AI Eats Electricity — and Then the Grid Fights Back

That 2,061 GW is GENERATOR interconnection to the transmission grid. It is explicitly not load interconnection, not distribution-connected, and not behind-the-meter. It is not the data-centre queue and must never be narrated as one.

Active natural gas capacity in US interconnection queues rose 86% during 2025 to 253 GW, while solar (773 GW, -19%), storage (749 GW, -16%) and wind (220 GW, -19%) all declined. Only about 10 GW of nuclear was active.

The IEA's own physical-bottleneck section lists bottlenecks in the plural, with transformers at 2-3 years, gas turbines at around 5 years, and grid connections at 5-10 years. Electricity is never singled out as THE binding one.

Wood Mackenzie's Q2 2025 benchmark put US power transformer lead times at 128.0 weeks and generator step-up transformers at 143.4 weeks — and both had DECLINED quarter over quarter, power transformers by 10.4 weeks.

Large power transformers were a long-lead item long before the AI cycle: a 2014 Department of Energy study already described them as requiring substantial capital and a lead time in excess of six months.

On GE Vernova's first-quarter 2026 earnings call, 22 April 2026, chief executive Scott Strazik said: 'in many of the cases with these projects, the gas turbines are really not the gating item. When you're talking about a three-year cycle from when a project starts, the EPC buildout, the permitting, the fuel availability.'

He said it twice on the same call, the second time unprompted: 'generally speaking, don't find our heavy-duty gas turbines to be the gating item on a directionally three-year cycle time right now.'

On the same call he put GE Vernova's own turbine lead time at 'directionally at about three years'. The turbines are not the gate because everything else takes exactly as long.

GE Vernova's gas equipment backlog and slot reservations reached 116 GW during the second quarter of 2026, up from 100 GW in a single quarter — but only 53 GW of that is firm backlog. The company expects to end 2026 with at least 125 GW under contract, 'making us mostly sold out through 30'.

Scope discipline on the quote: he said 'many', not 'most', and 'these projects' means his gas-power projects generally. Data centres are only around a fifth of that book.

On the BG2 podcast published 31 October 2025, Microsoft chief executive Satya Nadella said: 'the biggest issue we are now having is not a compute glut, but it's a power. And it's sort of the ability to get the builds done fast enough, close to power. So if you can't do that, you may actually have a bunch of chips sitting in inventory that I can't plug in. And in fact, that is my problem today. It's not a supply issue of chips. It's actually the fact that I don't have warm shells to plug into.'

Read the noun. What Nadella says is missing is not electrons — it is WARM SHELLS: finished buildings, with power and cooling, ready to plug into. That is construction, siting and energisation as much as generation.

NVIDIA's Form 10-Q for the quarter ended 26 July 2026 states: 'Securing land, power, and shell for data centers is a critical phase in the AI infrastructure buildout,' and discloses $56 billion of additional commitments — $36 billion of AI cloud agreements and $20 billion of third-party data-centre leases not yet commenced.

The same filing carries $3,529 million of notional guarantees on select AI cloud partners' land, power and shell data-centre lease obligations, with partners posting $712 million in escrow. NVIDIA discloses that the fair values of these credit derivatives 'were not significant', and the notional was essentially FLAT over six months.

In that same document NVIDIA also says: 'We will be shipping both Blackwell and Rubin systems in the future and are currently experiencing certain supply constraints.' It does not say which stage is constrained, does not quantify it, and does not rank it against power.

The IEA finds that 'in the second half of 2025 and beginning of 2026, global production capacity for high-end memory also emerged as a binding constraint on the rate of AI server production.'

And the sentence that decides its meaning: high-bandwidth memory production supports a maximum of around 25 GW of AI-ready servers per year through to 2027, 'broadly in line with the rate of capacity additions in the IEA's Base Case'. Memory and power pace each other.

The honest verdict on the top rung: there is testimony from a buyer, and there is filed corroboration that land, power and shell is a category worth guaranteeing with money. There is no filed disclosure attributing a specific delayed AI deployment to power.

About 549 GW of US capacity holds a draft or executed interconnection agreement without ever having reached commercial operations — 256 GW of solar, 161 GW of storage, 76 GW of wind and 45 GW of gas.

Even in that most-advanced cohort the buildable supply is overwhelmingly solar and storage. Gas is 8% of it and nuclear is absent.

Oncor, the largest Texas transmission utility, reported an interconnection request queue of 298 GW against an actual peak demand of 31 GW, with 44 GW under ERCOT evaluation in 'Batch Zero'.

A US data-centre policy tracker covering March 2025 to August 2026 records 56 formal public actions across 25 states: 21 moratoriums, 17 state legislative or regulatory actions, 7 enacted ordinances, 6 outright project rejections and 5 executive pauses. 34 were local and 22 statewide.

The chart counts ACTIONS, not megawatts. A ban in a rural county with no projects weighs exactly the same as a statewide interconnection freeze. Weighted by affected capacity it would collapse to roughly four jurisdictions.

PJM's Base Residual Auction cleared at $28.92 per megawatt-day for delivery year 2024/2025, then $269.92 for 2025/2026, $329.17 for 2026/2027, $333.44 for 2027/2028 and $325.00 for 2028/2029.

The last three prints cleared AT an administrative price cap, not at a market-discovered price. The fall to $325.00 for 2028/2029 is a LOWER COMPUTED CAP, not easing scarcity. The series flattens by construction.

New Jersey's Board of Public Utilities announced residential bill increases of 17.2% to 20.2% effective 1 June 2025. Its president, Christine Guhl-Sadovy, said: 'PJM's recent capacity auction results are the main driver of these increases.'

And the discipline on that number: by June 2026, EIA shows New Jersey residential prices up only 0.3% year on year. It was a one-time step, not a compounding trend.

And it runs the other way too. A Columbia University literature review finds load-growth effects on retail prices are mixed and context-dependent, citing an estimate that for PG&E every 1 GW of new data-centre load could REDUCE average household bills by 1 to 2 percent, by spreading fixed costs over more sales.

AEP has filed large-load tariff proposals in eight jurisdictions to make data centres pay their own way. Virginia approved Appalachian Power's in the second quarter of 2026: it applies to new loads of 150 MW aggregated or 100 MW individual, with contract terms up to 20 years and take-or-pay provisions written to protect existing customers.

New York Executive Order 62, signed 14 July 2026, holds DISCRETIONARY STATE environmental permits issued by the Department of Environmental Conservation for facilities of 50 MW or more. It does not touch local zoning, by-right development or federal permits, and it exempts applications the DEC had already deemed complete.

The order sets no expiration date. It lifts when the Department of Public Service submits its report on the final Generic Environmental Impact Statement, and no deadline is written for that study. The widely repeated 'one-year pause' is an outside characterisation, not a term of the order.

On 3 August 2026 Texas Governor Greg Abbott directed the Public Utility Commission and ERCOT to verify and audit every data centre advancing through ERCOT's interconnection process before any additional ones may move forward. ERCOT withheld its scheduled large-load classifications as a result.

The EIA cut its forecast for Texas electricity load growth in 2027 from 14% to 6% after that announcement, writing that it had lowered the forecast because of the pause.

AEP's authorised returns on equity sit in a 9.7% to 9.8% band across jurisdictions — West Virginia 9.75%, Kentucky 9.8%, Ohio 9.7% and 9.84%. Its FERC transmission vehicle REQUESTED a base return of 10.8%.

CoreWeave's Form 10-Q names specialised construction labour — electricians, mechanical engineers, HVAC technicians, commissioning specialists — as a shortage that could delay completion of new data-centre projects: 'Any inability to secure sufficient skilled labor on acceptable terms could delay the completion of new data center projects.'

The same filing gives power its own dedicated risk-factor heading, stating that limitations on generation, transmission and distribution 'may limit our ability to obtain sufficient power capacity for our potential expansion sites'. It names both, side by side.

Comfort Systems USA's revenue from Technology customers rose 105% to $1.917 billion and went from 43.0% to 58.7% of total revenue in a single year. Net income rose 91% to $441.6 million.

EMCOR's remaining performance obligations rose 43.9% to a record $17.14 billion, earned at a 10.6% operating margin.

Quanta Services grew revenue 41% to $9.56 billion and booked a record $53.4 billion backlog at a 7.3% consolidated operating margin — but that margin ROSE from 5.5% a year earlier, and operating income grew 88% against 41% revenue growth.

Margin capture across the electrical value chain in the quarter ended 30 June 2026: Vertiv 22.6% adjusted operating margin (+410bps), GE Vernova Power 18.8% segment EBITDA margin (+240bps), GE Vernova Electrification 18.4% (+390bps), Eaton 23.1% total segment margin (-80bps), Hubbell Utility Solutions 21.2% adjusted (-130bps), Hubbell Electrical Solutions 16.7% operating (-480bps), Atkore 22.2% gross (-120bps), Modine Data Centers 20.2% gross (-960bps), EMCOR 10.6% operating, Quanta 7.3% consolidated operating (+180bps), GE Vernova Wind -13.6% segment EBITDA (-630bps).

Year to date through 2 September 2026: Bloom Energy +138.1%, Powell +61.9%, Vertiv +58.5%, Quanta +43.4%, GE Vernova +37.8%.

The same complex sits far below its own post-2023 highs on the same day: NuScale about -84%, Oklo about -80%, Powell about -48%, Vertiv about -32%, GE Vernova about -23%.

Copper is only about 10% of the price of a power transformer. A 66% rise in the copper price raises the transformer price by about 7% — which the IEA judges 'unlikely to fundamentally undermine the economic viability of transmission and distribution projects.'

And the direction is wrong for the story anyway: copper mine production FELL 1.1% in the first half of 2026, to 11,341 kilotonnes from 11,468, with capacity utilisation dropping to 77.1% from 81.0%.

Aluminium is not the escape hatch either. Its conductivity is about 60% of copper's, and although a 3-to-1 price ratio was long treated as the substitution threshold, the ratio has exceeded 3.5 to 1 for much of the past five years without triggering a major tipping point.