Published: September 16, 2026 | Reading Time: ~10 minutes | Channel: technology
Everyone fixates on the wrong bottleneck. The AI conversation of 2026 has been a loud argument about chips, capex circularity, and whether the bond market will finance the buildout. Meanwhile, the actual choke point walked onto the stage with a press release: governors pausing grid hookups, capacity auctions clearing at the maximum allowable price, and a federal forecast that U.S. electricity demand will set records two years running. The AI industry doesn't have a compute problem or a money problem. It has an electrons problem — and electrons don't care how much venture funding you raised.
Here's the number that should reorganize your mental model of the AI trade: 474 gigawatts of interconnection requests are sitting in the Texas grid queue, roughly 90% of them data centers. That's more than five times the peak demand of the entire ERCOT system. Texas — the state that built its brand on saying yes — has stopped saying yes. If the most business-friendly grid in America is hitting pause, the constraint isn't regulation. It's physics wearing a regulatory costume.
On August 3, Texas Governor Greg Abbott directed the Public Utility Commission of Texas and the Electric Reliability Council of Texas to audit every new data center project seeking a grid connection — electricity and water demand, noise, lighting, tax incentives, ownership. New data center projects can't connect until that audit clears.¹ ²
The scale behind the directive explains the panic. In January, ERCOT's interconnection queue held 233 GW of projects waiting to connect. By August, that had more than doubled to 474 GW, with about 90% data centers.¹ ² ³ For context, that's more than five times ERCOT's total peak demand. A nontrivial share of those requests are what grid people call ghost demand — duplicate filings across multiple regions, developers holding queue positions while they shop for financing and customers. Goldman Sachs' Marcus research estimated in June that actual U.S. data center power demand is on track to grow from 31 GW in 2025 to 66 GW by 2027.³ Do the subtraction yourself: 474 GW of requests, 66 GW of real need. The queue is not a forecast. It's a lottery ticket stack.
And the ghost demand isn't harmless. Regulators can't tell the phantom 400 GW from the real 60, so they audit everything — and the audit takes months. Abbott's office said the ERCOT review is targeted for completion by December.³ Every month of audit is a month of delayed energization, delayed leases, delayed revenue recognition for anyone building in the state.
| Metric | Figure | Source |
|---|---|---|
| ERCOT interconnection queue, Jan 2026 | 233 GW | TechCrunch¹ |
| ERCOT interconnection queue, Aug 2026 | 474 GW (~90% data centers) | TechCrunch¹, WilmerHale/Hoodline³ |
| Queue vs. ERCOT total peak demand | >5x | TechCrunch¹ |
| Real U.S. data center power demand, 2025→2027 | 31 GW → 66 GW | Goldman Sachs Marcus³ |
| Data centers activating on time (within 4 quarters) | 72% historically | Goldman Sachs Marcus³ |
| U.S. data center share of electricity by 2030 | ~12% — roughly 6x the 2018 pre-boom share | Berkeley Lab via Fortune² |
| North American summer peak growth, next 10 years | +224 GW, 69% above prior forecast | NERC via Fortune² |
| Median time from grid request to commercial operation | 5 years | Berkeley Lab via Fortune² |
| PJM 2026/27 capacity auction clearing price | $329.17 per MW-day — maximum allowable, ~10x 2022 | PJM via EcoEngineers⁵ |
| PJM cleared capacity mix (latest auction) | Gas 45%, coal 22%, renewables 4% | PJM via EcoEngineers⁵ |
| Cumulative rate increases projected by 2030, PJM regions | +30–60% | EcoEngineers⁵ |
| U.S. record electricity demand | Both 2026 and 2027 | EIA via Reuters³ |
| Data center projects expected to miss their 1–2 year power timelines | 50–60% | Marsh (Kathryn Burke) via Fortune² |
| Korea semiconductor exports, August | Record $46.65B, +209% YoY (47.5% of all exports) | Korea MOTIR via CNBC⁴ |
Read that table as a system, not as trivia. Demand is compounding into a grid that expands in five-year increments.¹ The result is a widening gap between when AI companies promise capacity and when the grid can actually deliver it.
The demand side isn't hypothetical, either. Korea's August chip exports — $46.65 billion, up 209% year-on-year, 47.5% of the country's entire export book — are the physical evidence that the AI infrastructure demand is real, right now, upstream of every bottleneck downstream.⁴ Jeff Ng, head of Asia macro strategy at Sumitomo Mitsui Banking Corporation, estimates semiconductors accounted for nearly 80% of Korea's August export growth.⁴ Demand is ripping. Wires are the constraint. That mismatch is the entire story.
The standard rebuttal is that America has solved supply shocks before and will do it again. Fine — but run the timeline. U.S. power projects that came online in 2025 spent a median of five years between requesting a grid connection and commercial operation.¹ Generators are only half the problem. A gigawatt — enough to power roughly 750,000 U.S. homes — takes years of transmission, substations, and transformers before a single electron moves.¹ ² You can order a turbine faster than you can permit the wires to carry its output. This is not an argument about whether we'll build. It's an argument about which decade the electrons arrive.
Meanwhile the market is repricing scarcity in real time. PJM Interconnection — the RTO serving 65 million people from Chicago to Washington, D.C. — cleared its 2026/27 capacity auction at the maximum allowable price: $329.17 per megawatt-day, roughly ten times 2022 levels.⁵ MISO and Britain's capacity markets printed similar spikes.⁵ When an auction hits the price cap, the market isn't discovering a price. It's announcing a shortage.
Second-order effect nobody's talking about: the bottleneck has changed who's in charge. For a decade, data center developers picked sites; utilities accepted load. Now the queue is full, regulators are auditing, and grid operators are rewriting the terms. PJM has filed proposed rules with FERC that would require large new power users — read: AI data centers — to contract for their own new generation capacity or face curtailment during shortages.³ Texas is compiling audited dossiers on water use, noise, and ownership structures.¹ The party with the scarce asset sets the price, and the scarce asset is now the interconnection agreement.
And a third-order effect hiding in the reliability filings: NERC reported in July that sudden disconnections of large computational loads are creating new frequency-stability hazards, with multiple 2025 events where more than 1,000 MW of data center load dropped off the grid during transmission faults.³ AI's flagship workloads don't ride through disturbances the way industrial motors do. The grid was built for loads that flinch. Yours doesn't.
If you're building or investing in AI infrastructure:
The AI race quietly changed divisions this summer: it's no longer a compute race, it's a wire race — and the grid keeps score in years, not GPU-weeks. Underwrite energized megawatts, not announced ones; demand a signed interconnection date before you believe any capacity roadmap. The companies that internalize "electron time" will ship revenue on schedule. The ones still living on "chip time" will spend 2027 explaining their impairments.
All claims verified against Gold-tier (Reuters-reported EIA forecasts, PJM auction data, Korea MOTIR trade data) and Silver-tier (Fortune, TechCrunch, CNBC) sources. Each source URL was scraped and confirmed accessible on September 16, 2026. Two additional candidate sources (finance-commerce.com, utilitydive.com) returned 403 errors and were discarded per verification protocol. Bronze-tier data (EcoEngineers, citing PJM/Gartner/EIA primary documents) is labeled as such. Last verified: 2026-09-16.
Electricity doesn't do press releases. It does invoices — in years. ⚡
