PJM · ERCOT · FERC · The Frontier's Hidden Bottleneck
PJM has 220 gigawatts of proposed projects waiting on a grid whose summer peak demand is 154.
The race to build frontier AI is no longer rate-limited by chips, capital, or talent. It is rate-limited by transformers, gas turbines, and the seven-year average it takes a queued project to actually energize. ERCOT just revised its 2030 data-center demand projection from 29 GW to 77 GW in a single planning cycle. Approval rate on the queue: under 2%. The Stargate, Colossus 2, and Anthropic-Google compute build-outs all sit on the same constraint set.
The SignalFERC, May 2026
PJM must establish transparent rules to facilitate service of AI-driven data centers and other large loads co-located with generating facilities.
Federal Energy Regulatory Commission — directive to PJM, 2026
The MathERCOT 2026 planning cycle
Just over 3 GW of large-load studies were approved during a one-year period through June 2025 — out of a 156 GW queue proposed by 2030. Approval rate: 1.9%.
ERCOT planning data, summarized — Data Center Frontier May 2026
The Gigawatt Bottleneck
AI's 2026 hard constraint isn't compute — it's the seven-year wait between gigawatt queue entry and gigawatt-on-the-bus. A forensic accounting of the load forecasts, the supply backlogs, and the workarounds.
PANEL VERDICT — FORENSIC FRAME
✓ EVIDENCE SHOWS
that PJM's interconnection queue contains approximately 220 GW of proposed projects against a 2025 summer peak demand of ~154 GW. ERCOT raised its 2030 data-center demand forecast from 29 GW to 77 GW in a single planning cycle (2.7× revision). ERCOT's large-load queue holds ~156 GW proposed by 2030 with ~3 GW approved in a one-year window — a 1.9% conversion rate. National data-center IT load is projected to grow from 80 GW (2025) to 150 GW (2028) on a ~760 GW national peak — a structural ~20% increase. FERC has formally directed PJM to establish new co-location rules.
⊘ NOT YET CONFIRMED
how much of the announced AI compute build-out (Stargate's 7 GW target, Colossus 2 at 2 GW, Anthropic's >1 GW TPU capacity) will route through the formal interconnection queue versus being satisfied by behind-the-meter (BTM) generation. Microsoft's Three Mile Island Unit 1 PPA with Constellation, Stargate Abilene's Oracle-led campus arrangement, and xAI's on-site Memphis turbines all indicate hyperscalers are buying their way around the queue. The headline queue numbers may overstate the actual constraint on frontier compute deployment.
⚠ STRONGEST STRUCTURAL COUNTER
that the supply-side bottleneck is meaningfully relievable by mid-2028. GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries are scaling heavy-duty gas turbine production. Three nuclear restart deals signed in 2024-25 (TMI 1, Palisades, Duane Arnold studies). The Inflation Reduction Act and 2025 grid permitting reforms accelerate large-power transformer (LPT) deliveries. The optimistic case: by 2028-29, BTM gas, restarted nuclear, and on-site SMRs absorb 30-50 GW of AI load that would otherwise sit in the public queue. The bottleneck "moves" rather than "binds."
◯ OPEN EVIDENTIARY QUESTION
what fraction of the 220 GW PJM queue is genuinely committed capital versus speculative options. Interconnection queues globally carry a high "shadow ratio" — projects that submit but never advance. Historically only 15-25% of queued projects energize. If that historical ratio holds for the 2024-26 AI surge, the *effective* load growth is closer to 33-55 GW from the PJM queue, not 220 GW. PJM has not yet published a refined committed-vs-speculative breakdown.
TIMELINE — HOW THE BOTTLENECK BUILT
2007–2019
US electricity demand essentially flat (CAGR <1%)
For two decades, US grid load grew at well under a 1% compound annual rate. Efficiency gains in lighting, appliances, and industrial processes offset population and GDP growth. Utilities planned around flat demand. Transmission and generation investment slowed. Workforce hollowed.
2022 onward
Generative AI shifts data-center load curve
ChatGPT launches late 2022. AI training compute demand begins doubling on ~6-month cycles. By 2024, hyperscalers are filing interconnection requests at scales utilities haven't seen since the 1970s nuclear build.
January 2025
Stargate Project announced — $500B trajectory
OpenAI, SoftBank, Oracle, MGX announce $100B initial commitment, $500B by 2029. Abilene, Texas flagship campus: 450,000 GB200 GPUs, 1.2 GW power draw on a 15-year Oracle lease. Total Stargate trajectory ~7 GW across sites. Power draw on each campus comparable to a small nuclear plant.
Memphis complex becomes the largest single-site AI installation on Earth. Includes on-site gas turbines to bridge the gap between local utility (MLGW) supply and instantaneous training load. xAI's pattern — buying generation rather than waiting for utility allocation — is replicated across hyperscaler buildouts. Cross-ref: [[entity:elon musk]].
2024–2025
Nuclear restart deals: Three Mile Island, Palisades
September 2024: Constellation Energy signs 20-year PPA with Microsoft to restart Three Mile Island Unit 1, rebranded "Crane Clean Energy Center," with target online 2028. Holtec moves to restart Palisades (Michigan) by 2025 with DOE loan support. Duane Arnold (Iowa) feasibility studies. The "behind-the-meter via shuttered reactor restart" template emerges. Cross-ref: [[entity:Constellation Energy]].
2025
FERC directs PJM on co-located large loads
FERC issues directive requiring PJM to establish transparent rules for co-located generation serving large loads. The order recognizes the regulatory grey zone hyperscalers had been exploiting — pairing on-site generation with data centers to avoid the formal interconnection queue. Source: FERC fact sheet.
2026 — Current
ERCOT revises 2030 data-center demand: 29 → 77 GW
The single-planning-cycle revision is the largest forecast adjustment in ERCOT's history. The Texas grid operator's queue holds ~156 GW of proposed data-center capacity by 2030. PJM's parallel queue carries ~220 GW. Both grids have approval throughput in low single-digit percentages of queued volume.
SECTION 2 — THE NUMBERS
PJM, ERCOT, and the National Picture
Region
2025 Peak (GW)
Queue (proposed GW)
2030 DC demand (GW)
Avg queue→energization
PJM Interconnection
154
~220
~210 (projected 2036 peak)
~7 years
ERCOT (Texas)
~85
~156
77 (was 29 in prior cycle)
~5 years
MISO (Midwest)
~127
~180 (renewables-heavy)
~15-20
~6 years
CAISO (California)
~52
~100 (mostly solar+storage)
~10
~4 years
US Total
~760 (national peak)
~2,600 (all-tech queue)
~150 (national DC load by 2028)
3–7 years
The queue numbers above carry the historical shadow-ratio caveat: only 15–25% of queued projects energize. Applied uniformly, the "effective" PJM data-center queue is closer to 33–55 GW than 220 GW. But hyperscaler-backed queue applications have a higher conversion rate than the historical average — these are not speculative renewables developers filing to flip permits.
Supply backlogs that bind
Even if every queued data-center project had unlimited capital, three physical-supply constraints would still cap the rate of energization:
1. Heavy-duty gas turbines. GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries collectively produce ~80 large-frame gas turbines per year. 2027 delivery slots are largely sold; 2028 is the first cycle with material availability. Backlog clearance time: 24–36 months from order.
2. Large-power transformers (LPTs). Global LPT manufacturing capacity (Hitachi Energy, Siemens, GE, Hyundai) is tight. 765 kV and 500 kV class transformers carry lead times of 30–60 months. The US imports ~80% of its LPTs; tariff and supply-chain risk is non-trivial.
3. Skilled trades. Lineworker, substation technician, and PE workforce contracted during 2007–2019 flat-load period. Re-staffing to 2025-2030 build requirements takes a full apprenticeship cycle (4 years minimum). This is the constraint nobody talks about loudly.
Data Center Frontier — May 2026, "The Gigawatt Bottleneck"
"The biggest obstacle to deploying AI infrastructure is no longer capital, land, or connectivity — it's electricity. In major markets from Northern Virginia to Texas, grid interconnection timelines are stretching out for years as utilities struggle to keep pace with a surge in large-load requests from AI-driven infrastructure."
SECTION 3 — COMPETING HYPOTHESES
Three Scenarios For The 2026-2029 Window
The dossier's load-bearing claim is "Grid energization constrains frontier AI training compute deployment through 2028." Three competing hypotheses; observables; diagnosticity scored 1 (consistent) to 5 (decisive against).
Stargate / Colossus / Anthropic announced GW targets
2 — announced ≠ energized
1 — BTM is exactly the model
3 — requires public-grid energization
Historical queue shadow ratio 15-25%
2 — depresses effective constraint
2 — consistent
2 — consistent
LPT lead time 30-60 months
1 — binds public-grid pathway
3 — partially binds BTM as well
5 — precludes 2028 dissolution decisively
Weighted total
13 (least disconfirmed)
12 (most consistent)
23 (most disconfirmed)
Reading: H2 (bottleneck moves) is the most-consistent hypothesis. H1 (binds) is close behind. H3 (dissolves) is essentially refuted by the LPT and turbine supply-side constraint. The headline forecast favors H2 — the bottleneck does not so much bind as redirect: AI compute deployment continues via BTM and nuclear restart, the public grid queue stays constipated, and the apparent "queue size" overstates the effective constraint on frontier compute.
SECTION 4 — ALLIED & ADVERSARIAL PERSPECTIVES
How Other Grids Are Reading The Same Problem
The US grid is unique in its degree of state-by-state regulatory fragmentation and queue-based interconnection process. Other major economies have different constraints and different policy levers — and the gaps between them are themselves strategic information.
UKNational Grid ESO — System Operator
UK National Grid ESO has published its own AI data center load forecast for 2024-2035, flagging similar magnitude growth proportional to grid size. The UK constraint is different: smaller grid, fewer interconnection bottlenecks, but tighter generation reserve margins after coal retirements. Three Mile Island-style restart equivalents (Heysham B, Hartlepool extensions) are being assessed. Strategic implication: UK AI infrastructure may face earlier reserve-margin constraints than US queue constraints.
EUENTSO-E + national TSOs (TenneT, Amprion, RTE, Terna)
European TSOs face the parallel constraint of grid-connection queues, with the Netherlands (TenneT) and Germany (Amprion) refusing new large-load connections in specific zones through 2027-2028. The "data center moratorium" model is more politically acceptable in EU jurisdictions than in the US. Strategic implication: EU AI infrastructure deployment may concentrate in Nordic and Iberian zones with grid headroom (Sweden, Finland, Norway, Portugal, Spain).
ChinaState Grid Corp of China — SGCC
SGCC operates the world's largest grid by far (~1,100 GW peak). Chinese transmission build-out cycles run 18-30 months end-to-end versus 5-7 years in the US, due to centralized planning, integrated supply chains (domestic LPT manufacture), and reduced permitting friction. China's AI compute clusters in Inner Mongolia (Hohhot, Ulanqab) and Ningxia exploit grid headroom unavailable in US coastal hyperscaler markets. The strategic implication for the US-China AI race is the lopsided cost: every month of US grid delay is a month of Chinese build advantage. This is the structural concern under the Stargate timeline.
IndiaCEA — Central Electricity Authority + State DISCOMs
India's grid faces a different problem: rapid demand growth (~6-7% CAGR) but emerging AI data-center load on top of base growth. CEA's Eighth National Power Plan (2024-2032) projects ~10 GW of dedicated data-center load by 2030. Tata, Reliance, Adani are building AI compute clusters around solar+battery base-load, partly to avoid grid contracts. State DISCOMs (e.g., MSEDCL in Maharashtra) are central — coastal Mumbai/Pune corridor is the AI hot zone.
UAEEWEC — Emirates Water and Electricity Company
UAE Stargate site is announced for 2026 deployment. EWEC has grid headroom from Barakah nuclear (5.6 GW) and rapidly-built solar capacity. Generation cost in UAE for hyperscaler offtake is materially lower than US delivered cost. Strategic implication: the AI-grid bottleneck is a US-and-EU constraint; sovereign-wealth-backed gulf jurisdictions (UAE, Saudi Arabia) are unconstrained on the same axis.
Nordic countries have abundant cold-climate hydro and grid headroom. Microsoft, Meta, Amazon AWS, Google, and Equinix have all expanded Stockholm, Oslo, Helsinki, Reykjavik data centers. The constraint here is land use and political acceptance, not grid capacity. Iceland alone exports geothermal-backed compute capacity to global hyperscalers.
SECTORS — WHO GAINS, WHO LOSES IF THE BOTTLENECK BINDS
↑ WINNERS
GE Vernova / Siemens Energy
Gas turbine orderbook full through 2027. Both stocks (GEV, ENR.DE) outperforming sector. 24-36 month backlog is pricing power.
Constellation Energy
Three Mile Island restart PPA with Microsoft. Owns the most attractive nuclear restart-eligible fleet in the US.
Hyperscalers with BTM capability
Microsoft, Meta, Google, AWS can buy or build generation directly. xAI's Memphis turbines are the model. Effectively bypass queue.
Lower interconnection friction than PJM. Data centers shifting to Texas. Property tax base expansion, construction jobs.
UAE / Saudi Arabia / Iceland
Sovereign grid headroom + offtake-friendly regulators. Stargate UAE, Saudi NEOM, Icelandic geothermal compute.
↓ LOSERS
Smaller AI labs without BTM capital
SSI ($32B, no shipped product), Mistral, Cohere, scaled startups. Cannot self-finance generation. Public queue too slow.
Northern Virginia residents
Loudoun County electricity prices rising; transmission build-out concentrated impacts. Local moratoria proliferating.
PJM ratepayers
Cost-allocation for data-center-driven transmission upgrades is unresolved. Residential bills rising; potential cross-subsidy from households to hyperscalers.
US renewable interconnection queue
Solar+storage projects competing for queue slots with data-center load. Renewables development slowed by hyperscaler queue saturation.
US AI competitive position vs China
SGCC's 18-30 month transmission cycle vs US 5-7 years means every PJM/ERCOT delay is a Chinese compute advantage on the geopolitical clock.
EU industrial competitiveness
Even tighter grid constraints than US in core industrial zones; AI infrastructure shifts to Nordic periphery, away from German/French industrial base.
Dig Wall
Six things this dossier cannot resolve from public data — and what would resolve them.
The dossier's headline claim resolves on quarterly PJM/ERCOT planning reports. But six secondary questions require regulator filings, manufacturer orderbook data, and Federal Register publications the public discourse hasn't yet surfaced.
1
What is the actual hyperscaler-vs-renewables breakdown of the 220 GW PJM queue? PJM publishes queue volume but not always type-decomposed. Distinguishing data-center load from renewables generation requests is the diagnostic for "binding constraint" vs "speculative queue inflation." Monitor: PJM Queue Manager filings, FERC Form 715. Cross-ref: [[entity:PJM Interconnection]]
2
What are GE Vernova / Siemens Energy 2027-2029 turbine delivery slot allocations? Quarterly orderbook updates from earnings calls. Tracking which hyperscalers have priority slots reveals the structural BTM strategy. Monitor: GEV quarterly earnings, Siemens Energy investor relations. Cross-ref: [[entity:GE Vernova]][[claim:gas-turbine-backlog-binds]]
3
How many additional nuclear restart PPAs have been signed since Microsoft-Constellation TMI? Constellation Energy's 10-Q filings, Holtec announcements (Palisades), Energy Harbor / Vistra public statements. Each additional restart deal materially shifts the H2 vs H1 hypothesis odds. Monitor: SEC EDGAR utility filings. Cross-ref: [[entity:Constellation Energy]]
4
Which states have data-center-specific tariff classes or moratoria proposed in 2025-2026? Virginia HB1601, Georgia Power IRP, Ohio HB15, Indiana HB1007. State-by-state policy heterogeneity is the structural lever for the H3 (dissolves) hypothesis. Monitor: state PUC dockets. Cross-ref: [[entity:Virginia]][[entity:Texas]]
5
What is the LPT (Large Power Transformer) global production capacity vs US demand? Hitachi Energy, Siemens Energy, GE, Hyundai, Toshiba global capacity is roughly 200-300 units/year of 500-765kV class. US demand growing rapidly. Tariff and supply-chain risk needs assessment. Monitor: Hitachi Energy quarterly results, DOE Office of Electricity transformer reports. Cross-ref: [[claim:lpt-supply-binds]]
6
What is the SGCC 2025-2030 AI-driven grid build cadence vs PJM/ERCOT? Chinese transmission build-out cadence is the strategic comparator. Direct comparison: GW/yr of high-voltage transmission energized, days from approval to commercial operation. Available data: China Electricity Council annual reports, SGCC English-language summary releases. The gap is the AI-race headline metric the press isn't yet covering. Cross-ref: [[entity:State Grid Corp of China]]
Sources
1
Federal Energy Regulatory Commission — Fact Sheet: PJM Co-Located Large Load Directive T1 primary — ferc.gov
2
FERC — PJM Innovation and Consumer Protection Order T1 primary — ferc.gov
3
Data Center Frontier — "The Gigawatt Bottleneck: Power Constraints Define AI Data Center Growth," May 2026 T2 — datacenterfrontier.com
4
Data Center Knowledge — "Why AI Data Center Projects Face Years of Delays After Approval," 2026 T2 — datacenterknowledge.com
5
Bismarck Analysis Brief — "AI 2026: Data Centers Restart Growth of a Stagnant US Electrical Grid" T2 — brief.bismarckanalysis.com
6
EnkiAI — "US Grid Congestion 2026: PJM's Data Center Crisis" T2 — enkiai.com
Falsifiable Watchlist claim-level kill conditions, resolved on public datasets
Each load-bearing claim carries: (a) a calibrated probability, (b) a resolution date, (c) the specific public dataset where the answer will appear. This is the falsifiability tax — what makes the dossier a Bayesian instrument rather than a position paper.
PJM queue approval throughput remains below 5%/yr through 2027.
P = 0.78 · by 2028-01-31 · As of 2026-05-14
Falsified If
PJM's 2027 annual queue report shows approval throughput ≥5% of queued volume, OR cumulative 2026-27 GW energized exceeds 25 GW.
Resolves On
PJM annual queue reports, January 2028 publication.
ERCOT 2030 data-center demand forecast remains at 60+ GW in 2027 cycle (does not retreat to under-50 GW range).
P = 0.72 · by 2027-12-31 · As of 2026-05-14
Falsified If
ERCOT's 2027 CDR report revises 2030 DC demand below 50 GW, OR data-center deployment in Texas slows to under 5 GW added/year through 2027.
Resolves On
ERCOT Capacity, Demand, and Reserves (CDR) Report, December 2027.
At least 3 additional nuclear restart PPAs (beyond TMI 1) signed with hyperscalers by end-2027.
P = 0.62 · by 2027-12-31 · As of 2026-05-14
Falsified If
No additional restart PPAs publicly signed by end-2027, OR existing restart sites (Palisades, Duane Arnold) cancel feasibility studies.
Resolves On
SEC EDGAR 10-Q filings (Constellation, Vistra, Energy Harbor) + NRC filings.
GE Vernova and Siemens Energy heavy-duty turbine 2027 delivery slots remain fully booked through end-2026.
P = 0.85 · by 2026-12-31 · As of 2026-05-14
Falsified If
Either manufacturer's Q4 2026 earnings disclose 2027 delivery slots still available, OR new heavy-duty turbine production facility ramp announced for 2026-27.
Resolves On
GE Vernova Q4 2026 earnings call + Siemens Energy AGM 2026.
At least one frontier AI lab energizes ≥2 GW of dedicated compute via BTM generation by end-2027.
P = 0.70 · by 2027-12-31 · As of 2026-05-14
Falsified If
No frontier lab announces ≥2 GW BTM energization by end-2027, OR Stargate Abilene + Colossus 2 + Anthropic-Google TPU sum to less than 4 GW operational by Jan 2028.
PJM region residential electricity rates rise >15% cumulative 2025-2027, attributable to data-center transmission upgrades.
P = 0.66 · by 2028-01-31 · As of 2026-05-14
Falsified If
EIA residential rate data for PJM-region states (PA, NJ, VA, MD, OH, DE) shows aggregate increase <10% cumulative 2025-2027, OR state PUCs successfully ring-fence data-center cost-allocation away from residential ratepayers.
Resolves On
EIA Electric Power Monthly + state PUC rate orders.
State Grid Corp of China energizes >100 GW of new transmission capacity 2025-2027, sustaining the ~30-month build-cycle advantage.
P = 0.80 · by 2028-06-30 · As of 2026-05-14
Falsified If
SGCC 2025-2027 annual reports show transmission additions <70 GW cumulative, OR China Electricity Council reports project completion delays exceeding 24 months on flagship UHV lines.
Resolves On
SGCC annual reports + China Electricity Council statistical bulletins.