Powering the AI Boom: Where the Grid Breaks First (2026-2030)

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.