-connects-in-texas-thats

– How to measure where the grid breaks (the Constraint Index methodology) → https://www.theaigridreport.com/p/behind-the-score-how-we-measure-grid

– The queue looks like capacity, but most of it never becomes supply—and the withdrawal rate is where that illusion breaks → https://www.theaigridreport.com/p/the-withdrawal-rate

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.

Across the seven US interconnection queues we track, 14,856 projects have withdrawn.

That is approximately 2,093 GW 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 queue numbers suggest abundance.

The withdrawal numbers tell a different story.

Most models still treat the headline number as real.

From Texas to Everywhere Else

Last week, we looked at Texas.

ERCOT clears its queue. 0% withdrawal rate. Every project that enters gets through.

That is the exception. Everywhere else, the queue behaves more like a sieve.

→ CAISO: 75.6% withdrawn

→ NYISO: 81.2% withdrawn

→ ISONE: 75.3% withdrawn

→ PJM: 67.4% withdrawn

→ SPP: 64.0% withdrawn

→ MISO: 54.5% withdrawn

The median project in six of seven tracked regions never becomes a project.

The Spread Is The Story

0% in Texas and 81% in New York. That spread is not random.

It is a grade on process design — study timelines, fee structures, cost allocation rules, speculation filters.

The regions with the longest studies and the most ambiguous cost-allocation rules have the highest withdrawal rates.

The regions with faster processes and clearer rules have lower ones.

Withdrawal is not a supply problem; it is a process outcome. Most commentary still treats it as a supply problem.

What Withdrawal Measures

A withdrawal is not a cancelled project.

It is a completed study, an interconnection agreement negotiation, consultant fees paid, sometimes years of developer time, and then the developer walks away.

The costs are already sunk.

The grid study is already done.

The planners already slotted the project into their load-flow models.

When a project is withdrawn, all that work is unwound.

It is the most expensive form of queue churn.

The Open Question

If two-thirds of the projects in PJM’s queue never get built, why is the queue still growing?

The answer is structural—and uncomfortable.

Developers are not filing because they expect to build.

They are filing to preserve optionality — a queue slot, a study result, a position in the stack — while they watch the market.

The queue has become an options market on future interconnection rights.

Most of those options expire worthless.

The open question is whether the market should allow that to continue.

In This Week’s Edition

This week’s edition breaks down:

• why the spread from 0% to 81% is a grade on process design, not supply

• which generation types disproportionately die in the queue — and what that says about the energy transition

• how headline queue GW overstates the actual pipeline by a factor of three

• what a withdrawal-adjusted queue looks like for AI infrastructure planning

The full Constraint Index shows which regions have functional pipelines — and which ones are running on paper volume.

Continue reading

Read more

The Paradox

Texas has a 0% withdrawal rate.

Every project in the interconnection queue gets through.

→ No mass cancellations.
→ No multi-year attrition.
→ No system breakdown at the application stage.

On paper, it’s the most efficient grid in the United States.

But the queue is only one part of the system.

The grid behind it is a different question.

From Hardware Constraints to System Constraints

Last week, we outlined the emerging hardware constraint — transformers, substations, and long-lead equipment.

This week, the focus shifts.

Texas doesn’t fail at the hardware layer. It doesn’t fail at the queue.

It fails somewhere else.

Two Systems, Two Outcomes

The contrast with PJM is stark.

PJM:

• 9,251 active projects

• 67.4% withdrawal rate

• ~3.5-year average wait time

ERCOT:

• 1,832 active projects

• 0% withdrawal rate

• ~710-day average timeline

Two systems. Two completely different outcomes.

At first glance, the conclusion seems obvious:

PJM is constrained. ERCOT works.

But the data doesn’t fully support that.

The Signal Beneath the Surface

Demand pressure tells a different story. Demand pressure is measured as the ratio of generation ready to deliver power to peak load (which is a close proxy for the grid’s capacity to satisfy that load).

PJM demand pressure:

• ~3.0x

ERCOT demand pressure:

• 5.4x — the highest in the country

Nearly six times more capacity is attempting to connect than the system currently supports.

Where the Model Starts to Break

So:

• every project gets through

• timelines are shorter

• the queue appears functional

And yet:

• the system is under the highest pressure in the country

Those two things should not coexist.

The Open Question:

In This Week’s Edition

Most analysis stops at the queue.

That’s not where the problem is.

This week’s edition breaks down:

• why 0% withdrawal does not mean low risk

• where ERCOT is already showing signs of saturation

• how the constraint moves from the queue to the operating grid

• what the reliability tradeoff actually means for AI infrastructure in Texas

The full Constraint Index shows exactly where this pressure is building—and where it’s about to break.

Continue reading →

Read more

Fermi Inc. (FRMI) Had the Right Macro and the Wrong Everything Else

Three SEC filings in three days. Co-founder CEO out. CFO resigned “without Good Reason” and was placed on the board via family-trust designation. A restructuring specialist director was brought in. “Fermi 2.0” announced. Shares down 19% on the day, 75% from October’s IPO.

The obvious read is that an AI-power story cratered. The more accurate read is that the AI-power thesis was right and the single-name execution was wrong — and the distance between those two things is where the money is.

Fermi was supposed to be the vertically integrated answer to the hyperscaler power problem:

• A private 11-to-17 GW campus in the Texas Panhandle.

• Gas and solar in 2026, nuclear by 2032

• 18 million square feet of data center shell

• Pantex (the place where every nuclear warhead in the America stockpile is built, refurbished, and taken apart) next door

• A 99-year Texas Tech lease

• A Korean EPC partner

• Turbine orders with Siemens

• Executive Order branding.

Everything except the one thing that had to come first: a signed tenant.

This week’s filings aren’t a strategic evolution. Rather, they’re a founder removal dressed in a press release.

The incoming chairman is a BayPine / ex-Dell transactions operator.

The new director’s last five board seats include Rite Aid, Westmoreland Coal, and Dynegy.

This isn’t a build-the-grid dream team; it’s a sell-it-or-restructure-it team.

The full post-mortem — original thesis, where it broke, five management failures ranked by materiality, the read on “Fermi 2.0,” and what it all means for how we underwrite AI power going forward — is below for paid subscribers.

Physics over narrative, constraints not forecasts. Offtake is the gating input, not nameplate GW.

Read more

Subscribe now

The AI buildout doesn’t have a chip problem. It doesn’t have a capital problem. It has a transformer problem.

Half the data centers planned for 2026 won’t open on time. Not because the money ran out. Because a piece of equipment that costs less than 10% of the project — and takes five years to deliver — hasn’t arrived.

The country racing to dominate artificial intelligence can’t manufacture the electrical hardware to plug it in.

The Numbers

— 30–50% of US data centers planned for 2026 will be delayed or canceled (Sightline Climate, April 2026).

— Of 12 GW expected online this year, only 5 GW is under construction. 16 GW remains “announced” with no visible progress.

— Power transformer lead times: 128 weeks average. Generator step-up transformers: 144 weeks. Data center deployment cycle: under 78 weeks.

— Chinese transformer imports to the US surged from 1,500 units in 2022 to over 8,000 through October 2025 — a five-fold increase in three years.

The standard narrative says the AI constraint is chips, or capital, or models. That framing misses where the physical system binds.

The constraint is electrical equipment. Transformers step voltage down for data center use. Switchgear distributes power safely. Batteries smooth demand spikes that degrade chip equipment. Without these components, a data center is a building with no purpose — like a highway with no on-ramps.

And the US can’t make enough of them.

The Delay Wall

Here’s the math that should concern anyone underwriting AI infrastructure timelines.

Sightline Climate (a tactical intelligence platform for energy transition decision-makers) tracks 140 large data center projects targeting 16 GW of capacity by year-end 2026. Only 5 GW is under construction. The rest sits in a pre-construction limbo where announcements run years ahead of physical capability.

This isn’t new. In 2025, manufacturers delayed 26% of announced capacity and pushed back commercial operations for another 10%. The pattern is consistent and accelerating (per Techspot): for 2027, more than 25 GW has been announced, but fewer than 10 GW is under construction. For 2028–2032 (per Futurism), 37 GW of planned capacity has no firm completion date. Only 4.5 GW has broken ground.

Capital is available. Land is available. The electrical equipment to connect these facilities to the grid is not.

The gap between announced capacity and construction is widening, not closing.

The Equipment Bottleneck

Think of transformers like rebar in a building. Nobody talks about rebar. It’s maybe 3% of total construction cost. But if the rebar doesn’t show up, you don’t have a building — you have a pile of expensive concrete.

Transformers are the rebar of the AI buildout.

Lead times for high-power transformers have stretched from 24–30 months pre-2020 to as long as five years. AI data center deployment cycles run under 18 months. That’s a 3.3x gap between what the buildout needs and what the supply chain delivers.

Electrical infrastructure represents less than 10% of total data center cost — but a single delayed component halts the entire project.

The China Dependency

The US imports approximately 80% of its power transformer supply and 50% of its distribution transformer supply. In January 2026, US utility executives visited a Chinese transformer factory and found roughly half the units on the floor flagged for American delivery — some destined specifically for data center companies.

Chinese transformer exports to the US increased by 182% in the first two months of 2026 alone. China’s share of US battery imports remains above 40%. Its share of certain transformer and switchgear categories sits at approximately 30%. Power transformer demand is up 116% since 2019. Generator step-up transformer demand is up 274%.

A decade of reshoring initiatives has not moved the needle. The US faces a 30% supply deficit in power transformers and a 10% deficit in distribution transformers.

Despite every policy signal pointing toward decoupling, the physical dependency is deepening.

Here’s what most infrastructure analysis misses: the equipment bottleneck isn’t just a supply problem. It’s a demand convergence problem.

Six sectors are competing for the same transformers and switchgear simultaneously. US electricity demand is up 7% since 2020, reversing a decade-long decline. Solar deployments are up 400%+ since 2019. More than half of US distribution transformers are past their expected service life. Billion-dollar weather events rose from 8 in 2014 to 27 in 2024.

The AI buildout isn’t competing against a shortage. It’s competing against five other structural demand drivers for the same hardware — and the full scale of that collision is worse than most models assume.

Anyone modeling data center timelines without a transformer procurement sensitivity is not modeling the proper constraint. They’re modeling the press release.

Tariffs have made this worse, not better. Steel tariffs have raised transformer prices approximately 20%. Copper tariffs of 50% inflate costs for both imported and domestically manufactured units. The US has one domestic producer of grain-oriented electrical steel (steel in which the grain lines up with the direction of travel of the magnetic field) — AK Steel (Cleveland-Cliffs). The highest-grade material still has to be imported.

OEMs have announced $1.8 billion in capacity expansions since 2023. Siemens Energy committed $1 billion to US manufacturing. GE Vernova committed $600 million plus a $5.3 billion Prolec GE acquisition. These are real investments. But the demand growth rate means they’re running to stand still.

“It’s going to be a long battle. It’s not something that can be solved in the next five years.” — Ben Boucher, Wood Mackenzie

AI Grid Constraint Index — Preview

The equipment bottleneck doesn’t hit every region equally. Some ISOs have pre-ordered transformer inventory 36 months out. Others are still waiting on deliveries that were due in 2024.

That divergence is already showing up in our regional scoring. Regions with secured equipment pipelines and shorter interconnection queues are pulling ahead. Regions dependent on Chinese imports with no domestic backup are falling behind — and tariffs are accelerating the split.

This week’s Constraint Index update tracks:

• which regions have locked in transformer and switchgear delivery — and which are exposed to the 5-year lead time wall

• where the equipment bottleneck is already binding on energization timelines

• the two regions where our score moved this week — one tightening, one stabilizing

The full Constraint Index with regional scores, capital sensitivity notes, and the methodology breakdown is available to subscribers.

Most analysis stops at “demand is growing.”

That’s not where the problem is.

The problem is that six industries need the same hardware, one country makes most of it, and the policy response is raising the price of the hardware we can’t make ourselves.

This week’s full analysis covers:

• The full demand convergence breakdown — 455 GW of competing demand for a supply chain sized for a fraction of that

• Why the Gulf AI bet — $2.2 trillion in pledges — just hit a wall that has nothing to do with technology

• The Stargate Abilene case study: what happens when GPU cycles outpace grid timelines

• Which electrical infrastructure equities are positioned for the convergence, and why pricing power is structural

• The helium wildcard most semiconductor analysts aren’t tracking

→ Full analysis available to subscribers

Which data center project in your pipeline has the most unrealistic equipment procurement timeline? Serious question — we’re tracking this.

Subscribe now

Read more

If you’re investing in AI infrastructure, you’re not betting on compute.

You’re betting on whether the grid can deliver power on time.

Most capital is being deployed as if power is available, delay is temporary, and capacity will scale to meet demand.

None of those assumptions hold.

Interconnection queues are opaque.

Timelines stretch into years.

And the difference between a viable project and a stranded one often comes down to constraints that don’t show up in standard models.

This is the layer most investors are missing.

We built the AI Grid Index to make those constraints visible.

Not to describe the grid — but to measure where it breaks.

Subscribe now

Why This Matters

Not all megawatts are equal.

Not all interconnection queues represent real capacity.

And not all regions fail in the same way.

A large queue can signal opportunity — or it can signal congestion so severe that most projects never get built.

A fast-growing demand region can look attractive — until you realize the grid cannot absorb new load within any reasonable timeframe.

Most analysis stops at surface-level metrics.

This index goes one layer deeper — into the constraints that determine what gets built, what gets delayed, and what never connects.

This article explains how the AI Grid Constraint Index is built.

The full index—updated weekly across U.S. grid regions—is available to paid subscribers.

Behind the Scores: The Math That Powers the Index

Last week, we published a heatmap scoring nine U.S. grid regions across five constraint dimensions. If you missed it, you can view it here. Disclaimer: it is only for paid subscribers.

Every score in that map comes from observable data.

If you’re making infrastructure decisions based on those scores, you should know what they represent.

We publish the methodology for one reason: To make the system legible.

Two Data Streams. Five Scores.

The index combines two independent data sources:

Interconnection queue data and federal energy data.

From these, we measure five constraint types:

• Queue Depth — total gigawatts waiting to connect
  

• Wait Time — how long projects remain in the queue
  

• Withdrawal Rate — how many projects fail to reach completion
  

• Demand Pressure — queue size relative to actual grid demand
  

• Capacity Margin — how much (net) generation is leaving relative to demand

Each score isolates a different failure mode of the grid.

Together, they form a constraint profile — not just a snapshot.

Why This Matters

Queue data alone misses context.

A 200 GW queue on a grid serving 94 GW of demand (grid and transmission capacity similar) is very different from a 200 GW queue on a grid serving 14 GW.

The first is ~2x and the queue is likely to clear. The second is a fantasy: most projects will leave the queue, and never complete.

Each score isolates a different failure mode of the grid.

Together, they form a constraint profile — not just a snapshot.

What Each Score Actually Tells You

Queue Depth

High queue depth is often mistaken for growth.

In reality, it frequently signals congestion — more projects competing for limited interconnection capacity.

Wait Time

At four years or more, the queue stops functioning as a planning tool.

It becomes a bottleneck.

Withdrawal Rate

A high withdrawal rate reveals a broken process.

Most projects don’t fail because of capital — they fail because they cannot get connected.

Demand Pressure

Active queue GW divided by Average Demand.

This is where most models fail.

A large queue means nothing without demand context, but demand can’t outgrow the capacity of the grid to deliver the generation in the queue.

Capacity Margin

Retirements are the silent constraint.

Capacity leaving the grid can tighten conditions faster than new generation can replace it.

The Thresholds

Each score maps a continuous metric to a 1–5 scale using fixed thresholds, calibrated against 2024–2026 U.S. ISO data.

They are not arbitrary.

They were designed so that current conditions span the full range — with room for scores to move as constraints tighten or ease.

A score of 4 is “severe” because it is a system approaching failure.

A score of 5 is “critical” because it signals that the constraint is already binding.

The purpose of this scale is not classification — it is early detection.

Where Each Region Falls

Most investors are reading interconnection queues backwards.

They see large pipelines and assume future capacity.

In reality, large queues are often the clearest signal that the system is constrained.

Queue size is not a measure of growth.

It is a measure of competition for access.

The larger the queue, the harder it becomes to move projects through it.

In many regions, most of that “pipeline” will never convert into actual generation.

The implication is simple:

The biggest queues are often the worst places to build.

Queue size is not a measure of growth.

It is a measure of competition for access.

The larger the queue, the more constrained the system.

And in many regions, most of that “pipeline” will never convert into actual generation.

The implication is simple:

The biggest queues are often the worst places to build.

Interconnection Queues

Interconnection queues have quietly become the single biggest constraint in the AI infrastructure buildout.

Everyone is focused on how fast data centers can be constructed. That’s the wrong variable. The real gating factor is how long it takes to get power connected—and in most U.S. regions, that timeline is measured in years, not quarters.

This is the bottleneck most investors and operators are underestimating.

You can deploy capital, secure land, and build capacity relatively quickly. But without interconnection approval, none of it comes online. The result is a growing mismatch between AI-driven power demand and actual usable supply.

If these timelines don’t compress, the constraint won’t be compute—it will be electricity.

The Backlog Is Accelerating

The constraint isn’t just that interconnection queues are long—it’s that they’re getting longer, faster.

Each year, more generation enters the queue than actually gets built and connected. What looks like a delay problem on the surface is, in reality, a system that is falling further behind over time.

This is where the dynamic becomes structural.

When inflows consistently exceed completions, the backlog doesn’t stabilize—it compounds. Timelines stretch, uncertainty increases, and capital gets trapped in projects that can’t reach commercial operation.

If this trend continues, the limiting factor for AI infrastructure won’t be demand or capital.

It will be the grid’s ability to keep up.

The Backlog Is Multiples of Demand

The scale of the backlog isn’t just large—it’s disproportionate to the system it’s trying to connect to.

In several U.S. power markets, the amount of generation waiting in interconnection queues is multiple times larger than the current peak demand those grids serve.

Backlogs at this scale are structurally misaligned with reality.

In several U.S. power markets, the volume of generation waiting in interconnection queues is multiple times larger than current peak demand. That doesn’t represent future capacity. It represents a filtering problem.

Most of these projects will never be built, but they still consume study time, planning resources, and grid attention. The result is a system where congestion exists not just in infrastructure, but in the process itself.

The constraint isn’t just supply. It’s the system deciding what supply is real.

Where AI Can Be Built

The constraints are not evenly distributed.

While interconnection delays and queue congestion are widespread, some regions still have structural advantages—faster timelines, available capacity, and more flexible market design.

This is where the next wave of AI infrastructure will concentrate. Regions with shorter interconnection timelines and available power capacity will attract disproportionate investment, while constrained markets risk falling further behind regardless of demand. The geography of compute is increasingly becoming the geography of electricity.

This reveals something the heatmap cannot: how close each region is to tipping into the next level of constraint. Most people focus on score changes — that’s a lagging indicator. The real signal is proximity to the next threshold.

Where the Edge Is

The methodology is public. The data is observable. But the signal — where constraints are tightening before the market reacts — is not. Most people see the current state of the grid. We track how that state is changing, and where it is about to break. That’s the difference between reacting to constraints and positioning ahead of them.

The grid is tightening faster than most models update. By the time a constraint shows up in forecasts, capital has already been misallocated. We track those shifts in real time — where queues are building faster than they’re clearing, where timelines are extending before they’re reported, where regions are approaching constraint thresholds before the score changes. That’s where the edge lives.

Inside the Numbers

Every score in the index traces back to observable inputs.

Queue depth, wait times, withdrawal rates, demand pressure, and capacity margins are all derived from publicly available data—normalized and mapped against fixed thresholds.

We don’t publish the full dataset because the value isn’t in the raw numbers. It’s in how they’re interpreted.

What matters is not the exact figure—it’s how close each region is to constraint.

What We Don’t Know

No index is perfect. Queue data captures intent, not execution — most projects in the queue will never be built.

Non-ISO regions rely on proxy data that is directionally accurate but less granular.

EIA data lags real-time changes, and rapid shifts take time to appear. These are not flaws. They are conditions to understand when interpreting the signal.

Next Week: PJM — America’s Largest Grid Under Pressure

We go deeper into the largest U.S. power market—and what its constraints mean for AI infrastructure in the eastern corridor.

The AI Grid Report

Powering the infrastructure behind the AI economy.

This article explains how the AI Grid Constraint Index is built.

The full index—updated weekly across U.S. grid regions—is where the signal actually sits.

Subscribe now

Subscribe now

Oil prices can move overnight.

Electrification demand can follow quickly.

But the infrastructure required to support that demand—generation, transmission, interconnection—does not.

That mismatch is no longer abstract.

It is already determining where projects clear, where they stall, and where capital flows.

The assumption that demand drives outcomes is breaking.

The grid decides what scales.

Early Scale

• Data center campuses now require 500 MW–1 GW of continuous load

• U.S. interconnection queues exceed 2,000 GW

• Transmission build timelines stretch 7–12 years

Big Tech is deploying ~$600B+ into data center infrastructure against a system that cannot absorb it on the same timeline.

Demand is accelerating into a system structurally incapable of responding at the same speed.

Why This Matters

For nearly two decades, U.S. electricity demand was flat. That era is over.

Demand is no longer the constraint.

Absorption is.

Multiple forces are now pushing demand higher simultaneously:

• AI data center load

• Electrification of transport and heating

• Industrial reshoring

The disruption in the Strait of Hormuz has added a new forcing function.

Brent crude has moved above $100.

A chokepoint responsible for ~20% of global oil flows has been partially shut.

OIL SHOCK CHART - BRENT VS. HORMUZ

The Shock

The Strait of Hormuz—conduit for roughly 20% of global oil flows—is now open only to vessels with Iranian escort with transit tolls paid in RMB or stablecoin…or a backstage pass from the IRGC.

This is not a supply problem. It is a logistics constraint.

Producers can still produce, but they can’t move the product. Logistics constraints don’t resolve quickly. A short disruption drives temporary substitution; a prolonged disruption drives structural change. The market is attempting to price both at once.

Market Signal

WTI vs Brent Broke a Pattern.

What began as a dislocation in geography sent Brent Crude up, Asian crude higher, and West Texas Intermediate (WTI) lower.

Then Brent and WTI flipped. The replacement for lost Middle East supply wasn’t prompt North Sea production. It was US exportable crude.

WTI became the marginal scalable barrel.

But the US itself imports ~750k bpd from the Middle East and Atlantic Basin. With those imports stalled, domestic refiners went short — bidding up the same barrels the export market wanted.

WTI is now being pulled from both ends: Export substitution demand and tighter domestic balances.

The Shift

When fuel prices spike, electrification accelerates. That part is predictable.

The economics reprice quickly. Electric alternatives become more competitive as fuel prices rise. EV cost-of-ownership improves. Heat pumps and induction systems gain share. Import-dependent regions accelerate substitution faster.

Electrification is a demand response to volatility.

Demand is not the constraint; infrastructure is. And that’s where the market starts to break, because what’s happening now is not just a demand increase.

It’s a pricing problem.

The market is attempting to price two incompatible timelines at once:

• short-term disruption

• long-term structural shift

That tension is already visible in how crude is trading. Not because production has changed, but because logistics has.

Geography—not just supply—is beginning to determine price, and once geography drives pricing, infrastructure becomes the constraint layer capital has to navigate.

Most capital isn’t positioned for that.

It is still flowing toward demand density—not infrastructure readiness.

That mismatch is where the system fractures.

Electrification doesn’t remove constraints.

It relocates them.

From oil…to electricity.

THE SPEED GAP — DEMAND VS INFRASTRUCTURE

Capital Signal

Capital is concentrating in the most constrained parts of the grid.

PJM alone accounts for over $100B in data center investment, despite operating within one of the most congested interconnection regimes in the U.S.

Queue timelines extend multiple years, transmission limits are binding, and projects increasingly require significant network upgrades before power can be delivered.

This is not an edge case. It is becoming the baseline.

Projects modeled on ideal timelines are slipping years. A solar-plus-storage project expected online in 2027 can realistically energize closer to 2030.

Data center campuses are facing hundreds of millions in grid upgrade costs before a single megawatt is delivered.

This is not a demand problem. It is a system constraint.

Capital is underwriting growth as if infrastructure is neutral. It isn’t. It is the gating function—and that creates a structural mispricing be

The Transition

The gap is already visible. Demand compounds in months. Infrastructure responds in years. That gap doesn’t stabilize — it widens.

The grid is not failing. It is filtering, delaying, prioritizing, rejecting.

What matters now is not how fast demand grows. It’s where that demand can land — and where it can’t.

This Week’s Constraint Index:

The distribution is uneven. Buildable regions are limited, constrained regions dominate, and saturated regions are already rejecting new demand. This is not a scaling environment — it is a filtering system.

The constraint is no longer theoretical. It is already determining where projects get built, where capital flows, and which regions can actually absorb demand. Most analysis stops at “demand is growing.” That’s not the edge. The edge is knowing where the grid can support it — and where it can’t.

In this week’s issue, we break down:

• the regions where the constraint is already binding

• which markets can still absorb load—and why

• where capital is mispriced relative to infrastructure reality

This Thursday, we’ll break down the full methodology behind how the Constraint Index is constructed and what each component is actually measuring.

➡️ Full analysis available to subscribers

Subscribe now

Read more

If you enjoy The AI Grid Report, it would mean the world to me if you invited friends to subscribe and read with us. If you refer friends, you will receive benefits that give you special access to The AI Grid Report.

How to participate

1. Share The AI Grid Report. When you use the referral link below, or the “Share” button on any post, you'll get credit for any new subscribers. Simply send the link in a text, email, or share it on social media with friends.

Refer a friend

2. Earn benefits. When more friends use your referral link to subscribe (free or paid), you’ll receive special benefits.

• Get a 1 month comp for 5 referrals

• Get a 3 month comp for 10 referrals

• Get a 6 month comp for 20 referrals

Visit the leaderboard

To learn more, check out Substack’s FAQ.

Thank you for helping get the word out about The AI Grid Report!

Two Out of Three Power Projects in America’s Largest Grid Never Get Built

Two out of three power projects never get built.

Not because they fail. Because the system rejects them.

The AI boom has a constraint most of the market isn’t pricing.

It’s not chips.

It’s not capital.

It’s permission to connect to the grid.

In the largest U.S. power market, two out of every three projects that apply to connect eventually withdraw.

Not because the economics failed.

Not because financing fell through.

Because the process itself broke them.

That is the grid your AI infrastructure depends on. And that failure rate is not an anomaly. It’s the system working exactly as designed.

There are currently:

– ~2.3 terawatts of capacity waiting in interconnection queues

– ~1.2 terawatts installed across the U.S. grid

– More than 25,000 active projects

– A median wait time of 3–7 years

– Withdrawal rates exceeding 60% in some regions

The market is PJM Interconnection.

Thirteen states from New Jersey to Illinois.

65 million people.

The grid that powers the mid-Atlantic, the industrial Midwest, and the nation’s capital.

Its interconnection queue has a 67.4% withdrawal rate.

In California’s CAISO, the number is 75.6%.

Three-quarters of projects in the most aggressive clean energy market in the country never make it through.

The queue is nearly twice the grid.

2,300 gigawatts of generating capacity is waiting for permission to connect.

1,200 gigawatts is installed today—nationwide.

There are 9,251 projects in PJM alone (active + withdrawn + operational).

67.4% (of all projects that have entered historically) will withdraw.

CAISO: 75.6%.

ERCOT: 0.0%.

We’ll come back to ERCOT (below the paywall).

A single AI training campus requires 500 MW to 1 GW of continuous power.

That is a mid-sized city.

Microsoft, Alphabet, Amazon, and Meta are deploying roughly $650 billion annually into infrastructure.

Almost all of it is pointed at data centers.

A transmission line takes 7 to 12 years to complete.

A data center takes two to three.

Why The Queue Exists — and Why It’s Breaking

You can’t plug a 500-megawatt power plant into a transmission network the way you plug a laptop into a wall socket.

Before generation can deliver power, it must pass through a formal interconnection process — engineering studies that determine whether the grid can absorb the new capacity without browning out your Netflix stream.

Power flows are governed by physics. If you connect a new plant that destabilizes voltage or frequency, it can bring down an entire region.

The studies exist for a reason.

But the process was designed for a world where a handful of new plants connected each year. It was not designed to handle the current volume hitting the system.

The result is not a backlog. It is a structural filter — one that selects not for the best projects, but for the deepest pockets and the most patience.

The Scale of What’s Stuck

There are currently 25,018 projects in active interconnection queues across U.S. ISOs and non-ISO regions. Together, they represent 2,300 gigawatts of generating capacity — nearly double the entire installed capacity of the United States.

Not all these projects will be built. Many are speculative. Some are duplicative filings by the same developer hedging across grid regions. But the withdrawal rates tell you something the raw numbers don’t.

Developers enter, pay study fees, wait through multiple study cycles, and leave. In some regions, the majority never connect. The median project has been waiting years. Most will never reach commercial operation.

A hyperscale data center does not need 5 megawatts in a convenient location. It needs 100 to 500 megawatts of continuous, uninterruptible power in a specific place, on a specific timeline, with contractual penalties if the lights do not turn on when the lease begins.

That kind of demand is concentrated, inflexible, and time-sensitive. When it collides with a queue system designed for distributed, flexible, patient demand, the result is not a queue.

It’s a wall.

The Constraint Isn’t Evenly Distributed

What matters isn’t just that the grid is constrained.

It’s where.

Some regions are processing capacity.

Others are absorbing demand they can’t convert into power.

That difference determines where infrastructure actually gets built.

This is what we’re building:

The AI Grid Constraint Index

A system to track:

– where capacity is actually available

– where projects are getting through

– where the grid is becoming a bottleneck

Below is a simplified view.

Preliminary View — Full Index Available to Subscribers

Full Index + Regional Breakdown → Subscribers

The constraint is visible.

But it isn’t evenly distributed.

And that’s where the real signal is.

AIGR is normally a paid report.

But for our launch, we’ve unlocked this issue.

Subscribe for free to access full future article.

Referral Incentives

If you’re finding this useful, you can unlock premium access by referring others.

• 1 referral → 1 month free

• 3 referrals → 3 months free

• 5 referrals → 6 months free

If you’re reading this as a subscriber, your personal referral link is always included in the email version of this report.

If you’re not yet subscribed, we highly encourage you to join below to get access to your subscriber perks.

Subscribe now

Inside the Constraint

Subscribers get access to the full AI Grid Constraint Index:

– where projects are actually clearing

– where delays are compounding

– where capacity is still available

The AI Grid Constraint Index

Each month, this newsletter publishes a heatmap scoring nine US grid regions across five constraint dimensions: queue depth, wait time, withdrawal rate, demand pressure, and capacity margin. Scores run from 1 to 5.

Think of it as a traffic report for the power grid. Green means capacity is available. Red means projects stall. Dark red means the system is structurally failing.

The full methodology — exact thresholds, data sources, and scoring assumptions — publishes in Issue 02. If you are making infrastructure decisions against these scores, you should know exactly how they are calculated.

The constraint we just discussed isn’t evenly distributed.

It concentrates.

Some regions process capacity. Others absorb demand they cannot convert into power. The difference determines where infrastructure actually gets built.

This Week’s Reads

PJM scores the highest overall constraint. A 254 GW queue, 9,251 total projects, and a 67.4% withdrawal rate combine to produce the worst interconnection environment in the country. If your AI power strategy depends on the mid-Atlantic grid, the process itself is your primary risk.

ERCOT tells a different story — or appears to. Its queue holds 283 gigawatts of active projects across 1,832 applications, with a withdrawal rate near zero.

Projects in Texas enter the queue, move through it, and connect. But connection is not delivery. The substations, transformers, and transmission lines that serve current demand are being asked to absorb 5.6 times their capacity in new generation.

When the transmission path to load centers is saturated, connected projects hit curtailment, congestion, and negative nodal prices. Your wind farm clears the queue, hooks up to the grid, and discovers its power has nowhere to go.

Speed of process does not eliminate the constraint. It relocates it — from the queue to the operating grid, where it shows up not as withdrawal but as electrons that can’t reach the customer willing to pay for them.​​​​​​​​​​​​​​​​

CAISO presents a policy paradox. California mandates the most aggressive clean energy buildout in the nation. Its interconnection queue has a 75.6% withdrawal rate. Three-quarters of projects enter, study, wait, and leave. The policy ambition and the process reality are moving in opposite directions.

The Western non-ISO region — anchored by the Bonneville Power Administration across the Pacific Northwest and Mountain West — carries the single largest queue by volume at 369 GW. Much of this is wind and solar seeking access to transmission corridors that were built decades ago for hydroelectric power. The resource is there. The wires are not.

MISO faces a unique compounding risk. A 230 GW queue, significant withdrawal rates, and the highest retirement exposure in the country at 32% of demand equivalent. The Midwest is simultaneously losing old generation and failing to connect new generation.

ISO New England is the calmest region on the board. Just 15 GW in its active queue, a 1.0x demand pressure ratio, and modest retirement exposure. But New England’s constraint is different: it has limited capacity to absorb large-scale AI load in the first place. A calm queue on a small grid is not necessarily an opportunity.

The Mismatch

AI scales in: → quarters.

Power scales in: → decades.

Training clusters are deployed in months. Inference capacity is added in weeks. A hyperscale data center can be designed, permitted, and constructed in two to three years. A transmission line takes seven to twelve years from proposal to energization. A large natural gas plant takes four to six. Even solar and battery projects face three to five years in the interconnection queue before construction can begin.

The gap between those two curves is not a temporary mismatch. It is a structural feature of how these two systems operate. By the time the wire is approved, the data center could have been operating for a decade.

And the gap is widening. AI capital expenditure is accelerating. Grid reform is, at best, incremental.

The Compounding Pressure

The queue is only half the problem. While the new generation waits to get in, the old generation is walking out the door.

Across the United States, aging coal and natural gas plants are retiring on fixed schedules. These retirements are driven by economics, regulation, and in many cases, the physical age of the equipment. They are not discretionary. They are not contingent on what replaces them. They are happening regardless.

In MISO, 32% of average demand equivalent is scheduled for retirement — the highest of any region we track. Coal plants are closing on known dates. The solar and storage projects meant to replace them are stuck in a process that takes years to complete.

This creates compounding pressure. The queue delays new generation. Retirements remove existing generation. And demand — driven by data centers, electrification, and industrial reshoring — continues to grow. The three forces are moving in the same direction, and none of them is slowing down.

The Question

You have been modeling power as a solvable input. Model the demand, finance the plant, build the data center, connect to the grid.

The 2,300 gigawatt backlog disagrees.

The reality:

→25,000 projects. A 2.3 terrawatt backlog. A process that destroys two-thirds of the projects that enter it.

You can’t bid your way to the front of a physics-limited queue. You can’t buy transmission capacity that does not exist. You can’t accelerate an engineering study by offering to pay more for it.

Capital is already migrating. Developers are not asking where AI wants to build. They are asking where the grid will let it. The regions that answer that question correctly will attract the next decade of infrastructure capital. The regions that don’t will watch it leave.

The constraint is not priced. The reallocation has already begun.

This publication exists to answer that question before the market does.

Where AI wants to build is easy to find. Everyone knows. The data feeds tell you.

Where it can—that’s the only question that matters.

Next Week: Behind the Scores

Next week we’ll break down the full methodology behind the AI Grid Constraint Index — the thresholds, data sources, and assumptions that drive each score.

Because if these numbers are going to influence capital allocation, you need to understand exactly how they’re built — and where they break.

Introduction

The biggest constraint on artificial intelligence is not chips.

It’s electricity.

Training frontier AI models already requires hundreds of megawatts of power, and the next generation of compute clusters could demand gigawatts.

That is the equivalent electricity consumption of entire cities.

Yet while AI demand is scaling exponentially, electrical infrastructure expands slowly — often over decades.

The result is a growing power bottleneck that will shape where the next generation of AI infrastructure gets built.

Understanding this constraint is essential for investors trying to make sense of the AI boom.

AI Compute Is Becoming Gigawatt-Scale

The scale of AI training infrastructure has expanded dramatically in just a few years.

Early machine-learning clusters consumed only a few megawatts of electricity. Today’s frontier AI models require vastly larger compute environments supported by massive data center campuses.

The shift is occurring at remarkable speed.

Large AI training clusters now rival heavy industrial facilities.

The largest planned AI campuses are expected to require 1 to 5 gigawatts of electricity.

To put that in perspective:

• 1 gigawatt can power roughly 750,000 homes

• 5 gigawatts approaches the electricity demand of entire metropolitan regions

AI infrastructure is quickly becoming one of the most energy-intensive sectors of the global economy.

Electricity Infrastructure Moves Slowly

While AI demand is expanding rapidly, the physical systems that supply electricity evolve much more slowly.

Electrical infrastructure projects require years of permitting, construction, and grid integration.

Power plants and transmission networks cannot be scaled overnight.

Electricity infrastructure expands over decades, while AI data centers can be deployed in just a few years.

This mismatch between AI growth timelines and energy infrastructure timelines is now becoming one of the defining constraints of the AI economy.

While new AI facilities can be deployed in just a few years, the electrical systems required to support them often take a decade or more to expand.

As a result, AI infrastructure must concentrate in regions where sufficient electricity already exists.

The Geography of AI Compute

Because electricity infrastructure cannot expand quickly everywhere, AI infrastructure is clustering in locations that already possess abundant energy resources.

These regions share several important characteristics:

• reliable electrical grids

• large existing power generation capacity

• competitive industrial electricity prices

• favorable infrastructure permitting

Regions that meet these criteria are emerging as global AI compute hubs.

These regions already possess the large-scale electrical infrastructure necessary to support massive AI data centers.

In many cases they also benefit from:

• hydroelectric power

• natural gas availability

• cooler climates

• existing data center ecosystems

As a result, they are increasingly attracting the next wave of AI investment.

AI Data Centers Are Becoming Industrial-Scale Energy Consumers

The electricity requirements of AI infrastructure are now approaching those of major industrial sectors.

Large AI campuses can require power comparable to:

• aluminum smelters

• steel manufacturing plants

• large petrochemical complexes

In other words, AI is becoming an industrial energy consumer.

This shift represents a profound transformation in how digital infrastructure interacts with energy systems.

For decades, the internet economy consumed relatively modest amounts of electricity compared to traditional industry.

Artificial intelligence is changing that equation.

The Strategic Implication

The global race for AI will not be decided only by:

• semiconductor manufacturing

• software innovation

• venture capital

It will increasingly be decided by something far more fundamental.

Electricity.

Countries and regions that possess abundant power generation capacity, reliable grids, and scalable energy infrastructure will enjoy a structural advantage in hosting the next generation of AI compute facilities.

The power bottleneck is no longer theoretical.

It is already shaping the geography of the AI economy.

Conclusion

Artificial intelligence is often described as a software revolution.

But beneath the algorithms lies a much more physical reality.

AI runs on electricity.

As demand for compute accelerates, the world’s electrical infrastructure will become one of the most important constraints on technological growth.

For investors, policymakers, and technology companies alike, the lesson is becoming clear:

The future of artificial intelligence will be built not only in code — but in power plants, transmission lines, and electrical grids.

If you’re new here:

AI Grid Report explores how artificial intelligence is reshaping global energy systems.

Most coverage of AI focuses on models, startups, and software.

But the expansion of AI infrastructure depends on something far more physical:

electricity.

Training models, running inference, and operating hyperscale data centers all require enormous power supply and reliable grid infrastructure.

This publication examines the intersection of:

• AI infrastructure

• electricity demand

• power markets

• grid expansion

• infrastructure investment

If these topics interest you, consider subscribing.

Subscribe now

Now let’s look at the problem emerging at the center of the AI boom.

The Hidden Constraint Behind AI

Artificial intelligence is widely described as a computing revolution.

But every computation ultimately runs on electricity.

And the scale of power required by modern AI systems is growing at a pace that few energy systems were designed to support.

Consider a few numbers:

• A single hyperscale data center can consume 100–500 megawatts of electricity.

• The largest AI campuses currently planned may require 1 gigawatt or more of continuous power.

• That is roughly equivalent to the electricity consumption of 750,000–1,000,000 homes.

Multiply that by dozens of facilities being planned across the United States, Europe, and Asia.

The result is a surge in electricity demand that is beginning to challenge existing grid infrastructure.

Figure 1 — The global race to build AI infrastructure is accelerating as governments and technology companies compete for compute capacity.

AI Electricity Demand Is Exploding

Several recent projections illustrate the scale of the coming demand.

According to the International Energy Agency, global data center electricity consumption could more than double by the end of the decade.

Key estimates include:

• Data centers currently consume roughly 2–3% of global electricity.

• AI workloads could push that number significantly higher by 2030.

• In the United States alone, data center electricity demand could reach 9–12% of total power consumption within the decade.
  

  Much of this growth is driven by the rapid expansion of AI training clusters and inference infrastructure.

Training a large model can require tens of thousands of GPUs operating simultaneously for weeks.

But inference — running the model for users — creates the long-term electricity demand.

And inference must run continuously.

The Grid Was Not Designed for AI

Electric power systems expand slowly.

Building new infrastructure requires:

• multi-year permitting processes

• transmission construction

• generation development

• regulatory approval

Transmission lines alone can take 10–15 years to permit and build.

Meanwhile, the AI industry is expanding infrastructure on a much faster timeline.

This mismatch is already creating pressure.

Grid operators across several regions have reported:

• growing interconnection queues

• delays connecting new data centers

• local power shortages in high-demand areas

In some cases, utilities are warning that new data centers may need to wait years before sufficient electricity capacity becomes available.

Figure 2 — Global Data Center Electricity Demand Projections

The Next Bottleneck

For much of the past two years, the main constraint on AI development was chips.

GPUs became the scarce resource.

But the next constraint may be more fundamental.

Even with unlimited computing hardware, AI infrastructure still requires:

• reliable electricity supply

• grid capacity

• cooling systems

• long-term power contracts

In other words:

AI infrastructure is becoming tightly linked to the energy system.

And that means electricity markets, utilities, and infrastructure investors are moving toward the center of the AI economy.

The Bigger Shift

For decades, the technology industry largely treated energy infrastructure as a background condition.

Power was assumed to be available.

The rise of large-scale AI is changing that assumption.

Electricity is becoming a strategic resource for the digital economy.

The companies and regions capable of securing large, reliable energy supply may gain a significant advantage in building the next generation of AI infrastructure.

Which leads to the central question explored in the next article:

Where will the next wave of AI data centers actually be built?

What’s Next

If you’re interested in the intersection of:

• artificial intelligence

• electricity infrastructure

• energy geopolitics

• and long-term investment trends

consider subscribing to AI Grid Report.

Subscribe now

This publication explores how the global power system is being reshaped by the rise of artificial intelligence.

New articles are published regularly as we examine the infrastructure behind the AI economy.

Everyone is talking about AI.

The models.

The startups.

The trillion-dollar race to build the future.

But most people are looking in the wrong place.

Because the real constraint on AI isn’t software.

It’s the power grid.

AI is being treated like a software revolution.

Something you scale with better code, bigger models, and more investment.

But AI isn’t just software.

It is one of the most energy-intensive technologies ever deployed at global scale.

And the infrastructure required to power it — power plants, transmission lines, substations, and data centers — moves far more slowly than the technology itself.

In other words, the future of AI may not be determined by algorithms.

It may be determined by electricity.

The AI Economy Runs on Electricity

Artificial intelligence ultimately runs on physical infrastructure.

Applications depend on models.

Models depend on compute.

Compute depends on data centers.

Data centers depend on electricity.

Remove electricity from that chain and the entire AI economy stops.

This framework illustrates the full stack of the AI economy — and why electricity sits at the bottom of everything.

The Coming Electricity Shock

For most of the past two decades, electricity demand in developed economies barely grew.

Efficiency improvements offset new consumption.

LED lighting replaced incandescent bulbs.

Computers became more efficient.

Manufacturing moved overseas.

Utilities built their systems around this assumption.

Demand would remain relatively stable.

But AI is breaking that trend.

Training and running large AI models requires enormous amounts of computing power.

And computing power consumes electricity.

Lots of it.

This chart illustrates how AI could restart electricity demand growth after decades of stagnation.

If large-scale AI adoption continues, the electrical grid may face the largest new demand expansion in a generation.

The Power of One Data Center

Most people still underestimate how much electricity modern AI infrastructure consumes.

Large hyperscale AI campuses can require extraordinary amounts of power.

Not megawatts.

Hundreds of megawatts.

Sometimes even gigawatts.

This comparison shows that a single AI data center campus can consume as much electricity as a small city.

In some cases, it approaches the power consumption of major industrial facilities.

Which raises an uncomfortable question.

How many of these facilities will the grid actually be able to support?

The End of the Flat Electricity Era

For most of the past twenty years, electricity demand remained relatively flat in developed economies.

Utilities planned their infrastructure around that stability.

But AI changes the equation.

If AI adoption accelerates, electricity demand could begin rising again — potentially faster than the power system was designed to handle.

This chart shows how AI could become one of the largest new drivers of electricity demand in decades.

The implications extend far beyond technology companies.

They affect the entire global energy system.

The Physical Limits of AI

Building software is fast.

Building infrastructure is slow.

Training a new AI model can take months.

Building a new power plant can take years.

Building new transmission infrastructure can take a decade or more.

Permitting processes alone can take longer than the development cycle of most startups.

And data centers themselves face additional constraints:

• available land

• access to water for cooling

• grid interconnection capacity

• local permitting and political approval

Scaling AI is not just a technological challenge.

It is an industrial challenge.

Where AI Data Centers Will Be Built

If electricity becomes the bottleneck for AI, the geography of the AI economy may change.

The most important locations may not be where the best engineers live.

They may be where the most electricity is available.

Chart 5 — Where AI Data Centers Will Be Built

This chart highlights regions with favorable conditions for large-scale AI infrastructure.

Electricity availability may become one of the primary factors determining where AI infrastructure is deployed.

The Largest Electricity Machine Ever Built?

If AI adoption continues accelerating, humanity may be building one of the most electricity-intensive systems ever created.

Every major industrial system runs on energy.

Manufacturing.

Transportation.

Heavy industry.

The global internet.

AI could soon join that list.

And unlike many previous technologies, AI scales directly with computing power.

More models.

More inference.

More data centers.

More electricity.

The world has never built a digital system that requires this much continuous power.

Yet most of the AI conversation still focuses on software.

Very few people are asking the infrastructure question.

Can the grid keep up?

Why We Created AI Grid Report

This is exactly why we launched our publication, The AI Grid Report.

To track the intersection of:

• artificial intelligence

• electricity infrastructure

• energy markets

• geopolitics

• and the global race to power the AI economy

Because the future of AI will not be determined only by software.

It will also be determined by the physical systems that power it.

Final Thought

AI may reshape the global economy.

But before it can do that…

it has to plug into the grid.

And the grid wasn’t built for what’s coming.

Subscribe to The AI Grid Report

If you want to understand the energy infrastructure behind the AI revolution, subscribe to AI Grid Report.

We’ll be mapping the system that will determine how fast the AI future actually arrives.