AI data-center electricity demand through 2030: the power crunch and where the consensus number breaks

Research question. Is the projected surge in AI-driven data-center electricity demand real and investable at the magnitude the consensus assumes — or is it a forecast likely to overshoot — and what is the defensible demand range an investor should underwrite through 2030?

Executive Brief: The Decision Question and the Defensible Range

The central investment question is simple to state and hard to answer: will AI-driven data-center electricity demand materialize at the magnitudes headline forecasts suggest, or is the consensus number structurally inflated? This report's answer, delivered before the evidence is marshaled, is both. The surge is directionally real. The high end of the published range is not investable as stated.

Start with what the available evidence confirms. U.S. data centers consumed approximately 176 TWh in 2023, representing 4.4% of total U.S. electricity consumption, according to a 2024 report on U.S. data center energy usage.¹ That growth accelerated sharply: LBNL documents a compound annual growth rate of approximately 7% from 2014 to 2018, rising to 18% between 2018 and 2023.¹ The acceleration is real. It is not projection; it is measurement.

From that confirmed floor, forecasts diverge sharply. LBNL's scenario range for 2028 runs from approximately 325 TWh to 580 TWh, representing 6.7% to 12.0% of forecasted U.S. electricity consumption.¹ EPRI's February 2026 state-level pipeline analysis projects 384 to 793 TWh by 2030, having revised upward 60% from its 2024 estimate following 18 months of accelerated AI infrastructure announcements.² Goldman Sachs, in its February 2026 update, implies approximately 810 TWh for the U.S. alone by 2030.² Globally, the IEA's base case projects consumption roughly doubling to around 945 TWh by 2030, growing at approximately 15% per year.³ One synthesis of these forecasts places a defensible U.S. base case at 550 to 650 TWh, representing 11 to 13% of U.S. total, by blending LBNL's equipment method with EPRI's medium scenario and McKinsey's workload approach.²

That range, not the headline figures, is the number a disciplined investor should underwrite. The spread between the LBNL floor and Goldman Sachs ceiling exceeds 2x. That spread is not statistical noise. It reflects genuine conflicts in operative assumptions. Three of those assumptions move the range more than all others combined.

The first is the efficiency trajectory of AI workloads. LBNL's bottom-up model incorporates GPU power draws at 60 to 80% of rated capacity and projects power usage effectiveness (PUE) declining to approximately 1.4 through liquid cooling and hyperscale adoption.² High-end forecasters assume inference workloads grow faster than efficiency improvements can offset. The IEA documents that accelerated servers are projected to grow at 30% annually in its base case, while conventional servers grow at 9%.³ Which trajectory dominates is not resolved by current data. A Jevons-paradox dynamic, where efficiency gains lower per-query costs and therefore expand total query volume, is credible and cannot be dismissed with the evidence available. The bull and bear cases on this question represent a genuine disagreement, not a resolvable measurement gap.

The second assumption is the pace of interconnection clearance. The evidence base suggests over 2,000 GW of generation and storage capacity currently sit in U.S. interconnection queues, representing more than the total existing U.S. generation fleet.⁴ That figure is from a single source; it is not independently corroborated here. LBNL's own methodology explicitly avoids overestimation caused by tracking load for projects that have not yet selected a power provider.¹ Grid interconnection timelines now routinely extend up to four years.⁵ Electrical capacity is governed by physical grid limitations and utility planning cycles that often extend multiple years, unlike land or capital.⁶ The practical consequence: announced capacity is not energized capacity. Project realization rates matter as much as headline announcements.

The third assumption is the ratio of speculative to committed capacity. EPRI's pipeline method aggregates operational, under-construction, and announced projects before applying scenario-based realization rates.² A benchmark from Grid Strategies suggests only approximately 65% of announced projects may materialize by 2030.² The LBNL report announced by DOE noted that data center load is projected to double or triple by 2028, but that framing—which emphasizes that clean energy and grid modernization can meet the growth—does not grapple with the gap between queue entries and energized megawatts.⁷

These three uncertainties are not merely unknown. Each is uncertain in a specific, falsifiable way. The efficiency trajectory will move when GPU shipment data and operational utilization figures become available at scale. Interconnection clearance pace will move when queue completion rates are tracked systematically against 2024 filings. The speculative-to-committed ratio will move when project withdrawal rates in the 2025 to 2026 cohort are observable. The range narrows as each resolves. It does not collapse to a point estimate before they do.

One methodological note is material here. The IEA's global projections cited in this report appear across multiple source documents that share a common upstream origin. Their agreement is one voice, not independent corroboration. Similarly, the DOE press release summarizing LBNL findings is a downstream document. The primary evidence for the LBNL scenarios rests on the 2024 LBNL report itself.¹ LBNL's own framing is explicit: many financial analytics firm models are proprietary or methodologically opaque, and historical utility demand forecasts have consistently overestimated both peak and average demand.¹

The investment posture this report defends follows from that discipline. The directional case for material load growth is well-grounded in measured data. The quantitative case for the high end of consensus is not. An investor who prices assets on 810 to 945 TWh scenarios is underwriting assumptions the available evidence does not support. An investor who treats the floor as the ceiling is ignoring a decade of accelerating growth that runs through primary sources. The defensible range sits between them, bounded above by physical constraint and below by confirmed trajectory, and it moves when the three assumptions do.

What Is Actually Drawing Power Today: The Verified Baseline

The empirical floor begins with a single authoritative number. Lawrence Berkeley National Laboratory's December 2024 report, mandated by Congress under the Energy Act of 2020, estimates U.S. data center electricity consumption reached 176 TWh in 2023, equal to approximately 4.4% of total U.S. electricity that year. The LBNL report itself is the primary source for this figure, derived through the laboratory's bottom-up methodology.¹ That 176 TWh number is the most defensible current baseline available: according to the Congressional Research Service, it derives from LBNL's bottom-up methodology and represents the closest public-record approximation to a measured floor for U.S. data center energy consumption in 2023.⁸ It is not a projection. It is not a utility forecast. It is the closest approximation to a measured floor that the public record currently supplies.

LBNL's methodology matters because it explicitly avoids a systematic bias that afflicts other approaches. The bottom-up model calculates total electricity use from an installed base of actual equipment, rather than tracking load requests from projects that have not yet selected a power provider. ¹ That design choice is not incidental. LBNL notes that historical utility demand forecasts have consistently overestimated both peak and average demand, and that many financial-analytics models underlying industry projections are proprietary or methodologically opaque. ¹ An investor treating sell-side estimates as equivalent to LBNL's equipment-based accounting is not comparing two equally weighted sources.

The global picture is thinner. No institution with LBNL's methodology has produced a peer-equivalent global estimate. What the public record offers is a range. One reference aggregates IEA and other sources to suggest global data centers currently consume somewhere between 1% and 2% of worldwide electricity, translating to roughly 300 to 400 TWh annually. The source projects that IEA data show global data center electricity consumption could double by 2030, potentially reaching 600–800 TWh annually, though the source carries no primary methodology of its own.⁴ On global current consumption, the honest answer is that the floor is uncertain within a range of a few hundred TWh, and no single institution has produced a comprehensive bottom-up global count comparable to the detailed national-level methodologies that exist for individual countries.⁹

The growth trajectory is less contested than the absolute level. LBNL documents a compound annual growth rate of approximately 7% for U.S. data centers from 2014 to 2018, accelerating to 18% between 2018 and 2023. That acceleration reflects the rapid growth in GPU-accelerated servers for AI applications entering the installed base.¹ LBNL notes that total data center energy demand more than doubled between 2017 and 2023, driven by rapid growth in accelerated servers. The Belfer Center report documents the downstream consequence: hyperscale cloud providers have dramatically scaled capital expenditures in response to surging AI infrastructure demand, signaling that growth is being backed by real investment commitments. ¹⁰ Capital commitment at that scale is corroborating evidence that demand growth is real, not simply projected.

Now the disaggregation problem. How much of that 176 TWh is attributable specifically to AI and accelerated compute, as distinct from general cloud and colocation? The CRS report, citing a 2024 EPRI analysis, states that AI consumed 10% to 20% of data center energy at the time of that estimate. ⁸ That is the only available figure from a named institutional source that attempts the split. It implies AI accounted for roughly 18 to 35 TWh of the 176 TWh U.S. total in 2023, but EPRI's methodology for reaching that range is not disclosed in available public summaries. The available record cannot resolve the AI-versus-general-cloud split with precision. This is a measurement gap, not merely an unknown: the data to make the disaggregation cleanly do not exist in public form, because neither hyperscalers nor colocation operators report workload-level power consumption. Every forward projection that purports to model AI-specific electricity demand inherits this gap directly.

Unsourced claims circulate alongside the verified figures, and they must be separated. The assertion that a single GPT-4 training run consumed approximately 50 GWh, equivalent to the annual consumption of about 40,000 U.S. households, appears in aggregator documents but is stated without evidence and cannot be traced to a primary disclosure. ⁴ It is not part of the empirical floor. Similarly, the CRS report notes that an April 2025 report estimated training a specific large AI model required a total power draw of 25.3 MW, with power required potentially doubling annually, but does not name the model or the issuing institution. ⁸ These claims are real data points about what is being asserted in the discourse. They are not verified figures on which an investment thesis should rest.

The construction pipeline adds a second layer of evidence. CRS reports that computing capacity under construction in North America at the end of 2024 reached a record-high 6,350 MW, more than double the figure from a year earlier. ¹¹ These figures are grounded in regulatory filings and pipeline data, not forward projections alone. They are pipeline counts from regulatory filings and permit records. They establish that the buildout driving the forward demand curve is physically underway.

Two conflicts in the available record require surfacing here, before forward projections are considered. First, whether utility load forecasts reliably represent demand that will materialize. Grid Strategies finds that the five-year forecast for utility peak load growth has increased by more than a factor of six over three years, from 24 GW to 166 GW, with data centers accounting for approximately 55% of that growth, or roughly 90 GW. ¹¹ But the same analysis notes that data center market analysts from multiple independent firms suggest demand growth is unlikely to exceed 60 to 65 GW through 2030, implying utility forecasts overstate data center-driven load growth by about 40%. ¹¹ The source acknowledges directly that institutional projections of data center electricity demand vary widely, with utility load forecasts implying data center growth of roughly 90 GW while independent market analysts place the ceiling closer to 60–65 GW through 2030—a disagreement it describes as complicating medium-to-long-term grid planning. ¹⁰ A spread of that magnitude is not statistical noise. It reflects genuine disagreement about which announced projects will reach energized capacity.

Second, what the history of data center forecasting implies for confidence in current estimates. One independent analysis found that among 258 data center energy estimates from 46 publications between 2007 and 2021, only two sources provided reliable estimates. The same source corroborates this pattern, noting that its own 2022 review of 258 data center energy estimates from 46 publications between 2007 and 2021 found only two sources of reliable estimates, pointing to systematic methodological defects in the historical literature.¹² The same source notes that institutional projections of data center electricity demand range from about 200 TWh to over 1,000 TWh by 2030, a spread the World Resources Institute attributes to deep methodological uncertainty across competing forecasting approaches. ^{10 1} The methodological record does not disqualify modern estimates. LBNL's equipment-based approach represents a genuine improvement over the query-cost extrapolations that dominated earlier forecasting. But it does argue for treating the 176 TWh 2023 baseline as a well-grounded floor, and the 2028 and 2030 scenarios that follow from it as ranges conditioned on assumptions, not as measurements.

One disclosure is necessary before any forward figure is cited. The proposition that global data center electricity consumption is projected to roughly double by 2030, potentially reaching 600 to 800 TWh, appears in multiple documents that share content behind different URLs. ⁴ This is one source, not independent corroboration. Likewise, the proposition that U.S. consumption will grow from approximately 4.4% of total electricity in 2023 to between 6.7% and 12% by 2028 appears in a Belfer Center document and an LBNL-derived summary that share the same upstream LBNL report. ¹⁰¹ The 325 to 580 TWh range for 2028 is an LBNL scenario output. It is the best available forward range for the U.S., but it is one institution's model, not a convergence of independent analyses. The forward projections that subsequent sections examine depart from this verified floor: 176 TWh in 2023, growing at an accelerating rate, with the AI-specific contribution unmeasured and every figure above the 2023 baseline an estimate.

The Forecast Landscape and Why It Splits: A Driver Decomposition

The verified baseline established in the preceding section — 176 TWh of U.S. data-center electricity consumption in 2023, representing 4.4% of total U.S. electricity use — is where consensus ends.¹ From that shared starting point, published 2030 projections splinter into a roughly 2x spread that no averaging exercise can resolve. The LBNL bottom-up equipment model projects 325 to 580 TWh by 2028, with 13–27% CAGR extrapolations pointing toward 450 to 850 TWh by 2030.² EPRI's February 2026 pipeline analysis, revised upward 60% from its 2024 estimate following 18 months of accelerated AI infrastructure announcements, spans 384 to 793 TWh in 2030 across its low, medium, and high realization scenarios.² Goldman Sachs updated its global forecast to a 220% growth scenario by 2030, implying roughly 810 TWh for the U.S. share.² At the outer extreme, the World Resources Institute documents a projection reaching over 1,050 TWh, a figure that would represent about a quarter of all U.S. electricity generation in 2023.¹³ Meanwhile, many estimates cluster between 300 and 400 TWh by 2030.¹³ That is not a range. It is a catalog of disagreements that happen to share a unit.

One editorial disclosure is warranted here: two sources cited in the prior section for U.S. data-center consumption growing from approximately 4.4% of total electricity in 2023 share identical underlying content behind different URLs. They represent one document, not independent corroboration. The baseline figure stands, but readers should not treat its repetition across sources as convergent confirmation.

Grid Strategies' successive load-forecast compilations add a further dimension. Over three years, utility five-year peak load growth forecasts submitted to FERC rose from 24 GW in 2022 to 166 GW in 2025 — more than a six-fold increase — with data centers accounting for roughly 55% of that growth, or approximately 90 GW.¹¹ Grid Strategies' 2024 analysis found that U.S. electricity demand could rise by 128 GW over the next five years, driven primarily by data centers and manufacturing growth concentrated in six regions of the country.¹⁴ These FERC-filed figures are not independent of the analyst projections above; they reflect the same underlying buildout pipeline reported through a different institutional channel. That overlap matters when weighing whether the convergence of utility filings and sell-side forecasts constitutes genuine independent confirmation or shared assumptions dressed in separate formats.

The spread is not noise. It is the output of five distinct operative drivers that different forecasters model with different priors. Decomposing those drivers makes the disagreement legible.

GPU and accelerator power-draw trajectory. The LBNL bottom-up model anchors its estimates in equipment shipments, with GPU power draws assumed at 60 to 80% of rated capacity.² This is a conservative utilization assumption. Pipeline-based forecasters, including EPRI, convert announced IT capacity into energy use through implicit load factors; their high scenario reaches 132 GW of IT capacity.² The practical consequence: forecasters who trust announced shipment data as a ceiling lean toward the lower half of the range; forecasters who extrapolate from current GPU allocation trends without a utilization haircut lean toward the upper half. Neither group has primary telemetry from deployed systems. The actual power draw of AI accelerators under operational workloads remains a coverage gap.

Utilization rates of installed capacity. This driver produces the widest per-unit variance. A data center's installed GPU capacity and its consumed electricity diverge by whatever utilization rate one assumes. Grid Strategies estimates that alternative benchmarks from market analysts — Cleanview, TD Cowen, Bloomberg NEF, McKinsey, S&P Global, Wood Mackenzie — suggest data-center demand growth is unlikely to exceed 60 to 65 GW through 2030, implying that utility forecasts of roughly 90 GW overstate demand by about 25 GW.¹¹ That 40% gap between gross announced capacity and likely realized load is almost entirely a utilization and project-completion argument. LBNL makes this explicit: its bottom-up method avoids overestimation by tracking installed equipment rather than load for projects that have not yet selected a power provider.¹ Forecasters who start from announcements and apply no realization discount are the optimistic prior here; those who apply completion-rate haircuts converge toward the lower range.

Model and inference efficiency gains. Here the forecasters divide into two clearly different camps, and the disagreement is genuine rather than methodological. LBNL's data shows that during the early and mid-2010s, the shift from on-premises infrastructure to cloud facilities enabled near-flat electricity consumption despite large service-volume growth.¹ Efficiency optimists project a similar dynamic for the AI era. The devsustainability.com analysis rejects this analogy directly: because major AI labs are already deploying into hyperscale facilities, "the biggest historical efficiency lever — moving workloads out of inefficient enterprise data centers and into modern cloud infrastructure — has already been pulled."¹² Future gains will have to come from better accelerators, model architectures, higher utilization, better scheduling, and inference optimization.¹² That same analysis also notes that AI workload patterns are shifting from single-query inference through reasoning models toward agents that decompose problems into many model calls, tool calls, retrieval steps, and verification loops — complicating any fixed efficiency assumption.¹²

The World Resources Institute points to an additional complication: efficiency gains may trigger a rebound effect that ultimately drives total demand higher despite lower power-per-task.¹³ Goldman Sachs' upper-range scenario implies exactly that outcome. The IEA, by contrast, has modeled AI-enabled grid optimization and industrial efficiency as potential offsets that could match or exceed incremental data-center load. However, the available evidence on that claim rests on a single modeling exercise, not independent lines of evidence, and the empirical question of whether AI-enabled grid optimization will be deployed at sufficient scale and speed remains unresolved.¹²

Cooling overhead and PUE evolution. Average data-center PUE has declined from approximately 2.0 in the early 2000s to 1.5 to 1.6 for typical facilities today, with leading hyperscale operators reporting fleet-wide PUE of 1.10 to 1.12.⁴ Cooling can account for 40% of a data center's energy usage, making it a prime target.¹³ LBNL's scenarios assume PUE declining to approximately 1.4 through liquid cooling and hyperscale facility shifts.² McKinsey's workload-driven forecast assumes PUE improvement to 1.1.² The difference between those two PUE assumptions, applied across hundreds of gigawatts of capacity, generates a material share of the terminal-year spread. Forecasters who project aggressive liquid immersion cooling adoption are the optimistic prior; those who treat enterprise-segment PUE improvement as slow and costly are the pessimistic prior. Primary perspectives from chip designers and semiconductor manufacturers on realistic power reduction timelines are absent from the available record, a gap that limits confidence in either camp's cooling assumptions through 2030.

Capacity-addition completion rates. The Luminix synthesis reports that only roughly 65% of announced data-center projects may materialize by 2030 per Grid Strategies benchmarks.² EPRI's high scenario applies a 30% early-planning realization rate on top of under-construction and advanced-planning capacity; its low scenario is materially more conservative.² The World Resources Institute documents the mechanism behind non-completion: utilities are receiving speculative and duplicate interconnection requests, including early-phase projects unlikely to be built and multiple requests for the same facility, generating phantom load that distorts forecasts.¹³ Grid Strategies observes that the power industry does not have a clear understanding of how much demand will actually come from data centers.¹⁴ Forecasters who trust announced project pipelines at face value (EPRI high scenario, BCG) carry the optimistic prior on this driver. Those who apply realization discounts informed by historical interconnection completion rates carry the pessimistic prior.

These five drivers are not independent. They compound. A forecast that combines aggressive GPU power-draw assumptions, high utilization rates, no PUE improvement, and full project realization will produce a BCG-type outcome above 900 TWh. A forecast that applies conservative shipment data, 60 to 65% realization, and PUE declining to 1.2 will produce an LBNL-type outcome nearer 450 TWh. The divergence is not noise around a true mean. It reflects genuinely different analytical philosophies applied to genuinely uncertain inputs.

The Luminix synthesis, which positions itself as an investor-facing analysis and should be read in that promotional framing, proposes a defensible base case of 550 to 650 TWh by 2030, blending EPRI's medium scenario, McKinsey's workload-driven approach, and LBNL's trend-adjusted high, under assumptions of hyperscaler dominance at 85% of servers, PUE of 1.2 to 1.4, and roughly even training-to-inference split.¹⁵ Its sensitivity analysis suggests potential variance of plus 20% if inference exceeds training, minus 15% if GPU efficiency and PUE below 1.2 are achieved, and plus-or-minus 10% from project realization risk.² Note that this synthesis and a closely related document from the same publisher appear to represent the same underlying analytical work presented under different URLs; they are one analytical position, not two independent convergences.

One further structural issue compounds the terminal-year spread. The baseline itself is not settled. The LBNL report establishes 176 TWh for 2023.¹ Other sources report the global baseline at roughly 300 to 400 TWh for the same period.⁴ Because 2030 projections express growth as multiples of or additions to current consumption, any divergence in the starting point propagates through the entire forecast horizon. An investor comparing a 580 TWh LBNL 2028 projection with a Goldman Sachs 810 TWh 2030 projection is not just confronting different growth assumptions; the forecasts may depart from different empirical floors. The terminal-year spread is therefore wider in effective confidence interval terms than the headline range implies.

Methodological rigor is asymmetric across the forecasting community. The devsustainability.com analysis examined 258 data-center energy estimates published between 2007 and 2021 and found that only two sources provided reliable estimates.¹² Many historical estimates employed weak extrapolation, taking energy cost per query, multiplying by user growth, and calling it a forecast.¹² LBNL explicitly distinguishes its bottom-up equipment methodology from the proprietary or methodologically opaque models underlying many financial analytics projections, and notes that historical utility demand forecasts have consistently overestimated both peak and average demand.¹ That institutional critique is itself a prior: LBNL's method is more transparent, but its sensitivity to GPU shipment assumptions and utilization-rate inputs means it is not assumption-free. The World Resources Institute similarly cautions that data-center electricity demand was overestimated in the 2000s, when projections of power shortages after explosive 2000-to-2009 growth proved wrong as efficiency improvements held consumption flat through 2018.¹³

The three uncertainty drivers identified in the opening brief map precisely onto the largest assumption differences exposed by this decomposition. AI workload efficiency trajectory is the crux of the optimization-offset versus rebound-effect disagreement and drives the widest variance on both the PUE and efficiency-gain drivers. Interconnection clearance pace determines how much of the announced pipeline converts to energized capacity and thus drives the completion-rate driver most directly. The ratio of speculative to committed capacity — the phantom-load problem — is the mechanism behind the 25 GW gap Grid Strategies identifies between utility forecasts and market-analyst benchmarks.¹¹ None of these three is resolvable with data currently in the public record. Each is a falsifiable condition whose resolution over the next 12 to 24 months will shift the defensible range materially. That is precisely what makes them the right variables to monitor. The spread is now legible. It is not smaller for being understood, but it is navigable.

The Binding Constraint: Interconnection, Equipment, and Infrastructure Bottlenecks

The interconnection queue figure most frequently cited to anchor the scale of AI-driven power demand comes from Lawrence Berkeley National Laboratory: over 2,000 GW of generation and storage capacity currently sits in interconnection queues across the United States.¹⁶ That number appears in multiple places in the available record, but those appearances trace to a single upstream source rather than independent confirmation, so the figure carries one evidentiary voice, not many. More importantly, the queue total is not a demand estimate. It is a count of applications. The fraction representing committed projects with signed power-purchase agreements, secured deposits, and site control is unresolved in the available evidence. Grid Strategies' analysis suggests that utility forecasts overstate data center demand by roughly 25 GW relative to third-party market benchmarks, and historical interconnection tracking finds that only 13 percent of projects requesting interconnection between 2000 and 2019 had reached commercial operation by end of 2024, with 77 percent withdrawn and 10 percent still under review.¹¹¹⁶ Resolving the distinction between speculative queue entries and load-bearing committed projects, through FERC milestone data or deposit-backed queue analysis, would materially change any ceiling estimate derived from the raw queue figure. Without that resolution, the 2,000-plus GW number anchors the scale of stated demand while telling us almost nothing about how much of that demand will materialize.

What the queue does confirm is that interconnection is contested terrain. The timeline from PJM interconnection application to commercial operation has risen from under two years in 2008 to over eight years in 2025.¹⁷ ERCOT's large-load queue reached 410 GW by April 2026, with 87 percent comprised of data-center requests, exceeding the grid's peak demand by 4.8 times.¹⁵ In Northern Virginia's Data Center Alley, data centers drove 97 percent of 5,250 MW of load growth, with over 30 GW of data center demand queued through 2030 in the Dominion zone alone.¹⁵ These are not projections. They are queue records and utility filings. The buildout pressure is real. The question is how much of it can clear the physical and regulatory system before 2030.

Beyond the queue itself, the physical infrastructure governing what can actually be energized imposes a separate ceiling. Large power transformers above 100 MVA now carry lead times of 120 to 210 weeks, with prices up 79 percent.¹⁵ The 2026 Infrastructure Guide similarly reports transformer lead times extending to two to three years due to global manufacturing bottlenecks, and notes that substation capacity frequently caps usable delivery at 250 to 500 MW regardless of generation availability.¹⁸ These are the same TechPlus Trends figures, from the same author and domain, so they represent one voice on equipment constraints, not independent corroboration.

The transmission side is more clearly documented. The United States built 888 miles of new 345-kV and higher transmission in 2024. The DOE's 2024 National Transmission Planning Study identifies approximately 5,000 miles per year of high-capacity regional transmission as necessary to meet current demand trajectories.¹¹ The gap between those two numbers is not a rounding difference. Goldman Sachs suggests that transmission projects can take several years to permit and then several more to build, creating what its research frames as a potential bottleneck for data center growth in regions that are not proactive given lead times.¹⁹ Industry analyses place the permitting timeline for regional transmission lines at seven to eleven years.¹⁶ Nothing in the current regulatory or construction pipeline closes that gap before 2030.

Water availability has been proposed as a constraint of comparable severity. One study quantifies U.S. data center water requirements at 697 to 1,451 million gallons per day by 2030 and estimates $10 to $58 billion in new water infrastructure costs.¹⁵ The available record surfaces a genuine disagreement here. One position, reflected in sources discussing liquid cooling efficiency gains and geographic site selection, holds that water challenges are secondary and addressable through direct liquid cooling and geographic diversification.²⁰¹⁸ The dissenting position, resting on a single independent line of analysis, argues that water may become more constraining than power in some regions by 2028 to 2030 as data centers ramp to gigawatt scale. Neither position is decisive. The reconciliation supported by the available evidence is one of phasing: power is the binding constraint now, water becomes regionally binding as the build intensifies in water-stressed markets later in the decade.

Capital markets add a further layer of friction. Elevated interest rates, commodity inflation, and volatility in construction costs are reshaping project economics, while access to inexpensive debt and long-term power contracts is tightening.²¹ Goldman Sachs estimates approximately $720 billion in grid spending through 2030 may be needed to meet data center power demands.¹⁹ PJM's forward capacity auction for the 2025/26 delivery year cost $14.7 billion, up $12.5 billion from the prior auction's $2.2 billion. Grid Strategies estimated that faster generation interconnection could have saved consumers as much as $7 billion in that single auction.¹¹ The capital requirement is real. The question is whether it can be deployed fast enough to serve projects that are themselves racing against interconnection clocks.

The available record disagrees on whether interconnection bottlenecks are primarily procedural or primarily physical. Three sources point to FERC Order 2023, cluster-based study processes, and standardized financial penalties as reforms capable of reducing study timelines from 55 months to under 18 months, framing the problem as solvable through regulatory action within two to three years.¹⁶ A larger and more independently sourced body of evidence argues that physical infrastructure limitations, transformer shortages, substation saturation, and transmission line construction timelines, persist regardless of procedural reform.¹¹¹⁵²¹ The evidence weight is asymmetric. The optimistic position rests on essentially one evidentiary line echoed across three sources; the physical-constraint position draws on three independent lines. The reconciliation the evidence supports is that regulatory reform can shorten the study phase materially while leaving the construction phase, three to seven additional years for transformers, substations, and transmission, unchanged. Both statements can be simultaneously true. But through 2030, it is the construction phase that governs.

Behind-the-meter generation enters the picture as a partial bypass. Bloom Energy forecasts that 38 percent of data centers will use on-site generation by 2030, up from 13 percent in 2025, representing what the Luminix analysis describes as a 27-times increase in behind-the-meter deployment.¹⁵ RAND modeling cited in the same source estimates 49 GW of behind-the-meter net capacity by 2030 versus 33 GW front-of-meter. Those figures come from sources whose promotional framing colors their assumptions: Bloom Energy's forecast is its own market projection, and the Luminix report treats constraint sequencing as investment opportunity. A separate line of analysis finds that behind-the-meter natural gas costs exceed $120 per MWh, well above forward energy prices, and that FERC has restricted large behind-the-meter arrangements on reliability and cost-shifting grounds.¹⁵ The evidence suggests behind-the-meter will serve as a tactical workaround for some projects, not a strategic substitute for grid capacity. The Luminix analysis implies that if 25 to 38 percent of incremental data center load goes behind-the-meter, the grid-served increment drops to roughly 350 to 480 TWh, materially below headline base-case figures.¹⁵

Drawing these constraints together, interconnection delay stands as the most durable ceiling through 2030. Equipment lead times are long, but transformer manufacturing can be accelerated with capital, and multiple manufacturers are already expanding capacity. Financing constraints are real but cyclical and interest-rate-sensitive. Water challenges are regional and partially addressable through technology and siting choices. Transmission construction timelines of seven to eleven years for permitting alone place new high-capacity lines structurally outside the 2030 window for projects not already in advanced development.¹⁶ That argument is contestable: one could weigh equipment or capital constraints as more immediately binding in specific regions or project types. But the case rests on the relative addressability of each constraint on a sub-decade cycle. Equipment and financing are faster to resolve than regulatory and physical timelines for grid capacity.

This matters directly for the defensible demand range established earlier in this report. The 5-year forecast for utility peak load growth has increased by more than a factor of six over three years, from 24 GW in 2022 to 166 GW in 2025.¹¹ Data centers account for approximately 55 percent of that forecasted growth. Yet alternative benchmarks from data center market analysts suggest actual materialization is unlikely to exceed 60 to 65 GW through 2030, implying the utility gross forecasts overstate demand by roughly 25 GW.¹¹ The average size of a proposed U.S. data center doubled between 2023 and 2024, from 150 MW to 300 MW.¹¹ Larger projects face longer and more expensive interconnection processes. The same Grid Strategies analysis that documents the forecast inflation also documents that only about 65 percent of announced projects may materialize, favoring incumbents with sited power access over speculative builds.²

The interconnection clearance pace identified as a primary uncertainty in the executive brief is therefore the constraint most likely to determine where within the defensible range demand actually lands. Grid Strategies' data on historical completion rates, LBNL's queue withdrawal patterns, and PJM's capacity auction pricing all point in the same direction. A ceiling on how fast physical supply can clear demand exists regardless of what any load forecast assumes. That ceiling is set not by capital availability or equipment lead times, both real but shorter-cycle constraints, but by the regulatory and physical timelines governing new grid capacity. Until FERC milestone data or deposit-backed queue analysis can separate committed projects from speculative filings, the 2,000-plus GW queue figure signals the scale of the pressure without resolving how much of it translates to energized load before 2030.

The Overbuild and Bubble Question: Bear Signals Against Bull Rebuttals

Four bear signals challenge the consensus demand narrative, and four bull rebuttals answer them. Neither side can be dismissed. That is the honest position.

Signal one: speculative queue inflation. The prior section established that over 2,000 GW of generation and storage capacity sits in U.S. interconnection queues, representing more than the total existing U.S. generation fleet — a figure the Lawrence Berkeley National Laboratory supplies as the scale anchor for stated demand.⁴ The bear case is that this figure conflates genuine commitments with speculative entries. As the World Resources Institute reports, utilities are being flooded with early-phase interconnection requests filed at low cost to gauge connection timelines, producing both double-counting of projects and phantom load that will never be built.¹³ The WRI analysis further notes that power constraints have already extended data-center construction timelines by 24 to 72 months, and shortages of transformers, switchgear, and gas turbines are compounding the supply-side limitation.¹³ The Luminix synthesis offers a sharp illustration: ERCOT's large-load queue reached 410 GW by April 2026, with 87% comprised of data-center requests, exceeding peak demand by 4.8x.¹⁵ That ratio alone signals a structural overshoot of stated versus realizable demand.

The bull rebuttal is that announced capex and signed power-purchase agreements constrain the fiction. Utility five-year peak-demand forecasts grew from 38 GW in 2023 to 128 GW in 2024, a revision so large it reflects genuine customer commitments rather than speculative noise.¹³ The 2024 LBNL report explicitly avoids the overcounting problem by using a bottom-up model built from installed equipment rather than from interconnection requests, and identifies that methodology as what distinguishes its estimates from utility-reported queue volumes.¹ The disagreement between the two positions is reconcilable in structure, if not in magnitude: gross queue figures overstate demand, net realized demand will be lower, and the residual question is how much lower. The available evidence suggests a 20-to-40 percent haircut is plausible, but the precise ratio of speculative to committed capacity cannot be read from the current record.

Signal two: onsite gas generation overbuild. A news report covering the IEA's Energy and AI special report states that natural gas will grow by 175 TWh to meet data-center demand, with much of that growth in the U.S. driven by pro-gas policies and new behind-the-meter generation.²² The Luminix report states Bloom Energy forecasts that 38% of data centers will use onsite generation by 2030, up from 13% in 2025, and that RAND estimates 49 GW of behind-the-meter net capacity by 2030 versus 33 GW front-of-meter — though both figures are stated without supporting evidence in that source.¹⁵ Luminix's own implication is that if 25 to 38 percent of incremental data-center load goes behind the meter, the grid-served increment drops to roughly 350 to 480 TWh, materially less than the headline 550 to 650 TWh base case.¹⁵ The IEA's own report, as summarized by S&P Global, identifies significant uncertainties around AI adoption rates, efficiency improvements, and potential grid bottlenecks that could affect projected demand growth.²³

The bull rebuttal on this signal is that behind-the-meter deployment at those levels is economically constrained. The cost dynamics, regulatory pushback from FERC on large BTM arrangements, and emissions scrutiny mean that BTM will serve as a workaround for a fraction of projects rather than a strategic substitute for grid power. Even accepting the report's own range, the lower bound of grid-served demand (roughly 325 TWh) still represents a massive buildout relative to the 176 TWh baseline the 2024 LBNL report establishes for 2023.¹ The overbuild risk on gas generation is real, but it is a risk of misallocated capital within a genuine demand surge, not evidence that the surge itself is illusory.

Signal three: historical overshoot precedent. The WRI analysis provides the clearest historical anchor. Between 2010 and 2018, global data-center electricity use was essentially flat, even as computing demand soared, because efficiency improvements offset load growth. In hindsight, the WRI states, electricity demand was overestimated during that earlier period, with premonitions of power shortages that echo current concerns.¹³ The Dev Sustainability analysis of forecasting methodology reinforces the point: a 2022 review of 258 data-center energy estimates from 46 publications between 2007 and 2021 found only two sources of reliable estimates.¹² The field's track record on accuracy is weak.

The bull rebuttal here carries real weight. The 2024 LBNL report documents what changed. During 2010 to 2016, efficiency gains came from migrating workloads out of inefficient enterprise data centers into efficient hyperscale facilities, a one-time structural shift that kept demand flat despite explosive service growth.¹ The 2024 LBNL report also shows that data-center electricity consumption grew at 18% compounded annually between 2018 and 2023, accelerating sharply from 7% between 2014 and 2018.¹ AI workloads already deploy into the most efficient infrastructure available, as the Dev Sustainability analysis observes, so the historical efficiency lever — workload migration from bad to good infrastructure — has already been pulled.¹² The historical precedent is real, but the structural conditions that generated it are not present this cycle.

Signal four: the DeepSeek efficiency shock. The Dev Sustainability analysis captures the dynamic clearly. Efficiency gains in AI systems will have to come from better accelerators, better model architectures, higher utilization, better scheduling, and inference optimization.¹² The WRI analysis notes that experts have warned about a rebound effect: even with efficiency improvements, the industry might see increased demand rather than reduced consumption.¹³ This is the Jevons-paradox argument, and it cannot be dismissed with available evidence.

The Jevons-paradox rebuttal runs as follows. When the cost per inference falls because of algorithmic efficiency, demand for inference expands. Cheaper AI tasks generate more AI tasks. The Dev Sustainability analysis describes a workload evolution from chat through reasoning to agents, where software decomposes a single user request into many model calls, tool calls, retrieval steps, and verification loops.¹² Each efficiency gain expands the deployment frontier rather than contracting total consumption. The same source observes that agentic workloads and multi-step task orchestration are also driving CPU demand alongside GPU inference, broadening the energy base of AI beyond the accelerator segment that forecasters track most closely.¹²

This is where the evidence is insufficient to adjudicate the disagreement. The Jevons-paradox argument is logically coherent and historically attested in other sectors. The efficiency-shock argument is also coherent and empirically grounded in the algorithmic improvements already observed. The available record does not supply the data needed to determine which effect dominates, because the utilization rates, duty cycles, and actual inference volumes at deployed scale are not publicly reported. The Dev Sustainability analysis identifies this gap explicitly, noting that agentic workloads and shifting usage patterns complicate energy forecasts and that the source cannot tell us what total workload will be generated as inference scales.¹²

Here is the compounding risk that the evidence does warrant naming. The editorial conditions of this analysis require drawing the following conclusion explicitly. If inference displaces training as the dominant workload — Luminix implies a roughly even training-to-inference split in its base case, while other sources suggest inference already dominates¹⁵ — while algorithmic efficiency simultaneously reduces the compute required per inference task, the bull case faces pressure from two directions at once. A single efficiency shock compresses demand per task. A simultaneous workload shift toward lower-energy-per-query inference compresses the demand base those tasks draw from. Call this the workload-efficiency double compression. It is the compound form of the DeepSeek episode's significance: not one efficiency surprise, but a demonstration that efficiency surprises and workload-mix shifts can arrive together. Whether this compound effect is large enough to materially change the 2030 range depends on the efficiency trajectory assumption, which remains the most consequential single input to the forecast — exactly as the executive brief identified.

The methodological conflict between these positions is genuine and unresolved. The 2024 LBNL report states directly that historical utility demand forecasts have consistently overestimated both peak and average demand, and that many financial-sector models are proprietary or methodologically opaque.¹ Against that, the LBNL bottom-up approach itself, which projects a range of 325 to 580 TWh for 2028, requires utilization-rate and GPU-shipment assumptions that are rapidly evolving and difficult to forecast, as the report acknowledges.¹ The IEA's 945 TWh global projection for 2030 is reported without confidence intervals or scenario probabilities in the secondary accounts available.²²²³ The Dev Sustainability author summarizes the analytical situation fairly: extreme high-end scenarios are useful for getting an estimate quoted in the press rather than as realistic projections of actual consumption, while the economic incentive to improve efficiency will eventually manifest in actual consumption.¹²

What survives scrutiny from the bear case? Three things. First, the interconnection queue overstates committed demand by an unknown but material amount, meaning the physical ceiling on realized load is lower than queue volumes suggest. Second, historical efficiency improvements have repeatedly confounded power-surge predictions, and the structural mechanisms that made 2010-to-2018 demand flat are no longer available — but algorithmic efficiency is a new mechanism that was not present in that earlier period. Third, the workload-efficiency double compression is a falsifiable risk: if inference's share of AI compute rises toward the levels some sources suggest, while per-inference compute requirements continue falling, the compound effect would move realized demand toward the lower end of the LBNL range rather than the high end.

What survives from the bull case? One central finding. The 2010-to-2018 efficiency era was structurally different from the current moment. Demand growth has already re-accelerated. The LBNL report's 18% compound annual growth rate between 2018 and 2023 is not speculative — it is measured against actual installed equipment.¹ The Jevons-paradox dynamic, if it operates through agentic workloads and multi-step orchestration, could sustain that growth rate even under significant per-task efficiency gains. The overbuild risk on gas generation and the speculative fraction of interconnection queues do not change this directional finding. They change the magnitude.

The section's honest output is not a resolved range. It is a sharpened set of conditions. The workload-efficiency double compression carries the forecast toward the low end of defensible estimates. The Jevons-paradox expansion of workload carries it toward the high end. The efficiency trajectory assumption from the executive brief is the crux of both, confirming it as genuinely uncertainty-generating rather than resolvable by current data. An investor underwriting through 2030 must carry this as an open position, not a closed one.

Winners, Payers, and Second-Order Effects: Scenario Mapping Across the Supply Chain

The bull scenario is straightforward to sketch. If US data-center electricity demand materializes at consensus magnitude, the value flows are clear. Independent power producers holding contracted generation capacity, regulated utilities pursuing rate-base expansion, natural gas producers and midstream operators, nuclear developers, and grid-equipment manufacturers all capture disproportionate value. The LBNL baseline established earlier — 176 TWh in 2023, with projections ranging from 325 to 580 TWh by 2028 — anchors the scale of that prize.¹ Utility five-year peak load growth forecasts have risen more than six-fold over three years, from 24 GW to 166 GW, with data centers accounting for roughly 55% of that growth.¹¹ At those magnitudes, the infrastructure build is enormous.

The bull scenario, however, conceals a structural problem. Several utilities are seeking to socialize the capital cost of demand-driven infrastructure additions through rate-base expansion — building ahead of load and recovering costs from all ratepayers rather than from the industrial customers whose demand justified the construction. The analysis notes that some experts have already raised the alarm about current rate structures and cost allocation methods leading to the possibility of other residential or commercial consumers paying for infrastructure used only to service new data centers.¹³ If the demand that justified those additions does not arrive on schedule, ratepayers bear the stranded-cost risk. The utilities most exposed are those with the largest demand-driven programs.

PJM's capacity market has already revealed what that exposure looks like. Its forward capacity auction for the 2025/26 delivery year cost $14.7 billion, up $12.5 billion from the previous auction's $2.2 billion. Grid Strategies estimated that faster interconnection of queued resources could have saved consumers as much as $7 billion in that single auction.¹¹ PJM's 2027/2028 Base Residual Auction cleared at the FERC cap of $333 per MW-day but still fell 6,517 MW short of the reliability requirement — the first shortfall exceeding one percentage point in PJM history.¹⁵ These are costs that ratepayers absorb whether or not AI demand reaches the levels utilities projected when they justified the build.

The bear scenario sharpens this exposure. If demand falls materially short, hyperscaler and data-center REIT valuations compress. Utilities with aggressive rate-base expansion programs face regulatory scrutiny over stranded costs. The grid-equipment manufacturers and nuclear developers that positioned for a bull cycle find themselves holding capacity against a market that did not clear. The WRI points to a directly relevant historical precedent: between 2010 and 2018, global data-center electricity use was essentially flat despite soaring computing demand, because efficiency improvements offset load growth. In retrospect, electricity demand was overestimated, with concerns about power shortages echoing current anxieties.¹³

Against this backdrop, every interested forecaster's incentive matters. Utility integrated resource plans that justify rate-base expansion have a structural bias toward demand accommodation. The WRI explicitly notes that historical utility demand forecasts have consistently overestimated both peak and average demand — a finding LBNL's 2024 report echoes directly.¹¹³ Hyperscaler public siting announcements signal capability rather than commitment. Project Stargate, announced in January with $500 billion in stated AI data-center spending by 2029, had experienced a slower than expected start and had announced plans for only one small data center to open by year-end.¹³ Announced projects are demand signals, not consumption facts.

That framing is bullish by construction. The LBNL report observes that the underlying models on which financial analytics firms build their projections are proprietary or methodologically opaque.¹ Sell-side desks serve their coverage franchises. That is not disqualifying, but it must be legible.

Independent power producers, gas producers, and nuclear promoters all need the demand narrative to hold. On the behind-the-meter side, a projection that 38% of data centers will use on-site generation by 2030, up from 13% in 2025, and RAND's estimate that 49 GW of behind-the-meter capacity will exceed 33 GW front-of-meter by 2030, both carry the weight of assertions without independent verification in the available record.¹⁵ Bloom Energy sells on-site generation equipment. The promotional framing of that forecast is its own evidentiary context.

The IEA and LBNL are the closest available approximation of disinterested analysis. LBNL's bottom-up equipment model builds from commercial equipment shipment data, utilization assumptions, and PUE trajectories rather than from project-pipeline optimism or workload extrapolation.¹² That methodology avoids the overestimation that can arise from tracking data-center load for projects that have not yet selected a power provider.¹ The IEA's institutional mandate to advise OECD governments creates different incentives from sell-side coverage or vendor forecasting. Neither is free from methodological limitations, but both are more transparent about their assumptions than proprietary financial models. For an investor constructing a defensible range, these two anchors are the appropriate point of departure.

One finding, however, holds across scenarios. Large power transformers now carry lead times of 120 to 210 weeks with prices up 79%.¹⁵ The average size of a proposed US data center doubled from 150 MW to 300 MW between 2023 and 2024.¹¹ Grid interconnection timelines from PJM application to commercial operation have risen from under two years in 2008 to over eight years in 2025.¹⁷ Against that backdrop, behind-the-meter natural gas generation can be deployed substantially faster than grid interconnection can be secured. Cleanview tracked on the order of 50-plus GW of announced behind-the-meter or co-located generation associated with data center projects, with most identifiable equipment being natural gas-based.²⁴ The EIA projects 3.3 GW of combined-cycle natural gas generation online in 2026, another 3.3 GW in 2027, and 10.6 GW in 2028.¹⁵ Natural gas producers and midstream operators are near-term winners regardless of which long-run demand scenario materializes, because their product fills the gap while interconnection queues clear. This is a scenario-robust finding. It distinguishes itself from every other value-distribution claim in this section, which is scenario-contingent.

The conflict between interconnection optimists and pessimists matters here. Sources that frame queue delays as primarily a procedural problem — addressable through regulatory reforms such as cluster study redesign — represent one independent line of argument, treating the problem as administrative rather than physical in nature. Sources documenting physical infrastructure constraints — transformer lead times, substation saturation, and transmission construction timelines that run 7 to 11 years for permitting alone — represent three independent lines.¹⁵²⁴¹⁷ The weight of evidence favors the structural reading. Procedural reforms may reduce study-phase timelines, but they do not shorten the physical construction of transformers and transmission lines. For any stakeholder whose returns depend on demand materializing at the grid by 2030, that asymmetry is material.

The efficiency question divides the outlook for specific classes of asset. The IEA suggests that AI-enabled grid optimization could unlock 175 GW of transmission capacity through improved management, and that broader efficiency improvements in industrial and building sectors could yield 13 exajoules of energy savings. If those gains materialize at scale, they could offset or reduce incremental generation need. The competing position, supported by two independent analytical lines, holds that the prior decade's efficiency lever — moving workloads from inefficient enterprise data centers into efficient hyperscale facilities — has already been pulled. AI is being deployed into the most efficient infrastructure available, so further gains must come from better accelerators, model architectures, utilization scheduling, and inference optimization rather than from infrastructure consolidation.¹² The evidence does not resolve this disagreement. The IEA's 175 GW transmission unlock claim rests on adoption rates and technical performance that are not independently validated in the available record. For an investor, the unresolved nature of this conflict is itself the signal: efficiency-dependent theses carry more uncertainty than demand-supply gap theses.

The Luminix synthesis, which blends EPRI medium, McKinsey, and LBNL approaches to arrive at a 550 to 650 TWh base case for 2030, shares an upstream origin across both of its appearances in the available record. It represents one analytical voice, not independent corroboration.² That range implies the grid-served increment could fall to roughly 350 to 480 TWh if 25 to 38% of incremental load goes behind-the-meter — still a substantial buildout, but materially below headline figures.¹⁵ Under a bear scenario, the gap between announced capacity and realized demand is where stranded-cost risk concentrates. Utilities that built for the headline and recover for the base case are the most exposed. Ratepayers in their service territories bear the difference.

Synthesis and the Investable View: A Defensible Range with Falsifiable Conditions

The range established in the executive brief can now be populated with its load-bearing architecture. Three analytical threads run through the preceding sections: the verified 2023 baseline of 176 TWh consumed by US data centers, the roughly 2x spread among major forecasters, and the finding that interconnection delay gates physical delivery more severely than any other single constraint through 2030. This section assembles those threads into a defensible demand band, ties each boundary to its operative assumptions, and specifies what observable data over the next 12 to 24 months would move the band materially.

The defensible US range and its boundaries

The LBNL 2024 bottom-up model projects 325 to 580 TWh by 2028, representing 6.7% to 12.0% of forecasted US electricity consumption, driven by AI server shipments, GPU power ramps running at 60% to 80% of rated capacity, and PUE declining toward roughly 1.4 through liquid cooling and hyperscale consolidation.¹² The LBNL team reports a compound annual growth rate of approximately 7% from 2014 to 2018, accelerating to 18% between 2018 and 2023, with the 2023-to-2028 band ranging from 13% to 27%.¹ A synthesis that extrapolates LBNL's reported CAGR band forward to 2030 yields an approximate range of 450 to 850 TWh.²

The EPRI state-level pipeline analysis, which aggregates operational, under-construction, and announced projects under low, medium, and high realization scenarios, projects 380 to 790 TWh by 2030, representing 9% to 17% of US electricity use, converting IT capacity estimates of 56 to 132 GW to energy through implicit load factors and PUE assumptions.² That pipeline estimate represents a 60% upward revision from EPRI's 2024 forecast, driven by 18 months of accelerated AI infrastructure announcements.² Goldman Sachs projects approximately 810 TWh of US data-center electricity demand by 2030, at 14% to 16% of US total, based on a 220% global growth scenario with a 60% US share of incremental demand driven by hyperscaler capital expenditure surges exceeding $300 billion.²

Blending the EPRI medium scenario, McKinsey's 606 TWh workload-driven estimate, and the LBNL high-end figure, and weighting for hyperscaler dominance at roughly 85% of servers, a PUE range of 1.2 to 1.4, and an approximate 50/50 training-to-inference split, yields a defensible base case of 550 to 650 TWh by 2030.² A sensitivity analysis attached to this synthesis indicates potential variance of plus 20% if inference workloads exceed training, minus 15% if GPU efficiency and PUE below 1.2 are achieved, and plus or minus 10% from project realization risk.²

At the global level, the International Energy Agency projects that data center electricity consumption could double by 2030, potentially reaching 600 to 800 TWh annually.⁴ That figure comes from a single origin repeated across multiple URLs; it is one voice, not a corroborating chorus. The Belfer Center reports that institutional projections of data center electricity demand range from about 200 TWh to over 1,000 TWh by 2030, a spread that itself confirms the range cannot be collapsed to a point estimate.¹⁰

Why the range refuses to collapse

The three assumptions identified at the outset of this report remain unresolved by the available evidence. They determine where within the 550-to-650 TWh base case demand actually lands, and they determine whether demand falls below that floor or rises above the Goldman Sachs ceiling.

First, the efficiency trajectory of AI workloads. The evidence on this is contested. Two independent analytical lines argue that efficiency gains from the 2010s, which kept data center demand nearly flat despite surging service demand, drew primarily on a one-time structural shift from on-premises to hyperscale facilities.¹²¹³ That lever is now exhausted. The IEA, by contrast, has modeled 175 GW of transmission capacity potentially unlockable through AI-based grid management and 13 exajoules of energy savings through AI-enabled industrial optimization, figures that could theoretically offset incremental data center load.²⁵²⁶ The evidentiary weight here is asymmetric: the efficiency-exhaustion position rests on two independent lines; the IEA optimization claim traces to a single origin echoed across two sources. The range's lower bound depends on efficiency gains, whose adoption pace the available evidence supports only at a "suggests" level, with no operator deployment data in the underlying record. The floor is more assumption-sensitive than the ceiling. Chip-designer GPU power roadmaps and hyperscaler cooling capital expenditure disclosures are the specific data that would tighten the lower bound.

Second, the pace of interconnection clearance. This is the conflict where evidence weight is most asymmetric. The case that FERC Order 2023 and grid-enhancing technologies can reduce study timelines from 55 months to under 18 months draws on one analytical line echoed across three sources.²⁷²⁸²⁹ The case that physical infrastructure constraints will persist regardless of procedural reform draws on three independent lines: transformer lead times exceeding 200 weeks, substation capacity exhaustion, and transmission permitting timelines of 7 to 11 years.¹¹²¹¹⁶ Both positions can be true simultaneously: procedural reforms may shorten study phases while physical buildout timelines remain the binding constraint through 2030. The range ceiling is therefore set by interconnection delivery, not by announced demand or capital commitment.

Third, the ratio of speculative to committed capacity. The 5-year utility peak load growth forecast has risen by more than a factor of six over three years, from 24 GW to 166 GW.¹¹ Data centers account for roughly 55% of that forecasted growth, or approximately 90 GW.¹¹ Alternative benchmarks from data center market analysts suggest actual demand growth is unlikely to exceed 60 to 65 GW through 2030, implying utility forecasts collectively overstate data center demand by roughly 25 GW.¹¹ Only approximately 65% of announced projects may materialize by 2030, per Grid Strategies benchmarks.² Of sixteen GW-scale data centers scheduled to come online in 2026 and 2027 with aggregate demand of nearly 30 GW, nearly all hold initial permits and around half have begun construction.¹¹ The completion rate of that tranche is the near-term test of whether the speculative discount is 35% or closer to 10%.

The methodological conflict underlying the range

A genuine disagreement runs through the forecast literature and must be carried forward rather than silenced. LBNL argues that bottom-up equipment models are more defensible than pipeline extrapolation methods, and documents that historical utility demand forecasts have consistently overestimated both peak and average demand.¹ Counterarguments exist: modern forecasts represent a real methodological improvement over earlier work, given that a review of 258 data center energy consumption estimates found systematic methodological defects across the field, particularly regarding data availability and transparency.¹⁰ The reconciled position is that modern estimates are more rigorous, but that the uncertainty range from 325 to 580 TWh by 2028 reflects genuine epistemological limits, not analytical failure.¹

The 12-to-24-month observable indicators

Four data streams, observable within the next 12 to 24 months, would confirm or falsify the bull-case assumptions and shift the defensible range materially.

Interconnection approvals translating to energized capacity. The question is not how many GW sit in the queue but how many reach commercial operation. A realization rate tracking toward the Grid Strategies 60-to-65 GW market-analyst benchmark would compress the upper bound toward the LBNL base case. A realization rate tracking toward the utility 90 GW figure would validate the EPRI high scenario.

Hyperscaler capital expenditure execution rates. Amazon, Microsoft, Google, and Meta collectively spent over $200 billion on capital expenditure in 2024, a 62% year-over-year increase from 2023.¹⁰ Whether that rate sustains, accelerates, or normalizes through 2026 is a leading indicator of the committed versus speculative split. A material deceleration below 2024 capex levels would be a bear signal.

GPU shipment and utilization data. The average size of a proposed US data center doubled between 2023 and 2024, from 150 MW to 300 MW.¹¹ GPU shipment data at the chip-designer level, combined with hyperscaler disclosures of actual utilization rates, would tighten the utilization assumption that currently spans a wide band in the LBNL model.

Efficiency-benchmark trajectories. A reduction in grid-delivered PUE below 1.2 across newly commissioned hyperscale facilities, or documented liquid cooling adoption exceeding the rates embedded in the LBNL base case, would push realized demand toward the lower bound. The absence of such disclosures — which remain voluntary — would leave the floor assumption unresolved.

The residual uncertainty that must be carried

The overbuild risk cannot be quantified with confidence from the available evidence. The scenario in which demand falls materially short exposes ratepayers to stranded costs from unnecessary infrastructure investment; utilities with the largest demand-driven build-out programs carry the most concentrated bear-case risk.¹⁰ The direction of the demand surge is real. Its magnitude at the high end is not load-bearing. A range of 450 to 650 TWh for US data-center electricity demand by 2030, with the bull case approaching 810 TWh only if inference growth outpaces efficiency and interconnection delivery exceeds current trajectories, is what the evidence supports. The range is the honest output. Collapsing it to a point estimate would substitute a false precision for the genuine uncertainty the evidence requires an investor to carry.

Limitations and What We Could Not Determine

Four structural limitations bound this report's conclusions, and an investor who ignores them will mistake the range's precision for certainty it does not possess.

All 2030 figures are assumption-driven ranges, not forecasts. The defensible US demand range assembled in the synthesis section rests on three assumptions whose resolution remains genuinely open: the efficiency trajectory of AI workloads, the pace of interconnection and infrastructure clearance, and the ratio of speculative to committed capacity additions. Confidence intervals widen with the forecast horizon. The LBNL bottom-up model, the most methodologically rigorous tool available, projects only to 2028, with a scenario span of roughly 325 TWh to 580 TWh at that horizon. Extending those scenarios two additional years compounds the uncertainty. The range in this report is the honest output; a point estimate would be false precision.

The AI-workload disaggregation problem cannot be resolved by synthesis alone. The baseline section established that EPRI estimated AI consumed 10% to 20% of data center energy in 2024, but that figure rests on announced project pipelines rather than metered consumption. As the LBNL 2024 report concedes, its bottom-up model requires many inputs and assumptions developed from limited publicly available data. Grid Strategies' utility forecast analysis confirms that AI load is not separately tracked in any utility forecast it examined; data center projections bundle hyperscale, colocation, and AI facilities together. Every forward projection inherits this measurement gap. No synthesis of secondary sources closes it. Primary telemetry from deployed systems does not exist in the public record, and hyperscalers have no legal obligation to provide it. The CRS confirmed in 2025 that there are no legally binding energy standards that apply to private-sector data centers and that the EIA's 2024 attempt to collect cryptocurrency mining data ended with a court-ordered agreement to destroy what had been gathered. This is a structural absence, not a gap that additional research reports will fill.

Self-reported data carry unknown bias. Hyperscaler capital-expenditure announcements, siting disclosures, and renewable-procurement claims enter the forecast models as demand signals. A data center announcement is not electricity consumption; as the Dev Sustainability analysis put it, it is a claim on power, land, equipment, cooling, interconnection, and political permission. Grid Strategies found that alternative analyst benchmarks suggest utilities are overstating data center demand by roughly 25 GW relative to what market analysts expect to materialize. The direction of self-reporting bias is toward overstatement. Its magnitude is unknown, and no external audit mechanism currently exists.

Policy and efficiency shocks can invalidate any specific 2030 figure. This is not a generic caveat. The DeepSeek episode, examined in the overbuild section, demonstrated within the current forecast window that a single algorithmic development can alter the efficiency assumptions on which multi-hundred-TWh demand projections rest. The Jevons-paradox and efficiency-shock positions remain a genuine disagreement, unresolvable with current data. Either could be correct. A range that refuses to collapse to a point estimate is the epistemically honest response to that uncertainty, not a failure of analysis.

Two categories of limitation deserve separate treatment because they point in different directions. Some limits are endemic to the problem: without mandatory disclosure requirements, hyperscaler actual consumption will remain opaque, and the AI-workload disaggregation problem will persist regardless of analytical effort. Other limits are addressable. Mandatory reporting legislation, of the kind proposed in the Clean Cloud Act of 2025 though not yet enacted, would provide EIA with the legal authority to collect facility-level data. Regional granularity in interconnection data already exists at the ISO and utility level; linking it systematically to AI-specific demand drivers is a tractable research task. GPU utilization telemetry from major deployments, if disclosed, would substantially narrow the scenario range.

The research agenda is clear. The will to execute it is not.

Limitations

• Missing perspective. Data center operators' and hyperscalers' actual demand forecasts, investment priorities, and risk assessments—corpus relies heavily on third-party analyst projections (IEA, Goldman Sachs, LBNL) rather than primary disclosures from AWS, Google, Microsoft, Meta on their 2026-2030 capex plans and power procurement strategies.
• Missing perspective. Utility operators' and grid planners' feasibility assessments and willingness-to-invest analyses—corpus focuses on what demand projections imply for grid capacity but lacks utility company internal forecasts, cost-benefit analyses, or stated confidence in meeting projected timelines.
• Missing perspective. Regulatory and local opposition voices—strikebound, environmental, and community stakeholder perspectives on data center siting and grid expansion are largely absent; corpus does not include testimony, litigation records, or local government position statements that might constrain buildout.
• Missing perspective. Non-U.S. regional and developing-world electricity markets—corpus is heavily weighted toward U.S. and EU projections; perspectives from data center operators, grid planners, and policymakers in Asia-Pacific, India, Middle East, and Africa are either absent or filtered through Western-sourced forecasts.
• Missing perspective. Chip designers and semiconductor manufacturing perspectives on GPU/accelerator power efficiency roadmaps and actual deployment timelines—corpus cites equipment shipment assumptions but lacks primary perspective from NVIDIA, AMD, Intel, TSMC on realistic power reduction and performance scaling through 2030.
• Missing perspective. Grid reliability and stability engineers' assessment of operational risk from rapid load growth—corpus addresses interconnection queue bottlenecks and capacity planning but lacks voices of NERC, regional transmission operators, and grid stability specialists on frequency response, voltage control, and system resilience under high data-center penetration.
• Coverage gap — Demand destruction and elasticity scenarios. Corpus projects electricity demand under business-as-usual or efficiency-improved scenarios but does not systematically model demand destruction from high electricity costs, regulatory rationing, or data center operator shift to lower-cost jurisdictions. AI adoption slowdown due to model-scaling saturation or economic recession is mentioned but not quantified in forward projections.
• Coverage gap — AI workload composition and real-world power profiles. Projections treat AI electricity demand as an aggregate add-on but do not disaggregate inference vs. training, dense vs. sparse model architectures, or the actual duty cycle and utilization rates of GPUs/accelerators in production. Corpus lacks primary telemetry from deployed systems.
• Coverage gap — Environmental and resource constraints beyond electricity grid. Water availability, cooling technology feasibility, supply-chain constraints on data center materials, and embodied carbon are treated only tangentially. Corpus does not model whether water scarcity or semiconductor supply limits will be binding constraints on data center deployment before grid capacity becomes limiting.
• Coverage gap — Economic viability and return on capex for data center projects. Corpus documents $720 billion+ grid investment requirements and $1 trillion+ data center capex projections but does not assess whether AI revenue or efficiency gains justify these investments or whether expected returns have declined. Bubble-risk framework is outlined but not empirically executed.
• Methodology caveat. Corpus is dominated by executive summaries, secondary syntheses, and vendor/law firm thought leadership rather than peer-reviewed primary studies with disclosed methodologies and uncertainty quantification. Most demand projections (IEA, Goldman Sachs, LBNL) are cited through tertiary sources without access to original modeling assumptions, sensitivity analyses, or confidence intervals.
• Methodology caveat. Forecasts rely on historical data through 2024 but must extrapolate through 2030 in a period of unprecedented AI-driven technological change; historical accuracy of prior decade's data-center forecasts is not systematically assessed, limiting confidence in forward projections.
• Methodology caveat. Regional granularity is uniformly limited; projections aggregate to national (U.S.), continental (EU), or global levels. Interconnection queue and grid constraint data exist at ISO/utility granularity but are not systematically linked to AI-specific demand drivers by region.
• Methodology caveat. Corpus includes substantial bias-as-data sources (vendor white papers, law firm advisory notes, commercial research firm reports) that declare commercial interest but are weighted equally with government and academic sources in synthesis. Relative credibility and incentive structures are not systematically accounted for.
• Methodology caveat. AI adoption and workload assumptions are largely exogenous to the corpus—demand is projected as a function of announced capex and equipment shipments, not bottom-up modeling of model-scaling curves, inference-to-training ratios, or realistic utilization rates under operational constraints.
• Methodology caveat. Source acquisition was reachability-bound: 13 of 71 candidates could not be retrieved (18%), concentrated in electricchoice.com (2), image-ppubs.uspto.gov (2), powering-intelligence.epri.com (2). Literature hosted behind these publishers is systematically underrepresented; coverage of any question one such venue dominates may be skewed.
• Methodology caveat. 11 source(s) unreachable at acquisition were replaced by the next-ranked vetted candidates from the same discovery pool.

Bibliography

1. Shehabi, Arman, Sarah J. Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor. 2024 United States Data Center Energy Usage Report. Berkeley: Lawrence Berkeley National Laboratory, December 2024. LBNL-2001637. Accessed July 7, 2026. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf.
2. "Research and synthesize the leading publicly available forecasts for US AI/data center electricity demand through 2030 from…." Accessed 2026-07-07. https://www.useluminix.com/reports/industry-analysis/powering-the-ai-boom-where-the-grid-breaks-first-2026-2030/source/0.
3. International Energy Agency. Energy Demand from AI. Paris: IEA, 2024. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai.
4. Global Electricity. "Data Centers and AI Energy Consumption: The Surge in Electricity Demand." globalelectricity.org. Accessed July 7, 2026.
5. Ropes & Gray LLP. "Data Center Investment in 2026: AI Demand, Power Constraints, and Private Equity Trends." Ropes & Gray Viewpoints, 2026. https://www.ropesgray.com/en/insights/viewpoints/102mvfl/data-center-investment-in-2026-ai-demand-power-constraints-and-private-equity.
6. Data Center Resources. "How energy availability, interconnection delays, and execution complexity are redefining where hyperscale and AI infrastructure gets built in the United States." Datacenterltd.com. Accessed July 7, 2026. https://www.datacenterltd.com/articles-and-resources/the-data-center-bottleneck-no-one-wants-to-price-in.
7. U.S. Department of Energy. "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers." News release. Washington, D.C.: DOE, 2024. Accessed July 7, 2026. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers.
8. U.S. Congress, Congressional Research Service. "Data Centers and Their Energy Consumption: Frequently Asked Questions." R48646 (2025). Accessed July 7, 2026. https://www.congress.gov/crs-product/R48646.
9. SOCOMEC. "Understanding the Power Consumption of Data Centers." Accessed July 7, 2026. https://www.socomec.us/en-us/solutions/business/data-centers/understanding-power-consumption-data-centers.
10. Belfer Center for Science and International Affairs. "AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment." Harvard Kennedy School (2025). https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid. Accessed July 7, 2026.
11. Wilson, John D., Sophie Meyer, Zach Zimmerman, and Rob Gramlich. Power Demand Forecasts Revised Up for Third Year Running, Led by Data Centers. Grid Strategies LLC, November 2025. https://gridstrategiesllc.com/wp-content/uploads/Grid-Strategies-National-Load-Growth-Report-2025.pdf.
12. Author name unknown. "AI Data Center Energy in 2026." Development and Sustainability (Substack). Accessed July 7, 2026. https://www.devsustainability.com/p/ai-data-center-energy-in-2026.
13. World Resources Institute. "The Challenge of Forecasting Electricity Needs." WRI Insights. Accessed July 7, 2026. https://www.wri.org/insights/us-data-centers-electricity-demand.
14. Utility Dive. "Five-Year US Load Growth Forecast Surges 456%, to 128 GW: Grid Strategies." Accessed July 7, 2026. https://www.utilitydive.com/news/shocking-forecast-us-electricity-load-could-grow-128-gw-over-next-5-years-Grid-Strategies/734820.
15. "550-650 TWh Forecast Analysis." Accessed 2026-07-07. https://www.useluminix.com/reports/industry-analysis/powering-the-ai-boom-where-the-grid-breaks-first-2026-2030.
16. Hanwha Data Centers. "Data Center Grid Limitations: The Power Bottleneck." Blog post. Accessed July 7, 2026. https://www.hanwhadatacenters.com/blog/data-center-grid-limitations-the-power-bottleneck/.
17. "Power Requirements for AI Data Centers (2026): Complete Guide." Accessed 2026-07-07. https://techplustrends.com/power-requirements-ai-data-centers/.
18. Saameer. "The 2026 Infrastructure Guide." Tech Plus Trends. Accessed July 7, 2026. https://techplustrends.com/1gw-data-center-power-consumption-guide/.
19. Goldman Sachs Research. "AI to Drive 165% Increase in Data Center Power Demand by 2030." Goldman Sachs Insights. Accessed July 7, 2026. https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030.
20. Avid Solutions Inc. "13 Data Center Growth Projections That Will Shape 2026-2030." Accessed July 7, 2026. https://avidsolutionsinc.com/13-data-center-growth-projections-that-will-shape-2026-2030.
21. "The Binding Constraint: Current and Near-term Limitations on Data Center Development." Accessed 2026-07-07. https://www.powermag.com/webinars/the-binding-constraint-current-and-near-term-limitations-on-data-center-development/.
22. DataCenter Dynamics. "IEA: Data center energy consumption set to double by 2030 to 945TWh." Accessed July 7, 2026. https://www.datacenterdynamics.com/en/news/iea-data-center-energy-consumption-set-to-double-by-2030-to-945twh/.
23. S&P Global Energy. "Global data center power demand to double by 2030 on AI surge: IEA." Accessed July 7, 2026. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea.
24. Hade, Tim. "The Fastest Way to Ramp Compute Isn't a Power Plant Next to Every Data Center." Steel for Fuel (Substack), 2026. https://steelforfuel.substack.com/p/data-centers-need-the-grid-so-we.
25. International Energy Agency. "Key Questions on Energy and AI: Executive Summary." Report. Paris: IEA, 2025. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. Accessed July 7, 2026.
26. International Energy Agency. Energy and AI: Executive Summary. Paris: IEA Publications, 2024. https://www.iea.org/reports/energy-and-ai/executive-summary.
27. Novogradac & Company LLP. "Resolving the Interconnection Queue Bottleneck Along with Transmission Expansion is Critical for Timely U.S. Energy Deployment to Meet Demand." Notes from Novogradac (blog). Accessed July 7, 2026. https://www.novoco.com/notes-from-novogradac/resolving-the-interconnection-queue-bottleneck-along-with-transmission-expansion-is-critical-for-timely-us-energy-deployment-to-meet-demand.
28. Rocky Mountain Institute. "The Interconnection Queue Continues to Be a Barrier to US Economic Competitiveness." Accessed July 7, 2026. https://rmi.org/resources/interconnection-reform-ai-data-centers-generator-queues.
29. Loopback Systems. "Lawrence Berkeley Lab 2025 Interconnection Queue Report | Full Overview." Loopback Systems Insights. Accessed July 7, 2026. https://www.loopbacksystems.com/insights/lawrence-berkeley-2025-interconnection-queue-report-overview.

Editorial Perspective

An analytical reading generated by drs7 from the finished report — interpretive, drawn from the document's own content, and not part of the sourced findings above.

Non-obvious insights

1. The report's repeated source-provenance warnings — flagging that IEA figures 'appear across multiple source documents that share a common upstream origin' and that the Luminix synthesis 'represents one analytical voice, not independent corroboration' — collectively suggest that the apparent convergence of major forecasts around 550–800 TWh is largely an artifact of citation recycling rather than genuine analytical independence. The consensus is thinner than its surface breadth implies.
2. The behind-the-meter finding carries a structural implication the report never fully draws out: if 25–38% of incremental data-center load goes behind the meter, the grid-served increment falls to roughly 350–480 TWh — which means grid infrastructure investors and ratepayers could bear stranded-cost risk even if the underlying AI demand surge is entirely real. Demand materializing is not the same as grid demand materializing.
3. The report documents that PJM's 2027/2028 capacity auction cleared at the FERC cap but still fell 6,517 MW short of the reliability requirement — the first such shortfall exceeding one percentage point in PJM history. This is mentioned in the context of constraint analysis but its deeper implication is never named: the grid is already showing reliability stress before the projected AI demand surge has materialized, meaning the system may be fragile even on the low end of the demand range.
4. Natural gas producers and midstream operators are identified as 'near-term winners regardless of which long-run demand scenario materializes' because behind-the-meter gas fills the gap while interconnection queues clear. This means the energy sector's near-term financial beneficiary of the AI boom is structurally insulated from the forecast uncertainty that makes every other value claim in the report scenario-contingent — a concentration of scenario-robust return that the report names but does not emphasize as an asymmetric investment signal.
5. The Limitations section discloses that the EIA's 2024 attempt to collect cryptocurrency mining data ended with a court-ordered agreement to destroy what had been gathered. This implies that mandatory data collection on energy-intensive digital workloads faces not just legislative absence but active legal precedent working against it — making the 'addressable' category of limitations (mandatory reporting) substantially harder to achieve than the report's framing suggests.

Tensions and contradictions

The report's central tension is unresolved and self-aware: it defends a base case of 550–650 TWh while repeatedly documenting that the methodological and evidentiary basis for any figure above the LBNL floor is weak. It cannot simultaneously argue that the high end is 'not investable as stated' and that the base case itself is defensible, because the base case is constructed by blending EPRI medium, McKinsey, and LBNL high — sources the report elsewhere critiques for methodological opacity, promotional framing, and shared upstream origins. A secondary tension runs through the efficiency section: the report argues that the 2010–2018 efficiency lever (enterprise-to-hyperscale migration) has been 'pulled' and is no longer available, weakening the bear case — but it also argues that algorithmic efficiency (the DeepSeek dynamic) is a new and potent mechanism. These two claims together undercut both the bear and the bull cases without resolving which effect dominates, leaving the efficiency trajectory assumption genuinely open in a way the report acknowledges but cannot escape. A third tension: the report's treatment of LBNL as the most credible institutional voice coexists with the admission that LBNL's own model 'requires utilization-rate and GPU-shipment assumptions that are rapidly evolving and difficult to forecast' — meaning the anchor of the entire range is itself assumption-sensitive in ways the report cannot bound.

The bottom line

Do not underwrite the headline consensus number; underwrite the physical delivery constraint instead. The report's most durable finding — confirmed across the largest number of independent evidentiary lines — is that interconnection clearance timelines (now exceeding eight years from PJM application to commercial operation) set a harder ceiling on energized demand by 2030 than any demand forecast does. An investor whose return depends on AI electricity demand reaching the grid at the magnitudes utilities are forecasting is exposed to a physical and regulatory bottleneck that capital alone cannot shorten. The actionable implication is to treat interconnection-delivery risk as the primary investment variable, not demand magnitude — meaning assets with already-secured grid access, signed interconnection agreements, and sited power are worth a structural premium over announced-but-unconnected projects regardless of where in the 450–810 TWh range demand ultimately lands.

Open questions

The report raises but never answers what project-level FERC milestone data would actually show if systematically analyzed: specifically, what fraction of the 2023–2025 data-center interconnection cohort has progressed beyond the study phase to deposit-backed, site-controlled, construction-ready status. That single empirical question would allow the 'speculative to committed' ratio to be estimated from observable filings rather than from the Grid Strategies 65% benchmark, which is itself a single-source figure. The report also surfaces but cannot answer whether the Jevons-paradox expansion of agentic and multi-step workloads will outpace per-task efficiency gains — a question that requires utilization telemetry from deployed inference systems that does not exist in the public record and that hyperscalers have no obligation to provide. Most consequentially, the report never confronts what would happen to the entire investment thesis if AI revenue economics deteriorate before the infrastructure build completes: the $720 billion in grid spending and $1 trillion-plus in data-center capex are documented, but whether the AI applications generating demand for that infrastructure will produce returns sufficient to sustain it is explicitly listed as a coverage gap and never revisited. That is the bubble question the report frames but refuses to execute.

About this report

AI contribution disclosure. Section prose was machine-drafted from the curated source manifest identified above. The operator authored the research question, curated the source corpus, defined the outline, and approved or revised the result; responsibility for the content rests with the operator, not the drafting system.

• Citation integrity: 94.1% [89.5–96.8% 95% CI] (153 supported / 12 partial / 4 unsupported of 169 cited sentences; audited by an independent model from a different provider than the drafter)