Powering the AI Boom: Where the Grid Breaks First (2026-2030)
AI data centers will drive US electricity demand to 550-650 TWh by 2030, equivalent to 10-12% of total consumption, with forecasts varying widely from IEA's 425 TWh to BCG's 970 TWh due to differing assumptions on efficiency and growth. Grid constraints in key regions emerge as the first breaking point, risking supply shortages amid this surge.
- 01 The Kobeissi Letter highlights that AI data centers will quadruple power demand to 1,600 TWh by 2035 (4.4% of global electricity), creating a major energy shortage as compute scales without matching supply
- 02 Oxford energy professor Jan Rosenow argues AI and data centers contribute only 8% to rising electricity demand through 2030 per IEA, with most growth from transport, buildings, and industry electrification
- 03 OptionsPlay forecasts data center electricity surging from 448 TWh in 2025 to 980 TWh by 2030 (7.8% of US total), positioning power generation and grid as underowned AI picks alongside semis
- 04 Market analyst Jon Erlichman charts US data center electricity rising from 4.3% of total power in 2024 to 11.7% by 2030 per McKinsey, underscoring grid strain acceleration
- 05 Energy economist Tracy Shuchart warns PJM grid (US AI hub) faces data center demand outpacing new capacity through 2030, spiking auction prices amid calls for reforms
1. The Demand Picture: 550–650 TWh by 2030, and Why Forecasters Disagree by 2x
The range of published forecasts is extraordinary: from IEA's ~425 TWh (8% of US total) to BCG's ~970 TWh (19%), a spread so wide it's functionally useless for infrastructure planning. The disagreement is not noise—it reflects genuinely different modeling philosophies colliding with a genuinely unprecedented demand driver.
Three methodological camps explain the spread. Pipeline trackers (EPRI, Grid Strategies) count announced, under-construction, and operational data center sites, then apply realization rates. EPRI's February 2026 update yielded 384/596/793 TWh across low/medium/high scenarios depending on what share of early-planning projects materialize (Report 1). Equipment modelers (LBNL) count server and GPU shipments, apply utilization rates (60–80% of rated power), and simulate PUE thermodynamically—yielding 325–580 TWh by 2028 alone, trending to 500–700 TWh by 2030 (Report 1). Workload extrapolators (Goldman Sachs, BCG) model AI adoption curves—inference pervasiveness, agentic automation—and project backwards to power, producing the highest estimates (~810 TWh for Goldman's February 2026 revision assuming 60% US share of global growth) (Report 1).
The forecasts diverge on four specific assumptions. First, training vs. inference mix: LBNL models training consuming 50–53% of AI energy through 2028, while Goldman assumes inference intensity rises faster, adding 20–30% to the high end (Report 1). Second, GPU efficiency curves: LBNL assumes PUE declines to ~1.4 via liquid cooling and hyperscale shift; McKinsey pushes to PUE 1.1 (Report 1). Third, realization rates: Grid Strategies flags that utilities assume near-100% load factors without supply bottlenecks, while analysts cap at 50–96%—EPRI's medium case assumes roughly 65% of advanced planning materializes (Report 1). Fourth, hyperscaler vs. enterprise mix: LBNL projects 85% of servers in hyperscale/colocation by 2028, which means lower PUE and higher utilization than if enterprise data centers retained more share (Report 1).
Defensible base case: 550–650 TWh (11–13% of ~5,000 TWh total US generation). This blends EPRI's medium scenario (596 TWh), LBNL's trend-adjusted high, and Goldman's US-share estimate, assuming hyperscaler dominance (85% of servers), PUE of 1.2–1.4, and a roughly even training/inference split. The key sensitivities: +20% if inference explodes beyond Goldman's assumptions; -15% if GPU efficiency and PUE improvements accelerate beyond 1.2; ±10% based on pipeline realization (Report 1). This base case represents roughly tripling current data center electricity consumption from a 2024 baseline of 177–192 TWh (Report 1).
The critical insight the forecasts collectively reveal: even the low end (380–425 TWh) implies adding the equivalent of 40–50 large power plants' worth of baseload demand in five years—a buildout pace the US grid has not achieved since the 1960s.
2. The Five Bottlenecks, Ranked by Which Binds First
The bottlenecks do not strike simultaneously. They cascade in a specific sequence, and the binding constraint shifts from procedural (queues) to physical (equipment) to structural (generation/transmission) over the 2026–2030 window.
1. Interconnection Queues — Binding Now (2024–2026)
This is the first wall. ERCOT's large-load queue exploded to 410 GW by April 2026—87% data centers—exceeding peak demand by 4.8x (Report 2). PJM projects entering the queue now face 2028+ interconnection agreements, with median queue-to-operation at 55 months nationally and PJM projects operational in 2025 averaging 8 years in queue (Report 2). The evidence is already visible in outcomes: 50% of planned 2026 US data centers (12 GW) have been delayed or cancelled, with only 5 GW under construction (Report 2). FERC Order 2023 shifted RTOs from serial to cluster-study processing, but the transition itself created a "pause" in PJM through 2026, choking near-term throughput (Report 2). This bottleneck is procedural—it doesn't reflect physical impossibility, but it imposes 4–8 year delays that exceed data center construction timelines of 18–24 months (Report 2).
2. Equipment Supply Chain — Binding 2026–2028
Large power transformers (LPTs >100 MVA) now carry lead times of 120–210 weeks (2.5–4+ years), with prices up 79% (Report 2). GE Vernova's Q1 2026 data center orders equaled all of 2025, with a total backlog of $163 billion; Siemens and Hitachi are building new US plants but those won't come online until 2027–2028 (Report 2). Gas turbine slots are tight through 2029–2030 with 3-year leads (Report 4). The transformer shortage is the most underappreciated constraint because it is physical and global—even projects that clear queues cannot energize without step-up transformers, and half of US distribution transformers are past design life (Report 2). China imports surged 433% as a stopgap, introducing tariff risk (Report 2).
3. Generation Adequacy — Binding 2025–2027 in PJM, 2027+ Elsewhere
PJM's capacity auctions have already priced in scarcity: the 2027/2028 Base Residual Auction cleared at the FERC cap of $333/MW-day but still fell 6,517 MW short of the reliability requirement—the first shortfall exceeding 1 percentage point in history (Report 3). Reserve margins hit 14.4% versus a 20% target (Report 3). Data centers caused 45% ($21.3B) of $47.2B in cleared capacity costs across three auctions (Report 3). MISO's summer 2025/26 capacity price spiked to $666/MW-day from $30 the prior year (Report 2). ERCOT has no formal capacity market, but reserves drop below the 13.8% reference margin post-2028 (Report 3). The generation shortfall is not hypothetical—it is priced into forward markets today.
4. Transmission Capacity — Binding 2027+ but Building Slowly
FERC Order 1920 mandates 20-year scenario-based planning, but compliance filings only began in December 2025 (PJM/CAISO first), with MISO/SPP due June 2026 (Report 2). First projects from the new planning process won't emerge until 2028 at earliest, with buildout taking 10–15 years (Report 2). Only 55 miles of new high-voltage lines were built in 2023; congestion costs in PJM rose 64–78% to $1.7–7.3 billion (Reports 2, 3). PJM approved a $6.7B 765kV backbone for Virginia data centers in February 2025, but it is years from completion (Report 3). Transmission is the bottleneck that converts short-term queue problems into long-term structural constraints.
5. Natural Gas Pipeline and Water/Cooling — Binding Regionally, 2027+
These are secondary but increasingly acute in specific geographies. ERCOT has 58 GW of gas planning (half for data centers), straining distribution pipelines (Report 2). In Arizona, a proposed data center requiring 620 million gallons/year was rejected over water concerns (Report 2). Behind-the-meter gas turbines strain local pipeline distribution even when national production is ample (Report 2). Community opposition has delayed 20+ projects in 2025 (Report 2). These constraints are mitigable via dry cooling and pipeline expansion, but add 16–20 months of permitting in affected regions (Report 2).
3. Regional Risk Map: PJM Is Already in Crisis, ERCOT Is a Pressure Cooker, Everyone Else Is on a Fuse
The thesis that "AI power demand" is too coarse a frame is proven most clearly in regional comparison. PJM is experiencing a capacity market failure right now. ERCOT faces a speculative queue bubble with weather tail risk. MISO and SPP are slowly eroding toward negative reserves. CAISO and NYISO face localized transmission chokes. The binding constraint, timing, and severity differ meaningfully.
PJM — Severity: Critical / Binding Now
Northern Virginia's "Data Center Alley" (Loudoun County/Dominion zone) concentrates the crisis. Data centers drove 97% of 5,250 MW load growth in PJM's forecast; the Dominion zone alone has 30+ GW of data center demand queued through 2030 (Report 3). Capacity prices surged 833% to $270/MW-day, then hit the FERC cap of $333/MW-day, and the auction still came up 6,517 MW short (Report 3). Under-construction data center capacity fell 29% in Northern Virginia in 2025 (Report 3). PJM is proposing a 14.9 GW backstop procurement mechanism, with FERC eyeing a June 2026 decision (Report 3). The $6.7B 765kV transmission backbone is approved but years from completion. PJM is the canary—it shows what happens when a single region absorbs disproportionate data center load before the grid can respond. Severity escalates through 2030 without relief.
ERCOT — Severity: High / Binding 2028+
Texas has the largest absolute queue—410 GW of large-load requests, 87% data centers—but only 1.8% are energized (Report 3). The queue is overwhelmingly speculative: PUCT rejected inflated forecasts and imposed SB6 viability rules requiring financial commitments for 75+ MW loads (Report 3). ERCOT's advantage is fast-build natural gas: the Texas Energy Fund supports 10 GW of new gas, and behind-the-meter deployments (xAI, Crusoe, Oracle) can come online in 1–2 years versus 4+ years on grid (Report 4). But ERCOT's weakness is weather: reserves drop below 13.8% post-2028, and the 2021 freeze demonstrated catastrophic tail risk (Report 3). The peak forecast triples to 278 GW by 2029 (Report 3). ERCOT's constraint profile is fundamentally different from PJM's: it is a pipeline of speculative demand meeting a market that builds fast but cannot guarantee reliability in extremes.
MISO — Severity: Moderate, Rising / Binding 2027–2029
MISO's reserve margins are declining steadily: 7.9% for summer 2025/26, dropping to 4.3% by 2029, driven by coal/gas retirements outpacing replacement (Report 3). Data center growth is slower than PJM/ERCOT—20% of electricity by 2030—but the margin erosion is real (Report 3). The summer 2025/26 capacity auction spiked to $666/MW-day (Report 3). MISO's 242 GW queue (DPP-2025 cluster with 78 GW) signals multi-year waits (Report 2). The risk is less acute than PJM but more structural: MISO lacks the scarcity pricing mechanisms that force rapid supply response, and its queue reforms delayed processing.
SPP/Southeast — Severity: Moderate, Rising / Binding 2028–2030
SPP's reserve margins plummet from 20.7% (2025) to -1.6% by 2030 (Report 3). Duke territory sees 6.1% CAGR in industrial/data center demand, with 38 advanced projects totaling 5,610 MW peak (80% data centers) (Report 3). The $19.4B SPP transmission portfolio (including 765kV overlay) addresses congestion but won't be built quickly (Report 3). TVA and Duke operate in regulated territory with no capacity market safety net—bilaterals rule, and Duke's contracts mandate 50-hour/year curtailment for faster hookup (Report 3). Southeast utilities plan 10 GW of gas for speculative data center growth with only 0.2% probability of full realization (Report 3), creating overbuild risk.
NYISO — Severity: High Locally / Binding Summer 2025–2026
Downstate New York (NYC/Long Island) faces localized capacity requirements deficits from peaker retirements (1.5 GW ozone-season) and Central East transmission bottlenecks blocking upstate hydro/nuclear (Report 3). The Champlain Hudson Power Express (CHPE, 1.25 GW) resolves the immediate summer 2026 downstate shortfall, but 19 large loads (3+ GW including data centers) are queueing amid a proposed three-year moratorium (Report 3). NYC's MLCR is 79.2%, among the tightest in the country (Report 3). NYISO is a transmission-constrained island problem—upstate has power, downstate can't get it.
CAISO — Severity: Low-Moderate / Binding 2028+
California's data center load is smaller (1.8 GW by 2030), but Path 15 north-south congestion is binding, and Bay Area/Silicon Valley overloads are emerging (Report 3). CAISO's $7B plan for 38 transmission upgrades addresses the medium-term, including a $1.4B Silicon Valley project for large loads (Report 3). Gas peaker retirements (3.7 GW once-through cooling, 11 GW by 2034) create local reliability risk (Report 3). CAISO is the least acute for data centers specifically, but its transmission constraints compound if load grows beyond current forecasts.
The view: PJM is the only region where the grid is already failing to clear sufficient capacity for data center load. ERCOT is the region most likely to experience a reliability event because its fast-build advantage masks weather tail risk. MISO and SPP are slow-burning fuses that become critical around 2028–2029. NYISO and CAISO are locally constrained but nationally secondary.
4. Supply Response: What Actually Gets Built, and When
The supply-side response is a mix of the plausible, the aspirational, and the physics-constrained. Only natural gas and battery storage deliver material capacity before 2028. Nuclear restarts add meaningful power in 2027–2028. SMRs are a 2030+ story no matter what anyone announces.
Nuclear Restarts (2027–2029): ~2.2 GW, High Execution Risk
Holtec's Palisades (800 MW) is furthest along, with NRC approval for fuel loading and a restart target of early 2026, though it has slipped repeatedly due to steam generator upgrades (Report 4). No hyperscaler PPA is announced; it is financed by a $1.52B DOE loan (Report 4). Constellation's Crane/TMI-1 (835 MW) targets H2 2027, backed by Microsoft's 20-year PPA and a $1B DOE loan, but PJM transmission delays may prevent full grid deliverability until 2031 (Report 4). NextEra's Duane Arnold (615 MW) targets early 2029 via Google's 25-year PPA, with NRC licensing expected by January 2028 (Report 4). Net: ~800 MW possible in early 2026–2027 (Palisades), ~1.6 GW cumulative by 2028+ if Crane and Duane Arnold execute—but grid interconnection, not reactor readiness, is the binding delay.
SMRs and Microreactors (2028+ Pilots, 2030+ at Scale): Negligible Before 2030
Oklo's Aurora targets late 2027/early 2028 for a DOE pilot at Idaho National Lab, with Meta's 1.2 GW Ohio campus targeting first units by ~2030 (Report 4). NuScale has NRC design certification but no binding US customer or construction timeline (Report 4). Kairos broke ground on Hermes 2 (50 MW) in April 2026, targeting operations in 2030, but Hermes 1 construction was extended to 2029 (Report 4). Google's 500 MW fleet agreement targets 2030–2035 (Report 5). Net: zero GW before 2028, sub-500 MW of pilots by 2029, GW-scale not before early 2030s. HALEU fuel supply (US produces ~900 kg/year) is a hard physical constraint (Report 4).
Natural Gas CC/Peakers (2026–2028): Fastest Baseload Response
EIA projects 3.3 GW combined cycle online in 2026, another 3.3 GW in 2027, and 10.6 GW in 2028 (Report 4). Behind-the-meter gas dominates the near-term buildout—xAI's 422 MW Memphis facility, Google/Crusoe's 933 MW Texas project, Meta's 2 GW Louisiana combined-cycle (Report 4). Total gas development pipeline is ~252 GW, but turbine backlogs (5+ year delivery for new orders) limit throughput to pre-ordered equipment (Report 4). Net: ~3–7 GW grid-tied in 2026, 6–10 GW in 2027, 15+ GW in 2028+, plus substantial BTM additions. Gas is the only technology that can meaningfully close the generation gap before 2028.
Battery Storage (2026–2028): Fast Deployment, Duration-Limited
EIA forecasts 24 GW of utility-scale storage in 2026, rising to 67 GW cumulative by Q1 2027, led by Texas and California (Report 4). Behind-the-meter storage for data centers is surging, with 15 GW projected by 2030 (83% for data centers) (Report 4). But 4-hour duration limits baseload applicability for AI workloads—effective capacity (ELCC) is roughly 50% of nameplate (Report 4). Net: 20–25 GW nameplate in 2026, but only ~10–12 GW effective. Batteries are essential for peak shaving and renewable firming but cannot replace baseload generation for 24/7 AI load.
Behind-the-Meter Fuel Cells (2026–2028): Oracle's Bridge
Bloom Energy's solid-oxide fuel cells deploy in 55–90 days versus 5–7 years for grid interconnection (Report 5). Oracle's 2.8 GW Bloom deal (1.2 GW deploying) and AEP's $2.65B/1 GW agreement represent GW-scale committed capacity (Report 4). Bloom is ramping to 2 GW/year production capacity (Report 4). Net: 1–2 GW in 2026, 2–4 GW in 2027, scaling further in 2028+. Fuel cells are the fastest path for hyperscalers who need dispatchable on-site power now, at $150–180/MWh (Report 5).
5. Hyperscaler Self-Help: The Grid Gets Bypassed, Not Fixed
The most important structural shift in this entire analysis is the divergence between power demand that flows through the regulated grid and power demand that hyperscalers meet themselves. The grid bottleneck is accelerating a historic decoupling.
Bloom Energy forecasts that 38% of data centers will use on-site generation by 2030 (up from 13% in 2025), with 33% operating fully off-grid—a 27x increase (Report 5). RAND estimates 49 GW of behind-the-meter net capacity by 2030 versus 33 GW front-of-meter, meaning ~60% of net new data center capacity could bypass the grid entirely (Report 5). Goldman Sachs models grid demand CAGR of 2.8% with an additional 0.5% from BTM, indicating that behind-the-meter growth is already a significant fraction of the total (Report 5).
The hyperscaler strategies are converging on a common playbook:
Microsoft anchored the nuclear restart model: 835 MW Crane PPA (20-year, grid-delivered through PJM) plus investments in SMR development (Report 5). Meta assembled the largest corporate nuclear portfolio: 6.6 GW across Vistra existing plants (2.1 GW PPAs), uprates (433 MW), TerraPower Natrium SMRs (690 MW), Oklo Aurora (1.2 GW), and Constellation Clinton (1.1 GW)—bridged by 2 GW of on-site gas in Louisiana (Report 5). AWS pivoted from FERC-rejected behind-the-meter co-location at Susquehanna to a front-of-meter 1.92 GW PPA with Talen, routing through PJM grid (Report 5). Google pursued both co-located clean generation (AES partnership in Texas) and the Kairos SMR fleet (500 MW by 2035), while deploying 1 GW of demand response capacity (Report 5, Report 6). Oracle bet most aggressively on behind-the-meter: 2.8 GW Bloom fuel cells, 2.3 GW VoltaGrid modular gas, and the 2.45 GW Project Jupiter microgrid in New Mexico—effectively building a private utility (Report 5).
The implication: the base case demand of 550–650 TWh does not all hit the grid. If 25–38% of incremental data center load goes behind-the-meter (Report 5), the grid-served increment drops to roughly 350–480 TWh—still a massive buildout, but materially less than headline figures suggest. This changes the bottleneck calculus: interconnection queues matter less for hyperscalers with BTM strategies, but matter more for smaller operators who cannot afford private power infrastructure. The grid becomes a two-tier system: hyperscalers with self-help, and everyone else stuck in queue.
6. Value Chain Positioning: Who Benefits from Each Bottleneck
The five bottlenecks create five distinct value pools. Structural positioning advantages flow to companies that sit at the tightest choke points.
Interconnection queue congestion → Existing generators with capacity in constrained zones. Constellation (Crane/TMI-1, Byron, Braidwood, Calvert Cliffs adjacency), Vistra (Perry, Davis-Besse, Beaver Valley, Comanche Peak), and Talen (Susquehanna) hold the scarcest asset: already-interconnected dispatchable generation in PJM and ERCOT where new capacity cannot enter for 5–7 years (Reports 3, 5). PJM capacity prices at $333/MW-day (10x 2023 levels) flow directly to these incumbents (Report 3). Nuclear plant owners benefit disproportionately because their capacity factors (~92%) and carbon-free attributes command premium PPAs—Microsoft's Crane deal and Meta's Vistra agreements demonstrate willingness to pay above-market rates for firm clean power (Report 5).
Equipment supply chain → Transformer and switchgear manufacturers, electrical infrastructure. GE Vernova's $163B backlog and Prolec transformer subsidiary sit at the physical chokepoint where 50%+ of 2026 data centers are delayed (Report 2). Siemens Energy and Hitachi Energy are expanding US plants but output doesn't arrive until 2027–2028 (Report 2). Eaton and Hubbell supply switchgear and power distribution, where lead times have extended alongside transformers. Quanta Services and MYR Group build the substations and transmission lines that every project requires (Report 2). GE Vernova's gas turbine order book is tight through 2029–2030 (Report 4).
Generation adequacy → Dispatchable capacity developers and nuclear operators. NextEra (Duane Arnold restart, plus renewable/storage pipeline), Constellation (operating fleet plus Crane restart), and Vistra (gas/nuclear fleet plus Meta uprates) are positioned on both sides of the generation gap—they own existing capacity earning scarcity rents and are developing new capacity that commands premium contracts (Reports 4, 5). In ERCOT, fast-build gas developers with pre-ordered turbines can capture the 2028+ reserve deficit.
Nuclear fuel and SMR development → Upstream fuel suppliers. Cameco (uranium), Centrus (HALEU enrichment—US produces only ~900 kg/year against surging SMR demand), and BWXT (naval/commercial reactor components) occupy structural positions in a supply chain that cannot scale quickly (Report 4). Oklo and NuScale/NNE are pre-revenue technology developers whose value depends entirely on execution against timelines that have historically slipped.
Gas pipeline infrastructure → Midstream operators near data center hubs. Williams ($5B pipeline expansion), Energy Transfer (Oracle/Texas pipeline integration), and Kinder Morgan ($9.1B backlog) benefit from behind-the-meter gas demand that strains local distribution even when national supply is ample (Reports 2, 5). The constraint is not gas production—it is last-mile pipeline capacity to data center campuses.
Behind-the-meter power → Bloom Energy and modular gas providers. Bloom's position is structurally unique: 55–90 day deployment versus 5–7 year grid timelines, with $7.65B+ in hyperscaler commitments (Oracle 2.8 GW, AEP 1 GW, Brookfield $5B) (Reports 4, 5). No other technology delivers dispatchable, permitted, on-site power at this speed. The risk is that Bloom's fuel cell economics ($150–180/MWh) become uncompetitive if gas prices spike or nuclear restarts deliver cheaper baseload.
7. What Breaks First: The Constraint Timeline, 2026–2030
The sequence matters because each constraint triggers adaptation that shifts pressure to the next bottleneck. This is the cascading failure model, not a single-point crisis.
2026 Q1–Q2: Generation adequacy alarm bells
PJM's 2027/2028 auction results (December 2025) revealed a 6,517 MW shortfall—already public and already repriced (Report 3). PJM proposes a 14.9 GW backstop mechanism; FERC targets June 2026 decision (Report 3). MISO's $666/MW-day capacity price spike signals parallel tightening (Report 2). Palisades restart likely slips again into mid-2026 (Report 4). Half of planned 2026 US data center capacity (12 GW) is delayed or cancelled due to equipment and queue constraints (Report 2).
2026 Q3–Q4: Transformer and equipment shortage peaks
GE Vernova's Q1 2026 data center orders equaling all 2025 signals a 2026–2027 delivery crunch (Report 2). Projects that cleared interconnection studies in 2024–2025 cannot energize because transformers ordered in 2023–2024 haven't arrived. This is the period when the gap between "approved" and "operational" data centers becomes most visible. Bloom Energy ramps toward 2 GW/year production; Oracle's first Project Jupiter fuel cells come online as the BTM alternative (Report 4).
2027 H1: First nuclear restart—and first hyperscaler energization failures
Palisades (~800 MW) is the most likely first restart, contributing to Michigan/MISO grid but without hyperscaler PPA (Report 4). Crane/TMI-1 targets H2 2027, but PJM transmission delays may limit initial output or force a FERC waiver (Report 4). This is the quarter when hyperscalers who committed to 2027 campus openings in PJM discover that neither grid capacity nor equipment is available—the first public acknowledgments that major campuses cannot energize on schedule. BTM gas and fuel cells absorb the overflow.
2027 H2–2028: Transmission constraint crystallizes
FERC Order 1920 compliance filings are in, but first projects are years from construction. PJM's $6.7B 765kV backbone is in permitting. The realization hits that interconnection reform (queues) and generation buildout (gas/nuclear) solve nothing if transmission cannot deliver power from where it's generated to where data centers are sited (Report 2). ERCOT's reserves begin dropping below the 13.8% reference margin (Report 3). SPP approaches negative reserve margins (Report 3).
2028–2029: Gas buildout ramps, SMR pilots appear
EIA's 10.6 GW of combined cycle capacity comes online in 2028 (Report 4). Oklo's Aurora pilot at INL potentially achieves first criticality (Report 4). Duane Arnold restart reaches ~615 MW for Google (Report 4). Behind-the-meter capacity approaches 25–38% of data center load nationally (Report 5). The question shifts from "can we build enough?" to "who pays for it?"—PJM data centers have already absorbed $21.3B in capacity costs across three auction cycles (Report 3).
2029–2030: New equilibrium or continued tightening
If the base case demand materializes (550–650 TWh), gas/nuclear/BTM have added 30–50 GW of capacity against 40–60 GW of incremental demand. Reserve margins stabilize in regions with fast gas buildout (ERCOT, Southeast) but remain stressed in PJM and MISO. Kairos Hermes 2 targets operations in 2030 (Report 4). The first GW-scale SMR commercial orders, if they happen, are placed for 2032+ delivery. Transmission projects from 2026–2027 planning cycles begin construction.
8. What Would Invalidate the Thesis
Three factors could materially reduce constraint severity. They are not speculative—each has real evidence behind it—but their combined impact is uncertain.
Demand response and workload flexibility — Highest probability, earliest impact (2026–2028)
Google's 1 GW demand response commitment (March 2026) demonstrated that ML training workloads can be shifted off-peak, cutting peaks 25% for hours at a time (Report 6). Duke University modeling shows US grids could absorb 76–100 GW of new flexible data center load with less than 1% curtailment (under 50 hours/year), potentially avoiding new generation for 5–10 years (Report 6). EPRI's Flex MOSAIC framework classifies data center loads as 18–55% flexible via power capping and rescheduling (Report 6). This is the most likely mitigant because it requires software, not steel—but it demands that hyperscalers accept curtailment, which conflicts with the SLA culture of enterprise AI. If flexibility is mandated in interconnection agreements (as Duke Energy now requires 50 hours/year), it could reduce effective peak demand by 15–25%.
GPU and inference efficiency — Medium-high probability, 2027–2029 impact
Nvidia's Blackwell delivers 25x inference efficiency over Hopper; the forthcoming Vera Rubin platform claims 10x further improvement in inference throughput per watt (Report 6). Inference costs are falling 40x per year on a fixed-performance basis (Epoch AI), meaning the same AI task consumes dramatically less power each hardware generation (Report 6). Groq's ASICs achieve 5–20x performance per watt versus Nvidia GPUs on specific tasks (Report 6). The counterargument is Jevons paradox: cheaper inference drives more usage. Report 6 flags this explicitly—efficiency could cap power per task at 2026 levels, but aggregate demand likely rises. The net effect is probably a 15–30% reduction from the high-end forecasts, not a cancellation of growth.
Accelerated transmission and FERC reforms — Medium probability, 2028+ impact
FERC Order 1920 mandates grid-enhancing technologies (dynamic line ratings adding 40–80% capacity to existing lines) before approving new construction (Report 6). PJM's co-location rules (December 2025 FERC order directing PJM revisions by February 2026) could unlock GW-scale data center connections at existing power plants (Report 5). But the timeline is structural: first projects from the new planning process won't emerge until 2028, with full buildout taking a decade or more (Report 2). Federal siting authority—which could accelerate transmission across state lines—remains politically dormant (Report 6).
Honest assessment: The combination of demand response (2026–2027), inference efficiency (2027–2028), and behind-the-meter self-help (already underway) could reduce grid-served demand growth by 30–40% versus the gross base case. This turns a potential crisis into a severe but manageable challenge—tight reserve margins and elevated prices rather than widespread blackouts. But it does not eliminate the bottleneck sequence: queues, equipment, and generation adequacy bind regardless, just at lower severity. The thesis that different grids hit different breaking points at different times remains robust even under optimistic efficiency assumptions. What changes is the magnitude, not the structure.
Get Custom Research Like This
Start Your ResearchSource Research Reports
The full underlying research reports cited throughout this analysis. Tap a report to expand.
Report 1 Research and synthesize the leading publicly available forecasts for US AI/data center electricity demand through 2030 from EPRI, McKinsey, BCG, Goldman Sachs, IEA, Lawrence Berkeley National Lab, and Grid Strategies. Produce a data table comparing each forecast's methodology, key assumptions (training vs. inference split, GPU efficiency curves, PUE assumptions, hyperscaler vs. enterprise mix), and 2030 demand estimate in TWh and as a percentage of total US electricity consumption. Explain the specific modeling choices that drive the widest divergences and conclude with the most defensible base-case estimate and its key sensitivities.
EPRI's state-level pipeline tracking reveals data centers could drive 9-17% of US electricity use by 2030: they aggregate operational, under-construction, and announced projects into low/medium/high scenarios based on realization rates (e.g., low assumes 25% advanced planning online; high adds 30% early planning), converting IT capacity (56-132 GW) to energy via implicit load factors/PUE, yielding 380-790 TWh—60% above their 2024 forecast due to accelerated AI builds.[1][2][3][4]
- Low: 56 GW IT capacity, ~380 TWh (9%)
- Medium: 96 GW, ~590 TWh (13%)
- High: 132 GW, ~790 TWh (17%)
- % based on EPRI's US-REGEN model total demand (~4,200-4,600 TWh).[5]
- No explicit training/inference split, GPU curves, or hyperscaler mix; focuses on project pipelines vs. equipment shipments (e.g., aligns with LBNL to 2028).[6]
For entrants, EPRI's pipeline method highlights execution risk: only ~65% of announced projects may materialize by 2030 per Grid Strategies benchmarks, favoring incumbents with sited power access over speculative builds.[7]
LBNL's bottom-up equipment model projects 325-580 TWh by 2028 (6.7-12% US total), driven by AI server shipments where training overtakes inference (50-53% AI energy) amid GPU power ramps (60-80% rated) and PUE declines to ~1.4 via liquid cooling/hyperscale shift.[8]
- Servers: 85% hyperscale/colocation by 2028 (vs. enterprise decline); AI servers 80% hyperscale.
- Extrapolating 13-27% CAGR to 2030: ~450-850 TWh.
- Operational: AI training 75-85% uptime (high), inference 37.5-42.5%; PUE simulated thermodynamically (e.g., air chillers 1.4-1.6, liquid lower).
- Matches EPRI through 2028 despite methods (shipments vs. sites).
Competitors must match LBNL's efficiency moat: hyperscalers' 90-99% UPS/liquid cooling yields 20% lower PUE than enterprise, gating AI-scale inference.[8]
McKinsey's workload-driven forecast sees US data centers at 606 TWh (11.7% total) from AI tripling (124 GW incremental global), assuming PUE to 1.1 via chips/cooling but hybrid training/inference racks pushing density.[9]
- 25 GW (2024) to 80+ GW capacity; inference shifts edge/cloud.
- No explicit split; emphasizes GW-scale campuses for training.
New players face $500B+ capex barrier per McKinsey: efficiency (PUE 1.1) and hybrid designs favor hyperscalers scaling AI from training to inference.[9]
| Forecaster | Methodology | Key Assumptions | 2030 TWh | % US Total (implied ~5,100 TWh EIA AEO2026)[10] |
|---|---|---|---|---|
| EPRI[2] | Pipeline (sites/construction/announced) to IT GW → TWh | Realization rates; implicit LF/PUE | Low: 380 Med: 590 High: 790 |
9/13/17%[3] |
| LBNL (ext.)[8] | Bottom-up: shipments × power × util × PUE | Train>inf by '28; GPU 60-80% rated; PUE~1.4; 85% hyperscale | 450-850 | 9-17% |
| McKinsey[9] | Workload/capex growth (AI/HPC) | PUE 1.1; train/inf hybrid; 80+ GW | 606 | 11.7% |
| IEA[11] | Global model; US ~40% share | +240 TWh on 185 (2024); eff gains | ~425 | ~8% |
| Goldman Sachs[12] | Server shipments × intensity | 60% global in US; AI 39% | ~750 | ~15% |
| BCG (est.)[13] | Compute modeling | High AI/workload; global analogs | ~970 | ~19% |
| Grid Strategies[7] | Utility FERC agg. (adj. down) | High LF critique; benchmarks 65 GW | ~65 GW (~500 TWh est. LF=90%) | ~10% peak growth |
Widest divergences stem from pipeline (EPRI: sites) vs. shipments (LBNL: GPUs) vs. workloads (McKinsey/GS: AI adoption): EPRI/Grid cap at announced (risking overbuild), LBNL moderates via util/PUE, high-end (BCG/GS) assumes unchecked inference explosion outpacing eff gains (e.g., GS +220% global ignores LF<100%).[7][8]
Defensible base case: 500-600 TWh (10-12% US total), blending EPRI med/McKinsey/LBNL high—aligns bottom-up (equip eff) with pipelines, assuming hyperscaler dominance (85% servers), PUE~1.2-1.4, train/inf ~50/50. Sensitivities: +20% if inference >train (GS risk); -15% faster GPU eff/PUE<1.2; ±10% realization (Grid adj.).[10]
Entrants: target inference efficiency (70% lifecycle per LBNL) and co-locate with renewables for PUE edge, as pipelines favor power-secured hyperscalers.
Recent Findings Supplement (April 2026)
EPRI's February 2026 Update Resets the High End of Forecasts
EPRI's "Powering Intelligence 2026" report, released February 25, 2026, sharply raised its US data center electricity projections by 60% versus its 2024 estimates, driven by 18 months of accelerated AI-fueled project announcements; it uses state-level commercial development pipelines (operational, under construction, advanced/early planning) rather than equipment shipments, assuming most new capacity ramps quickly with type-specific load factors (higher for AI/hyperscalers), cooling/overhead loads, and PUE implicitly around 1.2-1.4 based on industry trends—yielding low/medium/high 2030 demand of 384/596/793 TWh (9/13/17% of ~4,400 TWh total US generation), aligned with LBNL's 2028 band but extending to 2030 via project realization rates (e.g., high assumes all construction/advanced +30% early planning).[1][2][3][4][5][6]
- Capacity: 56 GW low (384 TWh), 96 GW medium (596 TWh), 132 GW high (793 TWh); 2024 baseline 35-44 GW / 177-192 TWh (4-5%).
- Consistent with LBNL to 2028 (325-580 TWh) despite methods divergence: EPRI favors hyperscaler/AI buildout data over LBNL's bottom-up server shipments.[7]
For competitors: Project pipelines undervalue efficiency gains or overstate realization without EPRI's granular state data; sensitivities include planning-stage dropouts (bearish) or crypto/AI surge (bullish).
Goldman Sachs February 2026 Revision Ties Growth to Hyperscaler Capex and Inference Intensity
Goldman Sachs updated its global data center forecast upward to 220% growth by 2030 vs. 2023 (from prior 175%), implying ~543 TWh baseline +905 TWh growth to 1,350 TWh total (~60% US share, or ~810 TWh US including baseline ~200 TWh), modeling server shipments (TMT team revisions), genAI efficiency (low-double-digits annual) offset by power-dense AI servers/GPUs (e.g., +68% per-server for Rubin-era), rising inference mix (greater intensity vs. training), and capex surges; US capacity to 95 GW by 2030 (~10-12% of ~5,000-5,600 TWh total US electricity).[8][9]
- US skew: 60% of incremental demand (vs. prior 50%), driven by hyperscaler reinvestment >$300B upward to 2026-27.
- Assumptions: Inference > training power (debated pervasiveness/automation); non-AI ~10% annual growth; PUE unstated but implied via capacity-to-energy.
For entrants: Server-level modeling misses site constraints; key sensitivity is inference efficiency breakthroughs (could cut 20-30% off high-end).
| Forecaster | Date | Methodology | Key Assumptions | 2030 US DC Demand (TWh / % Total US) |
|---|---|---|---|---|
| EPRI (Low/Med/High) | Feb 2026 | State-level project pipelines (const./adv./early planning %) to IT capacity, then load factors/PUE | AI/hyperscaler high load factor; ramp rates; cooling/overhead; 60% > prior EPRI | 384 / 9%; 596 / 13%; 793 / 17% (total US ~4,400 TWh)[5][6] |
| LBNL | Dec 2024 (cited 2026) | Bottom-up IT shipments (servers/storage/net), ops util., PUE/WUE sims | GPU/AI growth (infer. 60%→50%); PUE 1.15-1.35; 50% util.; hyperscale dom. | 325-580 (2028: 6.7-12%); trend to ~500-700 (2030 est., my inference)[7] |
| Goldman Sachs | Feb 2026 | Server shipments, AI mix/efficiency curves, capex | Inference intensity↑; GPU power-dense; low-DD eff.; ~60% US growth share | ~810 total (~543 base+growth; 14-16%)[9] |
| Grid Strategies (agg. utilities) | Nov 2025 | Utility/RTO FERC filings aggregation | High DC load factors (~96%); ~55% growth from DC (90 GW peak) | ~900-1,000 (16-18% of 5,591 TWh total; likely -25 GW/40% DC overstate)[10] |
| BCG | Mar 2026 | Bottom-up peak capacity modeling | AI buildout to firm power gap; hyperscaler focus | 50-80 GW gap (implies ~400-650 TWh @ high util.; 8-12%)[11] |
Divergences Stem from Pipeline Realization vs. Equipment Constraints
Widest gaps (EPRI low 384 TWh vs. high 793 TWh; utilities overstated per Grid Strategies) arise from project completion (EPRI: 25-100%+ planning stages) vs. bottom-up limits (LBNL/Goldman: GPU shipments/util./PUE 1.15-1.35, inference 40-60% mix); utilities assume near-100% load factors without supply bottlenecks, while analysts cap at 50-96%—EPRI/LBNL converge mid-500s TWh as EPRI includes crypto/enterprise beyond hyperscalers.[10][7]
For market players: Over-reliance on announcements risks 40% shortfall; PUE drops or inference optimization most sensitive (10-20% swing).
Grid Strategies Highlights Utility Optimism on DC Load Factors
November 2025 aggregation of utility/RTO forecasts shows 166 GW peak growth (20% total), 90 GW (55%) DC at ~96% load factor driving energy to 5,591 TWh total US (+32%), but likely 25 GW/40% DC overstatement vs. analysts (e.g., 65 GW max via chips); no split assumptions detailed, but implies hyperscaler-heavy with minimal efficiency curves.[12][10]
Implication: Regional queues undervalue delays; competitors need interconnection reforms.
No Major Post-2025 Updates from McKinsey, BCG, IEA
McKinsey (pre-2026 cites) ~35 GW US DC power (~300-400 TWh est.); BCG Mar 2026 flags 50-80 GW firm power gap (no TWh); IEA Apr 2026 global double to 945 TWh (US ~40-50% share, ~400 TWh base); methods emphasize AI inference but lack new US-specific TWh.% no direct post-Oct 2025 shifts.[13]
Defensible Base Case: 550-650 TWh (11-13% of US Total)
Mid-range synthesis (EPRI medium 596 TWh, LBNL trend-adjusted, Goldman US share) at 550-650 TWh / 11-13% (~5,000 TWh total US per Grid/IEA trends), balancing pipelines (bullish AI/hyperscaler) with shipments/PUE (conservative util. 50%, inference eff.); most defensible as EPRI/LBNL methods converge despite differences, utilities high-bias corrected.[10]
- Sensitivities: +20% if planning >65 GW (Grid adj.); -15% PUE/inference gains; policy (e.g., queues) ±10%.
To compete: Focus midcase, hedge via behind-meter nuclear/gas-CCUS (BCG rec.); no recent regulatory shifts noted.
Report 2 Research the five specific constraint categories facing US electric grids under accelerating data center load: (1) interconnection queue depth and wait times across PJM, ERCOT, MISO, CAISO, and NYISO as of 2024–2026; (2) transmission capacity gaps and FERC Order 1920 implementation status; (3) generation adequacy signals including PJM capacity auction price spikes and reserve margin trends; (4) equipment supply chain constraints including publicly reported large power transformer lead times (2–4+ years) and backlogs at GE Vernova, Siemens Energy, and Hitachi Energy; and (5) natural gas pipeline capacity and water/cooling constraints in arid data center hubs. For each bottleneck, identify when it binds first, which regions are most exposed, and what evidence exists that it is already delaying data center energization. Produce a ranked summary of which constraint bites hardest and soonest.
1. Interconnection Queue Backlogs: PJM and ERCOT Lead with Multi-Year Delays Already Impacting Data Centers
PJM and ERCOT have transformed legacy "first-come, first-served" serial queues—originally designed for small-scale generation—into massive cluster-study processes under FERC Order 2023, where projects are batched for simultaneous impact analysis to equitably allocate upgrade costs; however, explosive data center "large load" requests (non-generator) now overwhelm these systems, forcing 4-8 year waits that exceed data center construction timelines (18-24 months), pushing developers toward behind-the-meter generation or speculative multi-queue filings.[1][2]
- End-2024 generator queues: MISO (2,213 projects/447.5 GW), ERCOT (1,447/346.1 GW), CAISO (638/272.8 GW), PJM (1,942/211.5 GW), NYISO (402/78.5 GW); U.S. total active queue fell 12% to 2,290 GW amid reforms, but withdrawals hit 700 GW.[1]
- ERCOT large-load queue exploded to 410 GW by April 2026 (87% data centers, up from 233 GW end-2025, 63 GW end-2024), exceeding peak demand (85 GW) 4.8x; PJM Cycle 1 (2026) received 811 requests amid paused new gen queues until 2026.[3][4]
- Median queue-to-operation: 55 months (2024 completions); PJM projects operational in 2025 averaged 8 years; ERCOT large-load grew 300% in 2025 alone.[1][2]
Implications for Competitors/Entrants: Data centers in PJM/ERCOT face energization no earlier than 2028-2030; pursue co-location (e.g., PJM's EIT fast-path, 10/month) or off-grid gas turbines; MISO/NYISO/CAISO less acute but trending up (NYISO 6 GW large loads, avg 6.5-year wait).
2. Transmission Gaps Persist Despite FERC 1920; Compliance Lags Limit 20-Year Planning for Data Center Hubs
FERC Order 1920 mandates 20-year scenario-based regional transmission planning (e.g., high load growth, electrification) with benefit-cost tests for selection, forcing proactive identification of needs like data center clustering in Virginia/Texas; but compliance filings due June 2025 were extended (e.g., PJM to Dec 2025, MISO June 2026), delaying implementation amid only 55 miles new HV lines built in 2023, creating 100s GW gaps that cascade into queue restudies.[5][6]
- U.S. queues reflect transmission scarcity: 81% solar/storage but high withdrawals (77% PJM) due to upgrade costs; PJM approved $6.7B 765kV backbone Feb 2025 for VA data centers.[1]
- Compliance: PJM/MISO filed Dec 2025 (extensions); CAISO Dec 2025; NYISO Apr 2026; ERCOT (non-FERC) separate; first cycles start 2026-2027.[7]
- Evidence: PJM load forecast up 5 GW (data centers); only 5,000 circuit-miles added 2024, mostly reliability-driven.[8]
Implications for Competitors/Entrants: Gaps bind 2027+ in PJM/ERCOT; lobby for Order 1920 state involvement; site in SPP/MISO for JV transmission (e.g., JTIQ 345kV lines by 2029).
3. Generation Shortages Signal Imminent Reliability Risks: PJM/MISO Auctions Spike on Data Center Load
PJM's Reliability Pricing Model auctions procure capacity 2-3 years ahead via descending-clock bids, but data center-driven load forecasts (non-curtailable) have eroded reserve margins below targets, spiking prices ~10x as retirements/dequeued renewables fail to offset; MISO followed suit, exposing PJM (14% reserves) and MISO first.[9][10]
- PJM 2025/26: $270/MW-day (from $29), $14.7B total; 2026/27: $329 cap; 2027/28: $333 cap, 14.4% reserve (short 6.6 GW vs 20% target) due to +5 GW large loads.[10]
- MISO summer 2025: $666/MW-day (from $30), reserves drop to 2.6 GW surplus; ERCOT peak forecast triples to 278 GW by 2029.[11]
- Trends: PJM IRM up to 20%; only 19% historical queue completion rate.[1]
Implications for Competitors/Entrants: Auctions bind now (PJM 2025/26); self-supply via FRR or bilateral gas PPAs; MISO/ERCOT next by 2027.
4. Transformer Supply Chains: 3-5 Year Lead Times Already Delaying 50%+ of 2026 Data Centers
Large power transformers (LPTs >100 MVA) require custom-wound copper/steel cores shipped globally, with U.S. capacity strained by data center substation needs; post-2020 demand +116% created 128-week (2.5yr) averages (up to 4-5yr for 500kV), backlogs at GE Vernova/Siemens/Hitachi, forcing 40-50% of 2026 U.S. data centers to delay/cancel as procurement precedes construction.[12][13]
- Lead times: 120-210 weeks (2024 NERC); prices +79%; GE Vernova Q1 2026 data center orders = all 2025; Siemens/Hitachi new U.S. plants 2027-2028.[14]
- Evidence: Half 2026 data centers delayed (Bloomberg); WoodMac 30% 2025 deficit; China imports +433%.[15]
Implications for Competitors/Entrants: Procure now for 2029; vertical integrate (e.g., GEV Prolec); dry-cooling/off-grid to bypass.
5. Gas Pipelines and Arid Cooling: Secondary Constraints Emerging in ERCOT/Arizona
Natural gas pipelines operate at firm capacity limits, with data center on-site turbines straining distribution (not production); arid hubs like AZ/TX face water-for-cooling shortages (620M gal/yr per site), delaying permits vs. humid VA/PA.[16][17]
- ERCOT: 58 GW gas planning (half data centers); Kinder Morgan $9.1B backlog; 40% U.S. data centers gas-powered.[18]
- AZ: Data centers 80% APS growth; Tucson "Project Blue" 620M gal/yr rejected; dry-cooling proposed.[17]
Implications for Competitors/Entrants: Pipeline expansions lag 2-3yr; prioritize humid East/South; hybrid air-cooling.
Ranked Constraints: Hardest and Soonest Impacts
- Interconnection Queues (Bites Now, PJM/ERCOT Worst): 4-8yr delays already stalling data centers; evidence: ERCOT 410 GW queue, PJM 8yr avg.[3]
- Generation Adequacy (2025/26 Auctions): PJM 14% reserves, prices capped; MISO $666 spikes.[10]
- Transformers (2026-2029 Builds): 50% delays; 3-5yr leads bind physical energization.[15]
- Transmission (2027+ via 1920): Gaps exacerbate queues; compliance ongoing.
- Gas/Water (Regional, 2027+): ERCOT/AZ exposed but mitigable via imports/dry-cooling.
Confidence: High on queues/auctions (direct LBNL/PJM data); medium on transformers (analyst reports); low on exact DC delays (inferred from leads/queues). Further utility filings needed for DC-specific evidence.
Recent Findings Supplement (April 2026)
Interconnection Queue Depths and Wait Times
PJM and ERCOT lead in exposure to data center-driven delays, with queues now processing large loads separately but still facing 3-7 year timelines due to backlog clearance and transmission studies; MISO's 242 GW queue (DPP-2025 cluster with 78 GW) signals multi-year waits until 2036 for some projects, while CAISO's strict intake cut 67% of submissions.[1][2]
- Queued Up 2025 (Dec 2025): PJM paused new requests until 2026, transition queue ~30 GW; MISO delayed 2023 window to 2024; CAISO Cluster 16 delayed, no 2024 requests.[1]
- ERCOT large load queue jumped 300% in 2025 to 226 GW (73% data centers), with batch studies ongoing but new rules risking curtailment.[3]
- Evidence of delays: PJM proposes load queue for data centers; 50% of 2026 U.S. data centers (12 GW planned) delayed/canceled, only 5 GW under construction due to 5-year waits.[4]
/grok:render
Implications for entrants: Prioritize regions with "executed" agreements (e.g., ERCOT/SPP joint queues); co-locate with generation or accept behind-the-meter gas to bypass, but expect 2-3 year head start needed for viable projects.
Transmission Capacity Gaps and FERC 1920 Status
FERC Order 1920 compliance filings began Dec 2025 (PJM/CAISO first), but first projects not until 2028+5-10 years buildout; gaps persist as data centers add 166 GW peak load forecast by 2025, overwhelming proactive planning in PJM/ERCOT where queues exceed risked generation capacity 3x.[5][6]
- PJM filed long-term process Dec 2025, extending to 20-year horizon with state input; MISO/SPP due June 2026; NYISO April 2026.[7]
- Gaps: PJM ARM risks below target by 2029; ERCOT peak to 139 GW by 2030 strains existing lines, quintupling congestion without upgrades.[8]
- Delays evident: 183 GW large loads with agreements =22% U.S. peak, but PJM utility commitments 3x accredited queue.[9]
Implications for entrants: Fund transmission co-investments via standardized agreements; early movers in compliant regions (PJM) gain, but 15-year horizon means off-grid gas now essential for 2026-28.
Generation Adequacy and Capacity Signals
PJM capacity auctions hit caps ($333/MW-day 2027/28), procuring 14.8-14.4% reserve vs. 20% target (short 6.5 GW), with data centers driving $23B costs across 3 auctions; MISO summer 2026/27 at ~$400/MW-day (down from $666 2025/26), but tightening.[10][11]
- PJM: 2027/28 short 6,517 MW UCAP; data centers caused $21.3B (45%) of $47.2B cleared costs; proposing price collar extension to 2030.[12]
- Trends: PJM peak growth 4.8%/yr to 2035; NERC LTRA (Jan 2026) flags PJM below IRM 2029.[8]
- Delays: Shortfalls force data center backups (e.g., DOE winter 2025/26 directive); auctions signal scarcity biting 2026/27.[13]
Implications for entrants: "Bring-your-own" generation mandatory in PJM; bid into auctions early or face 262% capacity cost hikes passed to loads.
Equipment Supply Chain Constraints
Large power transformer (LPT) lead times hit 3-5 years (130-200 weeks vs. 50 pre-2022), with GE Vernova Prolec backlog $5B (+25% post-acquisition), Siemens/Hitachi expanding U.S. plants (2027-28 online); data centers drove GEV electrification orders > all 2025 in Q1 2026 alone.[14][15]
- GEV: $163B backlog Q1 2026, gas turbine slots tight 2029-30 (3-yr leads); Prolec adds data center line.[16]
- Shortages: 50%+ 2026 data centers (12 GW) stalled; U.S. imports from China surged 5x; GSU demand +274% since 2019.
/grok:render
- Evidence: WoodMac Q2 2025: 3-yr LPT/GSU waits; half U.S. distribution transformers past life.[17]
Implications for entrants: Lock transformer orders 3-5 yrs ahead (China/Mexico now); U.S. expansions too late for 2026—favor pre-ordered sites.
Gas Pipeline and Water/Cooling Constraints
ERCOT/Texas arid hubs face gas pipeline bottlenecks for behind-the-meter plants (100 GW pipeline, but 5 GW sites need new laterals turning into majors); water use 16-33B gal/yr by 2028 strains cooling, with 20+ projects delayed 2025 by opposition.[18][19]
- Gas: 40% data centers gas-powered; Williams $5B pipelines, but constraints push to TX/PA/NM; 1/3 new gas for on-site.[20]
- Water: Hyperscalers expect cooling/water as bottlenecks post-power; AZ/NM bills for on-site nuclear/gas.[21]
- Delays: Fermi 6 GW TX needs 90 turbines; permitting/community blocks 20+ sites.[22]
Implications for entrants: Site near pipelines (TX hubs); liquid cooling + CHP to cut water 50%, but expect 16-20 mo added permitting.
Ranked Constraints: Hardest and Soonest
- Interconnection queues (bites 2026, PJM/ERCOT worst): 5-8 yr waits already canceling 50% 2026 capacity; first bind for new loads.
post:0 /grok:render - Equipment (LPTs/transformers, 2026-28): 3-5 yr leads kill half projects; non-obvious China reliance amplifies tariffs risk.[15]
- Generation adequacy (PJM now, escalating 2027): Auction shorts/revenue spikes ($23B data center hit) force BYO or backups immediately.[11]
- Transmission/1920 (2028+): Filings underway but buildout 10-15 yrs; gaps widen short-term.[6]
- Gas/water (mid-2027, TX/AZ): Secondary for on-site shift, but permitting/local opposition emerging.[18]
Overall for competitors: Queues/equipment bind first/hardest—viable paths: executed PPAs, off-grid gas (33% by 2030), or pre-2023 sites; confidence high on 2025/26 data, but 1920 impacts inferred pending full rollout (additional RTO filings needed).
Report 3 Research the specific constraint profiles of PJM (Northern Virginia data center alley, capacity auction dynamics, queue collapse), ERCOT (Texas data center build rate, fast-build gas, weather risk), MISO (reserve margin trends, slower data center growth), CAISO (transmission constraints, gas peaker retirements), NYISO (downstate congestion, generation retirement), and SPP/Southeast including TVA and Duke territory (re-shoring plus data center demand growth). For each region, identify the primary binding constraint, the earliest projected stress event, and any publicly reported cases of data center projects facing energization delays or capacity allocation failures. Produce a ranked risk matrix showing constraint severity vs. timing by region through 2030.
PJM (Northern Virginia Data Center Alley)
PJM's Reliability Pricing Model (RPM) capacity auctions have failed to procure sufficient resources to meet reliability targets due to explosive data center load growth in Northern Virginia's Dominion zone, where real-time sales data and interconnection queues now stretch to 7 years, forcing developers to delay energization or seek colocation with existing generation; this mechanism—bidding three years ahead—exposes shortages early but amplifies costs as retirements and queue backlogs (140 GW generation queue) prevent new supply, with the December 2025 auction for 2027/2028 falling 6,516 MW short for the first time in history.[1][2]
- Capacity prices surged 833% to $269.92/MW-day (RTO) in 2025/2026 BRA, $329/MW-day in 2026/2027, and $333.44/MW-day in 2027/2028, all hitting FERC caps; data centers drove 97% of 5,250 MW load growth forecast.[1][3]
- Dominion queue: 30 GW+ data center demand by 2030; congestion costs up 64% to $1.7B in 2024; $6.7B 765kV backbone approved Feb 2025 but years from completion.[4][5]
- Public cases: Under-construction capacity fell 29% in NoVA (2025); projects entering queue now face 2028+ agreements; first-ever auction shortfall signals 24-55 GW gap by 2030-2035.[2]
Implications for competitors/entrants: New data centers in Loudoun/Dominion face 5-7 year waits; colocation (e.g., nuclear) or behind-the-meter gas bypasses queues but requires FERC tariff changes; non-obvious: backstop auctions (14.9 GW bilateral by 2026) allocate costs to loads, raising entry barriers unless flexible demand (curtailment) qualifies for credits.[6]
ERCOT (Texas Data Center Build Rate)
ERCOT's shift to batch interconnection studies amid 410 GW large-load queue (87% data centers) creates a "speculative bubble," where only 1.8% of requests are energized, delaying projects via stricter viability rules (SB6) and fast-build gas peakers, but weather risks (e.g., 2021 freeze) amplify shortages as reserves tighten post-2028 despite 10 GW Texas Energy Fund gas.[7][8]
- Queue exploded: 63 GW (Dec 2024) to 233-410 GW (late 2025); peak forecast 278 GW (2029), 368 GW (2032)—4x current 85 GW peak.[8][9]
- No formal capacity market; reserves drop below 13.8% reference post-2028; 70%+ queue data centers, but curtailment rules for non-firm loads.[10]
- Cases: No specific failures reported, but PUCT rejected inflated forecasts; 26% national delays (2025), queue reforms wash out speculation; Permian Tx upgrades ($30B) lag build rate.[4]
Implications for competitors/entrants: Fast-build gas (2.5 GW replacements) favors incumbents; entrants need firm commitments or off-grid/onsite gen to bypass 4x peak queue; weather risk means diversified backup or curtailment contracts essential by 2028 deficits.[11]
MISO (Reserve Margin Trends)
MISO's declining reserve margins (7.9% summer PY2025/26, dropping to 4.3% by 2029) stem from slower data center growth vs. PJM/ERCOT but coal/gas retirements outpacing solar/battery queue (296 GW backlog), with ERAS fast-track for 3 GW+ urgent needs highlighting allocation failures for non-dispatchable resources.[12][13]
- Peak load +35% to 163 GW by 2035 (data centers 20% electricity by 2030); surplus fell 43% to 2.6 GW (2025/26); PRA summer price $666.50/MW-day.[14]
- Queue reforms reduced to 174 GW (Nov 2025); elevated risk 2027+ per NERC; 8-14 GW data centers 2026-27 uncertain.[15]
- Cases: No specific delays; 43% CAGR data centers (2020-25), but "right-sizing" emerging; LOLE study shows winter risks rising.[13]
Implications for competitors/entrants: Slower growth eases pressure, but PRM target (7.9%) demands dispatchable capacity; ERAS prioritizes firm resources—flexible loads/data centers gain edge; compete via co-location or DR for accreditation.[16]
CAISO (Transmission Constraints)
CAISO's transmission planning now prioritizes reliability-driven upgrades ($7B for 38 projects, 2025-26 plan) as data center load (1.8 GW by 2030) exacerbates Path 15/Silicon Valley congestion amid gas peaker retirements (3.7 GW once-through cooling), delaying remote clean energy delivery and forcing local backups.[17][18]
- Load +15 GW (2035); data centers key in Bay Area overloads; no RA shortage yet, but queue paused/reformed (185 GW backlog).[17]
- $1.4B Silicon Valley project for large loads; high gas retirement sensitivity (11 GW by 2034) strains locals.[19]
- Cases: Utilities see rising apps; no public failures, but national 26-50% delays; TPP concurs post-forecast loads.[18]
Implications for competitors/entrants: Transmission-limited; large loads post-TPP need PTO concurrence—Bay Area first; flexible/interruptible preferred amid retirements; compete by co-developing Tx upgrades.[20]
NYISO (Downstate Congestion)
NYISO's downstate (NYC/Long Island) faces Local Capacity Requirements (LCR) deficits from peaker retirements (1.5 GW ozone-season by 2025) and data center/EV growth, with Central East congestion blocking upstate hydro/nuclear, projecting statewide shortfalls by 2034 absent CHPE (2026).[21]
- Gold Book: 3 GW+ large loads queued (data centers/Micron); NYC reliability need summer 2025 resolved via peaker retention til CHPE.[22]
- IRM 25.3% (2026-27); MLCR NYC 82.6%, LI 106.7%; aging fleet +3 GW data centers strain.[23]
- Cases: Queue 12 GW large loads (Jan 2026); 3-year moratorium proposed; no specific failures but transmission overload risks.[24]
Implications for competitors/entrants: Downstate LCR binding—data centers need firm imports or onsite; upstate viable but congestion caps; policy risk high (moratoriums); DR/SCR critical for accreditation.[25]
SPP/Southeast (TVA/Duke Territory)
SPP's reserve margins plummet from 20.7% (2025) to -1.6% (2030) as data center/re-shoring demand outpaces intermittent queue, with no capacity market forcing bilateral risks; TVA/Duke see 37 GW queued but delays from equipment/power shortages mirror national 30-50% slip.[26]
- SPP BA: PRM rises to 17% (2029), but gap widens; Southeast utilities plan 10 GW gas for speculative DC growth (0.2% probability).[27]
- TVA/Duke: Cluster queues (2025-26); Carolinas DC 10% sales by 2030; re-shoring (chips) adds uncertainty.[28]
- Cases: National delays 26% (2025); no region-specific failures; Duke high-load tariffs enforce 75% take-or-pay.[29]
Implications for competitors/entrants: No auction safety net—bilaterals rule; SPP winter PRM 36% demands firm capacity; Southeast overbuild risk (half DC speculative); on-site gen or flexibility de-risks.[30]
Ranked Risk Matrix: Constraint Severity vs. Timing (2025-2030)
| Region | Primary Constraint | Severity (1-5) | Earliest Stress | 2025-26 | 2027-28 | 2029-30 | Key Metric |
|---|---|---|---|---|---|---|---|
| PJM | Capacity Auction Shortfalls/Queue | 5 | 2025 Auction (DOM/BGE Caps) | ●●●●● | ●●●●● | ●●●●● | -6.5 GW Short (2027/28)[1] |
| ERCOT | Interconnection Queue Overload | 4 | 2028 Reserves | ●●●● | ●●●●● | ●●●●● | <13.8% RM Post-2028[10] |
| NYISO | Downstate LCR/Congestion | 4 | Summer 2025 (Peakers) | ●●●●● | ●●●● | ●●●● | NYC Deficit til CHPE 2026[22] |
| MISO | Declining RM (Queue) | 3 | 2027 Elevated Risk | ●●● | ●●●● | ●●●●● | 4.3% RM (2029)[15] |
| SPP/SE | Reserve Erosion | 3 | 2027 Deficit | ●●● | ●●●● | ●●●●● | -1.6% RM (2030)[26] |
| CAISO | Tx Constraints | 2 | 2026+ Overloads | ●● | ●●● | ●●● | Bay Area DC-Driven[17] |
Severity: 5=Immediate shortages/delays; 1=Ample headroom. Timing: ●=Stress level (5 bullets=peak). PJM leads near-term; ERCOT/SPP long-term. Confidence: High (NERC/PJM auctions verified); further queue tracking needed.[15]
Recent Findings Supplement (April 2026)
PJM (Northern Virginia Data Center Alley)
PJM's capacity market has entered scarcity pricing due to data centers adding over 13 GW of forecasted load in the Dec 2025 2027/2028 auction, outpacing supply amid a clogged queue transition and 3-4 year supply chain delays for transformers/gas turbines; this drove a 285% capacity cost spike to $10.39B in 2025 (from $2.69B in 2024), with congestion up 78% to $7.3B.[1][2]
- 2027/2028 Base Residual Auction (Dec 2025 results, Feb 2026 report): Cleared at cap $333.44/MW-day; procured 145,777 MW UCAP, 6,517 MW short of reliability requirement; 14.4% IRM vs 20% target (5.6 pp shortfall), first miss >1 pp, triggering investigation and potential backstop auction if persists.[3][2]
- Apr 2026: PJM proposes 14.9 GW two-phase backstop (bilaterals then auction) for data centers/large loads; FERC eyes June decision; governors push data centers bear full costs.[4]
- Queue reforms (post-2025): Cleared 14 GW in Cycle 1; 24 GW projects terminated since 2020; data centers face no explicit delays but contribute to "transition gap" with 82 GW peak demand growth forecast to 2040.[2]
Implications for Competitors/Entrants: Data centers must self-procure via co-location or backstops (15-yr contracts eyed); queue fast-track (RRI) helps but permitting/supply chains delay new gen 24-44 months; non-Virginia sites less stressed short-term.
NYISO (Downstate Congestion)
Downstate (NYC/Long Island) faces binding transmission bottlenecks from upstate renewables to load, worsened by peaker deactivations like Gowanus/Narrows barges (608 MW, Zone J); 19 large loads (3+ GW incl. data centers/Micron) queueing amid 1.2 GW retirements over decade.[5][6]
- NYSRC IRM Filing (Dec 2025): 2026-2027 IRM at 24.5% (up from 24.4%); deactivations cut IRM 0.69%; Zone J MLCR 79.2% (+3.6 pp), G-J 88.75%; CHPE (1.25 GW summer-only) boosts J by 8.47% but winter fuel risks rise to 14% LOLE share.[6]
- Reliability needs: Summer 2026 downstate shortfalls (410-650 MW, 6-8 hrs); NYC potentially deficient by 2033 without CHPE/Empire Wind; LOLE nears 0.1 days/yr by 2034.[5]
No reported data center delays; queue reforms refocus storage downstate.
Implications for Competitors/Entrants: Co-locate with upstate imports or storage; downstate land/congestion favors smaller/distributed loads; CLCPA mandates (70% RE by 2030) prioritize renewables but delay hyperscalers.
SPP/Southeast (TVA/Duke Territory, Re-shoring + Data Centers)
SPP forecasts peak doubling to 109 GW by ~2035 amid 26 GW large load requests since 2020 (9 GW data centers), tightening PRM (16% summer/35% winter Year 10); Duke sees industrial/data center demand up 6.1% CAGR next 5 yrs, queue at 5.6 GW peak.[7][8]
- SPP 2025 ITP (Nov 2025): 11 GW spot loads (Future 2); peak from 61.7 GW (2026) to 76.4 GW (2034); $19.4B portfolio (2,921 mi new HV incl. 765 kV overlay) for congestion/voltage; reserve met with scoped RE but load shed risks in extremes.[8]
- Duke (Feb-Apr 2026): 38 advanced projects (5,610 MW peak, 80% data centers); $102B capex thru 2030; contracts mandate 50-hr/yr curtailment for faster hookup.[9]
No specific delays/failures; HILL framework expedites large loads.
Implications for Competitors/Entrants: Regulated returns favor utilities; curtailment clauses enable entry but cap firm power; re-shoring (oil/gas/manufacturing) competes with data centers for southern capacity.
CAISO (Transmission Constraints)
CAISO's Apr 2026 draft plan flags Path 15 north-south congestion as binding, needing 500 kV line amid 15 GW load growth by 2035 (20 GW by 2040, incl. data centers); reconductoring/advanced conductors prioritized over peakers (no retirements noted recently).[10]
- Recommends 38 upgrades ($7B) for electrification/manufacturing/large loads; Greater Bay/Tesla corridors targeted.
No data center delays reported; queue not detailed.
Implications for Competitors/Entrants: Northern imports strained; on-site PPAs or storage needed; policy pushes RE but transmission lag favors distributed gen.
ERCOT, MISO: Limited New Data Post-Oct 2025
ERCOT: 226 GW large loads monitored (Dec 2025, ~75% data centers); may delay energization for stability but no specific failures/delays found recently.
MISO: No new reserve margin/queue updates; tightening noted in NERC LTRA (Jan 2026) but pre-existing.[11]
Implications: ERCOT fast-build gas mitigates weather risk short-term; entrants watch PUCT rules (e.g., SB6 financials for 75+ MW).
Ranked Risk Matrix: Severity (High/Med/Low) vs Timing to 2030
| Region | Primary Constraint | Earliest Stress | 2026-27 | 2028-30 | Notes |
|---|---|---|---|---|---|
| PJM | Capacity Shortfall/Queue | 2027/28 Delivery (ongoing) | High | High | Backstop eyed; 60 GW gap possible.[2] |
| NYISO | Downstate Congestion/Retirements | Summer 2026 | High | High | 3 GW loads queueing.[5] |
| SPP/SE | Load Growth/Transmission | 2030 (PRM negative) | Med | High | 26 GW requests; Duke curtailments.[7] |
| CAISO | Transmission (Path 15) | 2035 (15 GW load) | Med | Med | $7B plan eases.[10] |
| ERCOT | Large Load Queue | N/A (monitoring) | Low | Med | Potential delays but fast gas.[11] |
| MISO | Reserve Trends | N/A | Low | Low-Med | No new data; slower growth. |
Confidence: High on PJM/NYISO (direct reports); medium elsewhere (sparse post-Oct 2025 data). Additional RTO filings would refine 2026-30 projections.
Report 4 Research the publicly announced timelines and feasibility challenges for each major supply-side response to AI power demand: nuclear restarts (Three Mile Island Crane Clean Energy Center restart, Palisades, Duane Arnold), SMR and microreactor commercialization (Oklo, NuScale post-bankruptcy, NuScale NNE, Kairos Power timelines), natural gas combined cycle and peaker additions, utility-scale and behind-the-meter battery storage, behind-the-meter generation including gas turbines and fuel cells (Bloom Energy), and on-site nuclear or gas at hyperscaler campuses. For each category, identify what capacity in MW or GW can realistically come online in 2026, 2027, and 2028+, what the key permitting, financing, or supply chain barriers are, and which projects have publicly announced offtake agreements with hyperscalers or data center operators.
Nuclear Restarts
Constellation's Crane Clean Energy Center (formerly Three Mile Island Unit 1) leverages its existing 835 MW pressurized water reactor infrastructure—shut down in 2019 for economic reasons but maintained in operable condition—to restart via a $1.6B refurbishment including turbine, generator, and control system upgrades, enabling rapid reactivation compared to new builds; this data moat of pre-licensed, fueled-ready hardware secures Microsoft's 20-year PPA for full output to match AI data center loads in PJM, but grid interconnection delays from PJM's transmission queue threaten full deliverability until 2031.[1][2][3]
- Targeting H2 2027 restart (advanced from 2028), backed by $1B DOE loan; 835 MW capacity fully contracted to Microsoft via 20-year PPA extending to 2054.[4][5]
- Holtec's Palisades (800 MW PWR) achieved NRC approvals for fuel loading and transitioned to operational status in 2025, but delays pushed restart to early 2026; no hyperscaler PPA announced, financed by $1.52B DOE loan for operations to 2051.[6][7]
- NextEra's Duane Arnold (615 MW BWR, shut 2020 post-derecho) targets Q1 2029 via Google's 25-year PPA for majority output to Iowa AI/cloud ops; NRC licensing bundle by Jan 2028, $1.6B+ cost; training ramps up but workforce recertification lags.[8][9]
Capacity online: 0 GW in 2026; ~0.8 GW Palisades in early 2027; 0.8-1.6 GW cumulative by 2028+ (Crane + Duane delayed); barriers: NRC relicensing (2-3 yrs), grid queue (PJM to 2031), $1-2B financing per plant; competing entrants must navigate FERC capacity rights transfers and state water permits, favoring incumbents with DOE loans over greenfield nuclear.
SMR/Microreactor Commercialization
Oklo's Aurora fast-fission microreactor (75 MW/unit, recyclable fuel) uses factory-built modularity and DOE site permits at INL to target late-2027/early-2028 first deployment, bypassing traditional 2-step NRC process via combined license; Meta's 1.2 GW prepayment at ex-Portsmouth site accelerates via fuel recycling, but HALEU supply and novel coolant limit scale until 2030s.[10][11]
- NuScale's 77 MWe VOYGR (NRC-approved 2025) pivoted post-UAMPS flop (no bankruptcy); ENTRA1/TVA eyes 6 GW by 2030s, Romania FEED ongoing; manufacturing ramp but customer subscriptions lag.[12]
- Kairos molten-salt (140 MW) secured Google 500 MW fleet (first 2030); Hermes demo 2026 but commercialization 2030+; NRC pre-app favors but fuel/supply chain unproven.[13]
Capacity online: 0 GW 2026-2027 (demos only); <0.5 GW 2028+ (Oklo Aurora pilots); barriers: NRC COLA (2-4 yrs post-2026 apps), HALEU shortage (US ~900 kg/yr), $B-scale FOAK financing; new entrants need hyperscaler prepays (e.g., Meta/Oklo) and DOE NSDA to de-risk vs. NuScale's stalled US projects.
Natural Gas CC/Peaker Additions
Hyperscalers bypass 5-10 yr grid queues via behind-the-meter (BTM) CC/gas turbines, with turbine backlogs (GE/MHI to 2028-30) driving 195% price hikes; Texas leads with 80 GW pipeline (40 GW data center direct), enabling xAI/Oracle/Crusoe campuses online in 1-2 yrs vs. grid's 4+.[14][15]
- EIA: 3.3 GW CC 2026 (half under construction), 3.3 GW 2027, 10.6 GW 2028; BTM dominates (e.g., Google/Crusoe 933 MW Texas, Meta 7.46 GW Louisiana).[16]
- Peakers/simple cycle for peaks; total gas dev ~252 GW, $416B capex.
Capacity online: ~3-7 GW 2026 (CC+peaker), 6-10 GW 2027, 15+ GW 2028+; barriers: turbine orders (6 yrs delivery), permitting (local emissions), financing ($/kW rise); competitors gain via BTM (e.g., Crusoe) but face ratepayer pushback without hyperscaler offtakes.
Utility-Scale/BTM Battery Storage
Batteries enable fast-track grid access (18-24 mo deploy) via load-shift/flexibility, with EIA forecasting 24 GW utility-scale 2026 (Texas/California lead); BTM pairs solar (20-40% ELCC) for data centers, but 4-hr duration limits baseload AI.[17]
- Utility: 24 GW 2026 to 67 GW Q1 2027; BTM: 15 GW by 2030 (83% data centers), hybrids boost ELCC to 50%+.[18]
Capacity online: 20-25 GW 2026, 30-40 GW 2027, 50+ GW 2028+ (nameplate; effective ~50% ELCC); barriers: lithium supply, 4-hr limits (needs hybrids), capex ($255-366/kWh 2026); entrants prioritize BTM for data centers to arbitrage peaks.
BTM Generation: Fuel Cells/Gas Turbines
Bloom's SOFC (60% efficient, 90-day deploy) scales modularly for BTM primaries, with Oracle 2.8 GW (2026+), AEP 1 GW ($2.65B), Brookfield $5B framework; 99.99% uptime at 30 MW/acre beats turbines on permitting/emissions.[19][20]
- Gas turbines BTM (e.g., xAI 422 MW Memphis) for speed, but 24-36 mo backlogs.
Capacity online: 1-2 GW fuel cells 2026 (Bloom ramp to 2 GW/yr), 2-4 GW 2027, 5+ GW 2028+; gas turbines 5-10 GW/yr BTM; barriers: Bloom scale-up (Fremont to 5 GW pot.), fuel (nat gas/hydrogen); hyperscaler deals (Equinix/CoreWeave) lock supply.
On-Site Nuclear/Gas at Hyperscaler Campuses
Hyperscalers co-locate BTM nuclear/gas for "bring your own power": AWS $20B Susquehanna nuclear campus (960 MW), Meta 6.6 GW (Vistra/Oklo/TerraPower 2032+), Google Intersect $4.75B for adj. gen; Gray Oak/FANCO bridges gas-to-SMR.[21][22]
- Fermi 11 GW Texas (nuclear/gas/solar) via turbine acquisitions.
Capacity online: Gas 5-10 GW 2026-27; nuclear 0 GW til 2028+ (SMR pilots); barriers: NRC/FERC for nuclear, turbine supply; winners secure PPAs (Microsoft/Constellation) and BTM to avoid queues.
Recent Findings Supplement (April 2026)
Nuclear Restarts
Constellation's Crane Clean Energy Center (formerly Three Mile Island Unit 1, 835 MW) secured a $1B DOE loan in November 2025 to support its Microsoft-backed restart, but grid interconnection delays via PJM—potentially until 2031 due to transmission projects and transformer shortages—threaten the 2027 target, forcing FERC waiver requests opposed by PJM's market monitor; water withdrawal permitting for 73M gallons/day from the Susquehanna River is under public comment until May 4, 2026, amid drought concerns.[1][2][3]
- No 2026 online capacity; 2027 at risk (H2 target); 2028+ viable if waivers granted, with full Microsoft 20-year offtake.
- Barriers: Transmission queues (delayed to 2030+), FERC/PJM waivers, Susquehanna River Basin Commission water approval, NRC public meetings ongoing.[4][5]
- Competitors face similar grid bottlenecks, amplifying value of existing nuclear; hyperscalers like Microsoft prioritize restarts for firm power despite premiums (~$110-115/MWh vs. market).
Holtec's Palisades (800 MW) restarted decommissioning status with NRC approval for fuel loading, but slipped from end-2025 to early 2026 (possibly late March) due to steam generator upgrades and inspections; $1.52B DOE loan supports, with NRC extending license renewal filing to March 2028.[6][7]
- ~800 MW possible early 2026; no confirmed hyperscaler offtake yet, but co-op PPAs (e.g., Hoosier/Wolverine) via New ERA funding.
- Barriers: Component recertification, legal challenges; SMR-300 add-ons (600 MW total) target early 2030s post mid-2027 Part 2 permit.[8]
- First-ever full restart sets precedent, but delays highlight supply chain risks; entrants need DOE loans and co-op deals to compete.
NextEra's Duane Arnold (~615 MW) advanced via Google's October 2025 25-year PPA (majority output) and NRC relicensing filing; DOE monitors as restart candidate, targeting early 2029.[9][10]
- No 2026-2027 capacity; 2028+ (~615 MW).
- Barriers: Ownership transfer tied to PPA, NEPA review; no recent permitting updates.
- Google's commitment de-risks economics, but slower than Palisades; new entrants lag without hyperscaler anchors.
SMR/Microreactor Commercialization
Oklo's Aurora (15-50 MW units) gained Meta partnership for 1.2 GW Ohio campus (first ~2030), plus Switch MPA (up to 12 GW by 2044); DOE pilot accelerates Idaho test (2027-2028), with hyperscaler pipeline >14 GW but non-binding.[11][12]
- Negligible <2028; pilots 2028+ scaling to hundreds MW via data center co-location.
- Barriers: Fuel allocation, site permitting; no firm GW-scale PPAs yet.
- Meta/Equinix demand favors micro-scale for edge sites, but execution risks high vs. restarts.
NuScale (77 MW modules) pursuing 6 GW ENTRA1/TVA deal (no binding PPA/timeline) and RoPower FID (late 2026/early 2027); no bankruptcy mentions, NRC-certified but Idaho project canceled pre-2025.[13]
- No firm 2026-2028 GW; potential RoPower Phase 2 revenue 2027+.
- Barriers: FID delays, no hyperscaler offtakes announced.
- Lags Oklo on data center momentum; needs PPAs to compete.
Kairos/others: Sparse post-Oct 2025 updates; Google's 500 MW rolling (2030-2035) unchanged.[14]
- 2028+ pilots; supply chain/regulatory hurdles persist.
- Hyperscalers co-develop, but timelines >2030 limit near-term entry.
Natural Gas CC/Peaker Additions
Proposals surged 300% to 159-252 GW (AI-driven), but turbine lead times (5+ years) limit to ~5 GW 2025 actuals; 19 GW equipment available by 2028, RWE eyes 9 GW peakers by 2031 (MISO/PJM/ERCOT), We Energies 2 plants (2028-2029).[15][16]
- <10 GW 2026-2027 (under construction ~30 GW total); 2028+ ramps to 40+ GW/year.
- Barriers: Turbine backlogs (easing 2027), regulators reject cost-shifting (e.g., data centers pay 100%).
- Fastest baseload/peaking for hyperscalers, but emissions scrutiny favors BTM.
Utility/Behind-the-Meter Battery Storage
Record 15 GW utility-scale added 2025; 24 GW projected 2026 (TX/CA/AZ lead), cumulative 67 GW by Q1 2027; BTM surges for interconnection acceleration (e.g., Aligned 31 MW), GW Ranch 1.8 GW (2027 first power).[17][18]
- 24 GW utility 2026, +BTM ~10-15 GW; 2027-2028 doubles.
- Barriers: Tariffs/supply restructuring dip growth 2026-2027.
- Enables renewables but non-firm; hyperscalers pair with gas for viability.
Behind-the-Meter Generation (Fuel Cells/Gas Turbines)
Bloom Energy's fuel cells exploded: Oracle 2.8 GW MSA (1.2 GW deploying), AEP $2.65B/1 GW, Brookfield $5B global; hyperscalers favor for 55-day deploy vs. grid years.[19][20]
- 1-2 GW 2026; 2027-2028 multi-GW via backlog.
- Barriers: Fuel costs ($150-180/MWh); nat gas efficient/low-water.
- Fastest BTM for Oracle/Meta; $7.65B deals signal hyperscaler shift to on-site.
On-Site Nuclear/Gas at Hyperscaler Campuses
Meta 6.6 GW nuclear PPAs (Vistra uprates 2.1 GW extend, Oklo/TerraPower 2030+); Google Duane Arnold, Crusoe gas talks; Microsoft Nscale 1.4 GW off-grid gas (2028); Oracle Jupiter 2.45 GW Bloom fuel cells; SoftBank 10 GW Ohio (9.2 GW gas).[21][11]
- Gas/fuel cells 2026-2027 GW-scale; nuclear 2028+.
- Barriers: State laws bypass grid (WV), but emissions/permitting.
- BTM/on-site dominates (38% data centers by 2030); hyperscalers lock PPAs early, sidelining grid-tied entrants.
Report 5 Research how Microsoft, Meta, AWS, Google, and Oracle are publicly pursuing strategies to secure power outside or alongside the traditional grid, including: nuclear PPAs (e.g., Microsoft-Constellation Three Mile Island deal, Google-Kairos Power agreement), on-site generation projects, co-location at generation assets, custom substation and transmission investments, and behind-the-meter gas turbine or fuel cell deployments. Identify publicly available estimates of what share of hyperscaler load growth is expected to be served behind-the-meter vs. through the grid by 2028–2030. Assess which strategies have moved beyond announcement to construction or operation, and what structural advantages or risks each approach carries relative to grid-connected alternatives.
Nuclear Power Purchase Agreements (PPAs) and Restarts
Microsoft unlocked a dormant nuclear asset by committing to a 20-year PPA for the full 835 MW output of Constellation's Three Mile Island Unit 1 (rebranded Crane Clean Energy Center), providing dedicated carbon-free baseload power to its PJM-region data centers without competing in wholesale markets; this corporate anchor revenue finances the $1.6 billion restart, including turbine/generator upgrades, while federal loan guarantees de-risk the refurbishment.[1][2]
- As of early 2026, the site is 80% staffed with 500+ workers on technical refurbishments; NRC safety reviews ongoing, targeting 2027 online (accelerated from 2028).[2][3]
- $1 billion DOE loan approved November 2025 covers most costs, first tranche Q1 2026.[4]
AWS (Amazon) shifted from rejected behind-the-meter co-location to a front-of-the-meter 1.92 GW PPA with Talen Energy's adjacent Susquehanna nuclear plant (2.5 GW total), routing power through PJM grid and PPL transmission to avoid FERC blocks on direct ties; this hybrid ensures priority-like access while adding net-new capacity to the grid.[5][6]
- AWS acquired the 960 MW Cumulus campus outright in 2024 for $650 million; existing 300 MW co-location transitions Spring 2026 post-refueling/transmission upgrades, ramping to full by 2032.[7]
- FERC rejected expanded behind-the-meter in late 2024/early 2025 over grid impact concerns.[8]
Meta secured immediate baseload via 20-year PPAs for 2.176 GW from Vistra's operating Perry/Davis-Besse (Ohio) plants plus 433 MW uprates across those and Beaver Valley (PA)—the largest corporate-backed nuclear uprates—while fronting development costs for SMRs; this dual-track stabilizes prices now and scales capacity later for its 1 GW+ Prometheus Ohio supercluster.[9][10]
- Purchases start late 2026 (existing capacity), uprates early 2030s; prior 1.1 GW Constellation Clinton (IL) PPA begins 2027.[11]
- Oklo/TerraPower deals (1.2 GW/2.8 GW) in pre-construction/site characterization (Ohio/Pike County), targeting 2030+.[9]
Implications for competitors/new entrants: Restarted plants like Crane offer the fastest path (1-2 years vs. 5-10 for greenfield), but require hyperscaler-scale contracts ($1B+) to justify; FERC's co-location skepticism favors grid-tied PPAs, raising barriers for smaller players without lobbying muscle.
Small Modular Reactor (SMR) Deployments
Google's master agreement with Kairos Power commits to a 500 MW fleet (6-7 molten-salt SMRs) by 2035, with Hermes 2 (50 MW) as first commercial unit feeding TVA grid for TN/AL data centers; iterative demos (Hermes 1 non-power) validate factory-built scalability, bypassing custom large-reactor overruns via off-site module assembly.[12][13]
- Ground broken April 2026 on Hermes 2 (Oak Ridge, TN) post-NRC permit; operations 2030; Hermes 1 construction extended to 2029 amid first-of-kind delays.[14]
- TVA's first advanced reactor PPA accelerates commercialization.[15]
Oracle announced permits for three SMRs to self-power a 1 GW data center but pivoted to near-term behind-the-meter; no SMR construction underway, remaining in design/planning as fuel/supply chain hurdles persist.[16]
Implications for competitors/new entrants: SMRs promise 3-5 year timelines vs. 10+ for traditional nuclear, but demos like Hermes show delays; Google's TVA bridge reveals utilities' role in de-risking, favoring partners with demo sites over pure announcements.
Behind-the-Meter Gas Turbines and Fuel Cells
Oracle leads aggressive off-grid via layered deployments: 2.3 GW VoltaGrid modular gas (INNIO Jenbacher engines) across Texas, 1 GW additional West Texas gas for Vantage campuses, and up to 2.8 GW Bloom solid-oxide fuel cells (expanding prior deal); fuel cells replace turbines at Project Jupiter (NM, 2.45 GW microgrid), cutting NOx 92% and water use vs. combustion while enabling 90-day installs.[17][18]
- Deployments operational/rapid-scale (e.g., Bloom full DC in 90 days); gas fleets backed by Energy Transfer pipelines.[19]
Meta bridges to nuclear with on-site gas: $3.2B 2 GW Hyperion (LA) combined-cycle and 700 MW Ohio plant (expanded from 400 MW), Williams pipeline integration.[20]
Implications for competitors/new entrants: BTM gas/fuel cells deploy in months (vs. 5-7 year grid queues), but emissions scrutiny and fuel lock-in risk stranded assets; Oracle's microgrid pivot shows fuel cells' edge for water-scarce/AI sites.
Co-Location, Substations, and Transmission Investments
AWS exemplifies co-location by acquiring Cumulus adjacent to Susquehanna, initially behind-the-meter (FERC-blocked expansion) now front-of-meter post-2026 upgrades; Meta/Vistra uprates add 433 MW via equipment swaps, no new lines needed.[5]
Limited public substation investments; hyperscalers pledge grid funding (e.g., White House March 2026: full data center energy costs, no ratepayer hikes), but focus on BTM avoids them.[19]
Implications for competitors/new entrants: Co-location cuts transmission losses (5-10% savings) but invites FERC/utility pushback; BTM sidesteps but forfeits grid ancillary services.
Projected Behind-the-Meter Share of Hyperscaler Load Growth
Goldman Sachs estimates U.S. power demand CAGR hits 3.2% through 2030 (grid 2.8 pp + BTM 0.5 pp), with data centers driving 2 pp of grid growth but much incremental load met via BTM gas; Bloom forecasts data centers' on-site generation rises to 38% by 2030 (from 13% 2025), with fully off-grid facilities surging 27x to 27%.[21][22]
- BTM primarily gas now (75% of 56 GW pipeline), fuel cells scaling; hyperscalers prefer grid long-term but BTM bridges 2026-2030 queues.[21]
Implications for competitors/new entrants: BTM captures 15-20% of 2028-2030 growth (my inference from Goldman/Bloom splits), enabling 2-3x faster rollout but exposing to fuel volatility/emissions regs; grid-tied needs $50B+ capex commitments to influence utilities.
Structural Advantages and Risks
| Strategy | Advantages vs. Grid | Risks vs. Grid | Progress Beyond Announcement |
|---|---|---|---|
| Nuclear PPAs/Restarts | Firm 24/7 carbon-free (92% capacity factor vs. 35% solar); price stability 20+ years.[1] | 2-5 year delays; regulatory (e.g., FERC co-lo blocks).[8] | Microsoft/AWS/Meta: Refurbs/uprates in planning/exec (2026-27 start). |
| SMRs | Modular/factory build (3-5 yr vs. 10+); scalable to GW campuses.[12] | Unproven supply chain/fuel; demos delayed (Hermes 1 to 2029).[14] | Google: Hermes 2 ground broken 2026; Oracle: Planning only. |
| BTM Gas/Fuel Cells | 90-day deploy; no queues; microgrid resilience.[17] | Gas price volatility; NOx/water regs; stranded if nuclear scales.[19] | Oracle/Meta: GW-scale operational 2026. |
Overall for entrants: Hyperscalers' $10B+ nuclear bets + BTM war-chests create moats via first-mover supply (e.g., Bloom backlogs); risks cluster on regulation/fuel, favoring diversified hybrids. Confidence: High on near-term BTM (deployed), medium on nuclear (demos progressing but historical overruns ~50%).[21]
Recent Findings Supplement (April 2026)
Meta's Nuclear Expansion: Uprates and SMR Funding Accelerate PJM Capacity Without New Builds
Meta signed landmark nuclear deals in January 2026 totaling up to 6.6 GW by 2035, including 2.1 GW from existing Vistra plants (Perry/Davis-Besse in Ohio, Beaver Valley in Pennsylvania) via 20-year PPAs, plus 433 MW from the largest corporate-supported nuclear uprates in U.S. history (online early 2030s), 690 MW from two TerraPower Natrium SMRs (2032), and 1.2 GW from Oklo's Ohio power campus (prepayments for fuel/site development, first Aurora units by 2030).[1][2][3][4]
- Builds on June 2025 Constellation PPA for full 1.1 GW output from Clinton plant (online June 2027).
- Uprates extend plant life 20 years; Oklo/Meta prepayments de-risk Phase 1 (150 MW).
- Power feeds PJM grid for Meta's New Albany, Ohio AI supercluster (Prometheus, 1 GW online 2026).
Implications for competitors: Meta's hybrid (existing + new SMR) locks baseload at grid scale faster/cheaper than greenfield SMRs (e.g., Google's Kairos Hermes 2 groundbreaking April 2026 for 50 MW by 2030), but ties to PJM expose to rate hikes from data center load (94% of PJM peak growth). New entrants gain from uprate precedent but face fuel/supply risks vs. Oracle's fuel cells.
Oracle's Fuel Cell Pivot: 2.45 GW Microgrid Replaces Gas at Project Jupiter
Oracle expanded its Bloom Energy deal to 2.8 GW (1.2 GW initial deploying now) in April 2026, redesigning New Mexico's Project Jupiter (Doña Ana County) as a single 2.45 GW fuel cell microgrid—ditching gas turbines/diesel for solid oxide cells that cut NOx 92%, water use 99%, and noise, with construction on schedule (4,000 jobs).[5][6][7]
- 55-day prior deployment proves speed (vs. 5-7 year grids); natural gas/hydrogen fuel enables dispatchable 24/7 power.
- ERCOT queues (5-7 years) drove BTM shift; cells bridge to SMRs.
Implications for competitors: Oracle's microgrid sets BTM benchmark—faster/lower-impact than grid (avoids queues/upgrades), but gas emissions risk net-zero goals (vs. Meta's nuclear). Grid-tied hyperscalers like AWS face higher costs; BTM favors cash-rich players but exposes to fuel volatility.
Google and AWS Advance Co-Location: Land-Ready Sites Minimize Grid Strain
Google inked February 2026 20-year PPAs with AES for co-located clean generation at Wilbarger County, Texas data center (construction underway, land/interconnect secured); air-cooled to slash water, part of 7.8+ GW Texas adds.[8][9] AWS proposed April 2026 Calvert Cliffs (MD) campus next to Constellation nuclear (preliminary reviews done, no plans filed).[10]
- Google's "power-first" uses AES renewables (e.g., 545 MW Crescent solar by 2028) adjacent to DC.
- AWS eyes 2,000 acres/8 buildings for nuclear proximity.
Implications for competitors: Co-location de-risks via shared infrastructure (faster than standalone), but FERC/PJM rules needed for netting (Dec 2025 order mandates PJM revisions by Feb 2026).[11] Meta/Google lead; others risk grid delays (e.g., 4-7 years).
Restart Progress Stalls on Transmission: TMI Loan, But Delays Persist
Constellation secured November 2025 $1B DOE loan for Crane (ex-Three Mile Island Unit 1) restart (Microsoft 20-year PPA, 835 MW by 2027), with inspections on track but April 2026 transmission delays pushing timeline.[12][13][14] Google/NextEra Duane Arnold (IA) restart funded (online 2029, 600 MW to grid).[15]
- No operational shifts; uprates (Meta/Vistra) first "beyond announcement."
Implications for competitors: Restarts cheaper/faster than SMRs but grid-tied (PJM risks); BTM (Oracle) avoids. New FERC rules (June 2026 target) could standardize co-lo/BTM.[16]
BTM Load Share Forecasts: 25-33% Incremental Growth Off-Grid by 2030
RAND (June 2025, cited 2026): 49 GW BTM net capacity by 2030 (vs. 33 GW FTM), ~50% of 82 GW adds; plans total 149 GW BTM/151 GW FTM nameplate.[17] Bloom (Jan/Mar 2026): 33% data centers 100% onsite by 2030 (up from 10% forecast); 38% using onsite (from 13%).[18][19][20] McKinsey/S&P/Latitude (ongoing): 25-33% incremental DC demand BTM thru 2030; 48 GW queued (90% 2025 announces).[21]
- EPRI: DCs 9-17% U.S. power by 2030; gas dominant BTM.
Implications for competitors: BTM surges (Texas 30% market by 2028) bypass queues but forgoes grid revenue/ancillary; hyperscalers like Oracle win speed, grids lose (FERC safeguards cost-shift). Grid entrants need flex (e.g., Google DCFlex).
Report 6 Research the strongest publicly available counterarguments to the thesis that AI power demand will create severe, sustained grid constraints through 2030. Specifically investigate: (1) GPU and silicon efficiency improvement curves (e.g., inference efficiency gains from custom ASICs, smaller models, sparsity techniques) and whether compute demand per AI task is falling faster than demand is growing; (2) historical precedents where anticipated power demand surges did not materialize (e.g., early 2000s data center buildout, crypto mining cycles); (3) evidence that AI training compute may be plateauing or shifting to more efficient inference workloads; (4) accelerated transmission permitting and build scenarios under FERC Order 1920 or potential federal siting authority; (5) demand response and flexible load programs that could absorb data center load without new generation; and (6) any analyst reports or grid operator forecasts that project AI demand landing at the low end of the range. Produce a structured assessment of which disconfirming factors are most likely to materially reduce constraint severity and on what timeline.
1. Silicon Efficiency: ASICs and Techniques Outpace Raw Compute Growth, Shrinking Power per AI Token
Nvidia's Blackwell GPUs (B200) deliver up to 2.9x inference throughput per GPU versus H100s via FP4 precision and structured sparsity, enabling 70B-parameter models at $0.047/M tokens on spot instances—translating to 70-80% lower power per token for common workloads—while custom ASICs like Groq LPUs achieve 10-20x faster tokens/second than GPUs at sub-10ms latency by specializing matrix operations and eliminating general-purpose overhead.[1][2] This mechanism—precision reduction (FP16 to FP4), sparsity (N:M patterns skipping 50%+ zeros), and ASIC fixed-function units—collapses compute demand per task faster than AI usage scales, as inference (90%+ of future workloads) dominates training.[3]
- B200 FP4 sparsity hits 18,000 TFLOPS/GPU, 2.9x H200 on Llama 70B (MLPerf v5.1, Sep 2025).[1]
- Groq/SambaNova ASICs: 750-2000 tokens/sec on Llama 3.1 vs. Nvidia's 72-257, with 5-20x perf/watt.[2]
- Epoch AI: LLM inference costs fell 9-900x/year by task; fixed-performance inference prices drop 40x/year mid-range.[4]
Implications for entrants: Efficiency moats favor incumbents with data (e.g., Shopify's sales-derived lending), but open-source sparsity/quantization tools democratize 4x model shrinks; compete via edge ASICs for low-latency inference, viable by 2027 as inference surges to 40%+ of data center power.[5]
2. Workload Shift: Inference Efficiency Boom Overshadows Training, with No Plateau Yet but Diminishing Returns Emerging
AI compute pivots from training (one-off, bursty) to inference (continuous, 35% CAGR to 93 GW by 2030), where smaller quantized models (e.g., 4-bit Llama) run 1.5-4x faster on less power via techniques like KV-cache compression and speculative decoding—reducing effective demand as usage explodes without proportional power hikes.[5][6] Training compute grows 4-5x/year (Epoch AI), but R&D experiments consume 70-90% of total (not final runs), and post-training optimizations (test-time compute) yield gains without full retrains; no full plateau, but data exhaustion (~500T tokens indexed web) looms by 2030.[4]
- Inference to dominate: 40%+ data center power by 2030 (McKinsey), vs. training's 22% CAGR.[5]
- Efficiency: Quantization/pruning shrinks models 4x, speeds 1.3-1.7x; Google's Dynamo boosts inference 5x via memory optimizations.[6][7]
- Training trends: 4.5x/year growth continues to 2030, but open models hit 1026 FLOP by late 2025.[8]
Implications for entrants: Training's high barrier (GW-scale clusters) favors hyperscalers; inference's edge/ASIC shift opens niches for flexible providers—target metro sites for low-latency by 2027, leveraging 90%+ GPU utilization via batching.[6]
3. Historical Overforecasts: Data Centers and Crypto Show Demand Surges Flatten via Efficiency and Volatility
Early 2000s data center boom exploded 90% electricity use 2000-2005 amid dot-com, but 2010-2018 global use flatlined despite soaring compute demand, as virtualization/PUE drops (from 2.0+ to 1.2) offset growth—echoing today's fears, where pre-2020 projections missed efficiency.[9] Crypto mining: 2017-18 hype forecasted massive surges, but post-bear markets and China's 2021 ban saw U.S. use stabilize at 0.6-2.3% national electricity (EIA), with operators pivoting to AI without sustained grid strain.[10]
- 2000s: Use rose linearly 1.6%/year post-peak, vs. exponential fears.[11]
- Crypto: Volatile; post-halving shutdowns cut inefficient ops, demand flexible.[12]
Implications for entrants: Utilities overbuild risks stranding assets if AI hype fades; new players secure flexible contracts (e.g., curtailment clauses) to hedge, mirroring crypto's pivot.
4. Grid Buildout Acceleration: FERC 1920 Mandates 20-Year Planning, Unlocking Transmission via Scenarios and GETs
FERC Order 1920 (May 2024) forces 20-year scenario planning with seven benefits metrics (e.g., reliability deferral, production savings), evaluating grid-enhancing tech (GETs like dynamic line ratings, up 40-80% capacity) and high-performance conductors before new lines—compliance filings due 2026, first projects 2028-2033, preempting AI bottlenecks.[13][14]
- Requires three diverse scenarios, state input, ATT evaluation (DLR/APFC/switching).[13]
- Builds on Order 1000; reviews start 2026.[15]
Implications for entrants: Speeds interconnections for flexible loads; co-locate with generators under PJM rules (FERC Feb 2025), but fund upgrades—viable 2027+ for AI inference.
5. Demand Flexibility: Software Shifts AI Loads, Absorbing 18-55% Peaks Without New Gen
Google's 1GW demand response (2026) with TVA/Entergy shifts ML training (deferrable) off-peak, cutting peaks 25% for hours via job queuing; Microsoft power-caps coordinate scheduling/infra; studies show 18-55% flexibility in regulation reserves/emergencies, unlocking 100GW data centers sans new plants if <50 hours/year curtailed.[16][17]
- Google: ML workloads shifted, 1GW capacity (TVA, Indiana Michigan, etc.).[16]
- Duke: 100GW via modest peaks; Emerald AI: 25% cut 3hrs.[18]
Implications for entrants: Mandate DR in leases; hyperscalers lead—partner for "flex credits" reducing interconnection queues by 2027.
6. Low-End Forecasts: EIA/NERC Project Manageable Growth at 1-2%/Year Overall, with Risks Overstated
EIA AEO 2026: Electricity demand grows 0.9-1.6%/year to 2050 (post-2.1% recent), data centers key but servers to 818 billion kWh high-case (16x 2020, ~8-12% total by 2030); NERC LTRA 2025: 224GW summer peak +10yrs (conservative per Grid Strategies), but critiques note 2/3 data center risk from chips/finance.[19][20]
- EIA: Commercial (data centers) drives, but efficiency tempers.[21]
- NERC: Data centers ~55% growth, but Grid: Overstated 90->~65GW.[20]
Implications for entrants: Low-end (1%/year) viable with flexibility; monitor EIA for policy.
Assessment of Disconfirming Factors
Most Likely to Reduce Severity (High Confidence, 2027-2030): Efficiency (1) and flexibility (5)—mechanisms proven, scaling fast; could halve effective demand growth. Medium (2028+): Inference shift (2), transmission (4). Low: History (3), plateau (no evidence), low forecasts (EIA conservative but risks noted). Timeline: Material relief by 2028 via software/hardware; full by 2030 if DR/GETs adopt. Additional research: Chip forecasts, DR pilots.
Recent Findings Supplement (April 2026)
GPU and Silicon Efficiency Improvements
NVIDIA's Vera Rubin platform (announced January 2026) integrates advanced NVFP4 precision and co-designed CPU-GPU architectures to deliver 10x higher inference throughput and 10x lower cost per token versus prior generations, enabling sustained efficiency gains that could outpace some task demand growth through hardware-software fusion—though Jevons paradox risks offsetting this via expanded AI usage.[1][2]
- Rubin doubles NVLink-C2C bandwidth and triples memory capacity, reducing power per inference operation; Blackwell already claims 25x inference efficiency over Hopper.[3][4]
- ArXiv papers (e.g., ZipServ, March 2026) show lossless compression yielding 1.22x end-to-end inference speedup and 30% model size reduction on GPUs like RTX5090, with fused kernels cutting memory bandwidth needs.[5]
Implications for competition/entry: New entrants gain via inference-optimized ASICs (e.g., Trainium, TPUs), but incumbents like NVIDIA hold data moats; efficiency could cap power per task at 2026 levels if sparsity/quantization scales, but aggregate demand likely rises (my inference from Jevons mentions, no direct source).
Shift to Inference Workloads
AI compute is shifting from training (predictable bursts) to inference (bursty, user-driven), projected to comprise 2/3 of workloads by 2026 and 65%+ by 2029, potentially plateauing training FLOPs as models mature while inference efficiency improves via distillation/quantization—reducing overall power intensity if adoption doesn't explode via agents.[6]
- Inference market grows to $255B by 2030 (19% CAGR); 70% of data center demand from inference by 2030, with specialized chips favoring lower-power edge deployment.[6]
- Training-like baseload shifts to inference variability, enabling flexibility (e.g., non-urgent tasks rescheduled).[7]
Implications for competition/entry: Favors hyperscalers with global inference fleets (Google, AWS); smaller players enter via edge inference, but power savings hinge on no agentic explosion (low confidence, training plateau inferred from shift data).
Demand Response and Flexible Loads
Google integrated 1GW demand response into U.S. data center contracts (March 2026) by shifting ML workloads during peaks, turning inflexible loads into "ghost batteries" that unlock grid headroom—Duke University models show U.S. grids absorbing 76-100GW new flexible data center load with <1% curtailment (e.g., <50 hours/year), avoiding new generation for 5-10 years.[8][9]
- AI data centers offer 18-55% flexibility vs. average power via power capping/rescheduling; EPRI Flex MOSAIC classifies for grid ops.[10][11]
- Utilities/data centers collaborate on curtailment, saving 6-54% costs while stabilizing grids.[12]
Implications for competition/entry: Most viable near-term mitigator (2026-2028); entrants partner with utilities for DR incentives, but requires software for workload shifting—highest potential to reduce severity without capex.
Accelerated Transmission via FERC Order 1920
FERC Order 1920 compliance filings began December 2025 (PJM/CAISO), mandating 20-year planning with state input and three selection criteria (performance, cost, maximization)—early progress in PJM/SPP/MISO unlocks GW-scale transmission for data centers by 2028-2030, complemented by colocation rules (Dec 2025) favoring on-site gas/nuclear.[13][14]
- FERC directs large-load reforms by June 2026; PJM adds transmission options for data center-generator colocation.[15][16]
Implications for competition/entry: Timeline too slow for 2026-2027 crunch (years away); favors regions with early filings (PJM), but federal siting dormant—medium impact post-2028.
Grid Forecasts and Overstated Demand Risks
EIA AEO2026 (April 2026) projects U.S. electricity growth at 0.9-1.6% annually to 2050 (low end baseline), with data centers driving commercial sector but EVs/data centers only 10-25% of total demand; Enverus 2026 Outlook calls AI boom "overstated" as behind-the-meter/on-site gen mutes grid impact.[17][18][19]
- High-demand case hits 818 TWh data centers by 2050 (16x 2020), but baseline assumes efficiency offsets some growth.[18]
Implications for competition/entry: Low-end scenarios (efficiency/DR) enable entry without crisis; no grid operator "no severe constraints" quote, but Duke/EPRI headroom suggests mitigable (estimated from models).
Historical Precedents (Weak Recent Evidence)
Early 2000s data center/video fears kept demand flat via efficiency (cited Nov 2025); BTC mining warnings (2017-2021) didn't overwhelm grids (0.8% global electricity), as economics capped growth—AI differs in inflexibility, but miners pivot to flexible AI hosting.[20][4]
- No 2026-specific AI parallels; crypto stabilized predictably.[20]
Implications for competition/entry: Least compelling; AI's scale/steadiness worse than crypto, low confidence in repeat.
Strongest Disconfirmers (Likelihood/Timeline): Flexible loads/DR (high, 2026 via Google/Duke); inference shift+efficiency (medium-high, 2026-2028); transmission reforms (medium, 2028+). Combined, could halve effective constraints by 2028 (inferred, no direct source); low-end forecasts reinforce. Additional research on NERC 2026 LTRA needed for confidence.