Blog

Data Center Electricity: From Cloud Efficiency Era to AI Load Growth

Between 2008 and 2021, total US electricity consumption was broadly stagnant, rising from roughly 3,873 TWh to 4,116 TWh over thirteen years, under 0.5% per year on average, a marked contrast with prior decades of steady electrification-driven growth.^[3] Efficiency gains in appliances, buildings, and industrial processes offset much of the new load, leaving utilities planning around a low-growth environment. Data centers were an exception: LBNL estimates their electricity use rose from approximately 58 TWh in 2014 to 76 TWh by 2018, but computing output grew roughly tenfold over that same period. The shift from under-utilized enterprise servers to dense hyperscale racks, server virtualization, and improved power usage effectiveness (PUE) absorbed most of the workload growth without proportional power increases. This efficiency era kept the industry's power footprint broadly in check even as workloads exploded.^[1]

That efficiency offset began eroding after 2018 and broke definitively after 2022. The hyperscale cloud buildout of 2017–2020 was the first acceleration in raw power terms. Pandemic-era digital demand amplified it from 2020 through 2022. The commercialization of large language models from late 2022 onward triggered a third and qualitatively different wave: capital spending cycles in AI-specific infrastructure, GPU clusters, and purpose-built training and inference facilities that operate at power densities an order of magnitude higher than conventional cloud servers. Unlike the efficiency era, where more computing was delivered for roughly the same power, AI workloads consume power in near-direct proportion to model size and inference volume. LBNL estimates US data center electricity use grew at approximately 18% per year between 2018 and 2023, reaching a measured anchor of 176 TWh in 2023, which represented 4.4% of total US electricity consumption.^[1][15]

For context: 176 TWh is roughly equivalent to the annual electricity consumption of Australia. The IEA's 2030 global projection of 945 TWh is approximately equal to Japan's current annual electricity consumption. These are not marginal additions to the world's electricity system. They represent the emergence of a new industrial load category comparable in scale to major existing sectors, appearing within a much shorter timeframe than those sectors took to develop.^[1][2]

The demand growth story also contains an important internal tension. AI hardware efficiency is improving rapidly, but the public benchmark claims are workload-specific: NVIDIA reports up to 25× better energy efficiency for GB200 NVL72 versus H100 infrastructure on a defined LLM inference configuration, while training gains are closer to 4×.^[18] The exact improvement depends on model architecture, precision, utilization, networking, cooling, and deployment design. The question is whether efficiency gains at the chip and system level translate into lower total power consumption, or whether they reduce the cost of AI output, stimulate more demand, and ultimately increase total consumption. This is Jevons paradox, and it is the primary source of uncertainty in the lower bound of any demand forecast. Evidence so far, particularly the market reaction to DeepSeek-R1's January 2025 release and reported efficiency claims, suggests that cheaper AI inference can accelerate adoption faster than it reduces aggregate power demand.^[19]

The Gap Between Corporate Claims and Physical Grid Reality

Two concepts are routinely conflated when discussing how AI data centers are powered, and separating them is essential for any clear analysis. The physical fuel mix describes what actually generates the electricity serving the servers at any given hour: the combustion source, grid region, and dispatch stack at that specific moment. The contractual fuel mix describes what companies report in sustainability disclosures through renewable energy certificates and power purchase agreements, which can differ substantially from the physical reality. A third concept also matters: the architectural mix, meaning whether a campus draws from the shared grid, operates behind-the-meter gas generation, or combines co-located renewables, storage, and backup. All three dimensions shape a campus's actual energy footprint, and all three are discussed below.

The contractual mix is what appears in corporate ESG reports. Through renewable energy certificates (RECs) and power purchase agreements (PPAs), technology companies can match their reported electricity consumption with clean-energy purchases. Google, Microsoft, Amazon, and Meta each claim high or complete renewable matching on a contractual basis. These claims are legally valid under standard corporate accounting frameworks (GHG Protocol Scope 2). They do not describe the physical generation supplying the servers at any specific hour.

When a data center draws electricity from the regional grid, it receives whatever combination of generation sources is dispatched onto that grid at that moment. In Northern Virginia, which hosts the world's most concentrated data center cluster, including the core of Amazon's us-east-1 region, Microsoft's largest Azure footprint, and major Google and Meta facilities, the physical mix is gas-heavy. At 2 AM when a large AI training run starts, the electrons powering it come from whatever source is on the dispatch stack at that moment, which is frequently a combined-cycle natural gas plant. IEA data estimate that natural gas supplies roughly 40–42% of electricity physically serving US data centers, compared with approximately 24% renewables, 20% nuclear, and 15% coal.^[2]

The gap between contractual and physical reality is not hypocrisy. It is a structural reflection of where the electricity system currently stands. Always-on clean power sources capable of serving 24/7 AI workloads (nuclear, geothermal, firm storage-backed renewables) do not yet exist at the scale required to physically power the fleet. The PPA and REC market creates financial incentives to build new renewable capacity and provides a signal of corporate demand. But it does not change what electrons are physically powering the servers today.

The nuclear PPAs, geothermal contracts, and SMR agreements being signed now are the mechanism by which hyperscalers intend to close the contractual-to-physical gap over the next decade. Whether they succeed on the announced timelines, which this report argues they will not fully meet, determines whether AI infrastructure becomes a material contributor to decarbonization or a structural carbon liability through the 2030s.

Four Architectures, Different Trade-Offs

Natural gas appears throughout the AI energy story in forms that are technically and operationally distinct. A grid-connected data center drawing power from a regional utility, an on-site simple-cycle turbine, a utility combined-cycle plant, and a fuel-cell installation all use gas, but in ways that differ substantially in speed, efficiency, water use, local emissions, and infrastructure risk. Conflating them produces confusion about what is actually being built and what its consequences are.

Three Constraints, Different Developers Affected Differently

Industry analysts and grid operators have cited figures of 30–50% of planned data center projects facing meaningful delays, a range drawn from interconnection queue data, project tracker surveys, and utility forecasts.^[10][13] The figure is directionally credible but analytically incomplete without decomposing its causes. Three distinct constraints are at work, each affecting different segments of the market in different ways and requiring different responses. Treating them as a single "power problem" obscures which constraint actually applies to which project.

Bottleneck 1: Grid interconnection

Getting a new large electricity load connected to the regional transmission network currently takes 3–7 years in the US. PJM, the grid operator serving 65 million people across 13 states, has seen interconnection queue costs rise approximately seven times since 2023; a project applying today cannot reduce capacity payments before 2033.^[10][13] In Ireland, EirGrid has imposed a hard cap on new large data center connections in the Dublin region, meaning Google, Microsoft, and Amazon cannot expand their Irish footprint regardless of available capital. This is the clearest documented case of power physically stopping hyperscaler expansion.

The response has been the move to behind-the-meter generation: build on-site gas turbines or fuel cells and bypass the interconnection queue entirely. This structural shift is driving the gas turbine buildout at Meta, xAI, and Stargate; the fuel cell deployment at Oracle; and the co-located solar plus storage model at Google's Texas campus. Projects that cannot secure either a grid connection or viable on-site generation are genuinely stopped, not merely delayed.

Bottleneck 2: Community and regulatory opposition

Community opposition has transitioned from a project-by-project irritant to a structural constraint on the AI infrastructure buildout. As of early June 2026, 188 organized opposition groups were active across 40 states according to Data Center Watch (10a Labs); the broader tracking suggests total active groups had grown to approximately 833 across 49 states by the end of Q1 2026, more than doubling in three months.^[14] In Q1 2026 alone, at least 75 projects worth approximately $130 billion were blocked or delayed, equaling the total for all of 2025 in a single quarter.^[14]

The PW Digital Gateway project ($24.7B, Virginia) provides an instructive case study: legally cancelled not because of power unavailability, but because Prince William County failed to properly advertise a required public hearing. Three lawsuits from national conservation groups (National Parks Conservation Association, American Battlefield Trust) citing Civil War battlefield heritage impact sustained the cancellation through appeal. The primary cause was procedural and heritage-related, not energy-related.

The underlying community grievances are substantive. Electricity costs have risen 42% since 2019 in some markets. PJM's own modeling found that removing all data center demand from its forecasts would reduce capacity payments by $9.33 billion, a 64% reduction,^[13] indicating that residential and commercial ratepayers are bearing a portion of the grid infrastructure costs that large data center loads generate. The White House facilitated a Ratepayer Protection Pledge in March 2026, under which major AI companies committed to directly fund necessary grid infrastructure improvements. Whether this pledge proves enforceable and how it affects project economics are open questions.

Bottleneck 3: Equipment supply chains

Transformer demand increased 119% between 2019 and 2025, while manufacturing capacity has not kept pace.^[10] Lead times for large power transformers have stretched from 24–30 months to 2–4 years. Gas turbine slots at GE Vernova are sold out through 2030 at current booking rates.^[8][10] Cooling systems, switchgear, and high-voltage equipment all face extended lead times because demand across the entire infrastructure stack accelerated simultaneously. Every electron reaching a data center passes through a transformer; the transformer bottleneck is often a longer timeline constraint than the generation source itself.

Restart, SMR, and the Reality of Timeline Risk

Nuclear power commands the highest capacity factor of any generation source (92–95%), produces zero CO₂ during operation, and can deliver 800–1,600 MW from a single facility, matching the power requirements of even the largest planned AI superclusters through one contract, one counterparty, and one set of regulatory relationships. These properties make nuclear theoretically the closest match to AI data center power requirements. The question is not whether nuclear is theoretically attractive. It clearly is. The question is whether it will actually arrive on the timelines announced in press releases.

How nuclear PPAs actually function

When Microsoft signed a 20-year PPA with Constellation for Crane Clean Energy Center (Three Mile Island Unit 1), the contractual arrangement works as follows: the plant generates 835 MW into the PJM shared grid. Microsoft pays a fixed price per MWh through a Contract for Difference. If the PJM market price is below the agreed price, Microsoft pays Constellation the difference; if above, Constellation rebates the difference to Microsoft. The net effect is that Microsoft pays a fixed price regardless of market movements and receives a clean energy certificate for every MWh generated. No wire connects the reactor to Microsoft's data centers. The plant serves the shared grid; the PPA is a financial and attribution instrument.^[9]

Three Mile Island Unit 1: status as of mid-2026

The plant (Unit 1, entirely separate from Unit 2 involved in the 1979 accident, which was permanently closed) shut down in 2019 because natural gas made it economically unviable. Microsoft's PPA changed that economics. As of mid-2026, refurbishment is over 65% staffed; major systems including the main generator and turbines have been successfully tested. Target commercial operation is 2028, with Constellation noting it could be as early as 2027 if an early interconnection request with PJM is approved. As of mid-2026, more than 65% staffed and major systems successfully tested. Not yet producing power; operational targets remain subject to regulatory process.

SMR timeline risk: five independent constraints

Every hyperscaler SMR contract in the current pipeline is for a reactor not yet built in the Western world by a company that has never built one commercially. The 45 GW global offtake pipeline (IEA, April 2026)^[2] represents demand certainty from the buyer side, not engineering certainty from the supply side. Five constraints must each be resolved for SMRs to arrive on the 2030–2032 timelines being promoted.

Regulatory pathway. Advanced non-light-water designs such as Kairos Power's molten salt, TerraPower's sodium-cooled Natrium, and Oklo's microreactor all require the NRC to develop entirely new regulatory frameworks. NuScale, a conventional light-water design with the most familiar regulatory profile, took six years to obtain design certification from the NRC and still cancelled its commercial project in 2023 when costs rose from $58/MWh to $89/MWh. Oklo had its first NRC license application rejected in 2022; as of mid-2026 it had not yet filed a commercial license application.

HALEU fuel supply. Most advanced SMR designs require High-Assay Low-Enriched Uranium (HALEU), enriched to 5–20% vs. 3–5% for conventional light-water reactors. The US currently produces approximately 5 metric tonnes per year from a single Centrus facility in Piketon, Ohio, enough for approximately 11 Xe-100 reactors in total. A new domestic enrichment facility would produce first output around 2031 and cost approximately $5 billion. The fuel supply chain for the SMR pipeline being contracted today does not exist at the required scale.

First-of-kind cost risk. NuScale's project cancelled after costs rose 53%. Vogtle Units 3 and 4, conventional large reactors built to already-approved designs in Georgia, overran by approximately $17 billion and seven years. First-of-kind construction in nuclear consistently costs more and takes longer than estimated. The factory-built modular premise underpinning SMR economic projections has never been tested in the Western world at commercial scale.

Supply chain atrophy. Nuclear-grade heavy forgings, pressure vessels, and specialized components require manufacturing capability that contracted substantially during three decades of minimal new nuclear construction. Rebuilding that capability is occurring simultaneously with demand for it, creating a constraint that cannot be resolved quickly regardless of funding.

Workforce. The nuclear engineering and operations workforce shrank significantly during the same period. Training qualified nuclear operators, inspectors, and engineers to the required standards takes years and cannot be compressed.

The Underappreciated Option: Operating Plants, Project Financing, and a Public Company

Of all the clean energy technologies covered in this report, enhanced geothermal systems have undergone the most dramatic change in commercial status over the shortest period, while receiving proportionally the least analytical attention. Two years ago, EGS was a pre-commercial technology at Technology Readiness Levels 4–6. As of mid-2026, it has operating plants with over 600 days of continuous operation, non-recourse project financing from a top-tier banking syndicate, a publicly listed company (Fervo Energy, NASDAQ: FRVO) that raised $2.2 billion in its May 2026 IPO, and a 3 GW framework agreement with Google signed in March 2026.^[6]

How EGS works

Conventional geothermal requires hot water naturally present near the surface, typically in volcanic regions such as Iceland, California's Geysers, and Nevada. EGS creates an artificial reservoir in hot dry rock anywhere the subsurface is sufficiently hot. Two wells are drilled 4–5 kilometres down; water is injected under pressure to fracture rock between them; cold water pumped down one well emerges superheated from the other and drives a turbine. The rock is always hot. The loop is closed and essentially continuous. Zero fuel cost once established; no combustion emissions; water circulates in a closed loop, consuming approximately 0.1–0.3 liters per kWh.

Proven milestones and remaining risks

Project Red in Nevada, Fervo's 3.5 MW pilot with Google, has operated continuously for over 600 days, validating EGS physics at commercial field scale. Cape Station's well tests have achieved 16 MW per production well, triple Project Red's output and exceeding levels the National Renewable Energy Laboratory did not expect until 2035, a decade ahead of schedule. Fervo reports reducing drilling time by approximately 70% between Project Red and Cape Station, applying oil-and-gas drilling techniques including horizontal drilling and multistage stimulation.

The non-recourse project financing for Cape Station Phase I, underwritten by RBC, Barclays, HSBC, and JPMorgan without Department of Energy loan program backing, is a commercially significant milestone: it means the banks assessed the project's subsurface risk, revenue contracts, and technology as sufficient to underwrite at project level. This is not proof that every future EGS site will perform similarly; each well is a partial subsurface experiment, but it establishes that institutional capital now considers EGS bankable rather than speculative.

The central remaining risk is geological replicability. Cape Station's exceptional well performance at a specific Utah site does not guarantee similar results at other sites. Induced seismicity is a documented risk: the Pohang geothermal project in South Korea triggered a magnitude 5.5 earthquake in 2017, linked directly to EGS stimulation. Well costs of $10–20 million each mean that a site with unexpected subsurface conditions can absorb significant capital before viability is confirmed.

Bloom Energy's SOFC: A Speed Solution With an Unresolved Carbon Question

Bloom Energy's solid oxide fuel cells are among the fastest-deployable primary power options for data centers. The modular 325 kW boxes can be installed more quickly than a new utility interconnection or a combined-cycle plant, require no grid connection, use zero operational water at the point of generation, and produce very low local air pollutants compared with combustion turbines. The fundamental trade-off is fuel: every commercial SOFC deployment for data centers today runs on natural gas, and the pathway to lower-carbon operation depends on technologies and markets that do not yet exist at commercial scale.

How fuel cells differ from gas turbines

Both architectures use natural gas delivered by pipeline. The difference is the conversion process. A gas turbine combusts the gas, with hot expanding gases spinning turbine blades. A Bloom SOFC does not combust; it runs the gas through an electrochemical reaction at 700–900°C that converts it directly to electricity. This eliminates NOx, SOx, and particulate emissions, produces a concentrated CO₂ exhaust stream better suited to potential carbon capture, requires zero operational water, and achieves 55–65% electrical efficiency versus 35–40% for a simple-cycle gas turbine. The trade-offs are higher cost per MWh, a slow cold restart measured in hours rather than seconds, and significant logistical complexity when assembling hundreds of 325 kW boxes for very large installations.

Scale and carbon accounting dynamics

Oracle's initial 1.2 GW SOFC order and master agreement supporting up to 2.8 GW is the largest single-customer commitment in Bloom's history and the clearest demonstration of hyperscale deployment intent. AEP has an agreement for up to 1 GW, with an initial 100 MW order. Equinix has deployed Bloom across 19 US co-location facilities, making it the largest deployment by site count and demonstrating the technology at the 5–10 MW per site scale typical of co-location infrastructure.

A notable pattern in the current deployment landscape: the four largest hyperscalers (Google, Microsoft, Amazon, and Meta) have limited direct public Bloom contracts despite using fuel cells indirectly through utilities and co-location providers. A plausible explanation is carbon accounting. These companies carry public commitments to 100% renewable electricity. A large direct natural-gas-fed fuel-cell contract would make gas a more visible component of their disclosed power strategy. Oracle, facing less legacy pressure from public clean-energy pledges, moved first and most decisively. The others can access fuel cell capacity through indirect channels while maintaining lower direct carbon-accounting visibility.

Decarbonization pathways

Three potential pathways exist for reducing the carbon intensity of current SOFC deployments. Carbon capture is the most realistic near-term option: Bloom's exhaust is approximately ten times more CO₂-concentrated than gas turbine exhaust, making capture economically more viable. A pilot with Chart Industries is underway, with potential commercial availability 2027–2029. Biogas blending is viable now without hardware change, but US biogas production equals approximately 2.5% of gas supply, insufficient to serve more than a small fraction of current installations. Green hydrogen, while theoretically allowing zero-CO₂ operation using the same hardware, requires green hydrogen at the Department of Energy's target price of $1/kg; current cost is approximately $4–6/kg, the US hydrogen hub program was largely defunded in October 2025, and the IEA cut its 2030 green hydrogen forecast 25% in 2025. Green hydrogen for large-scale SOFC operation is a post-2031 scenario at earliest under current policy conditions.

Coverage, Residual Firm Power, and What Storage Cannot Do

Battery storage is the fastest-growing segment of US electricity infrastructure. The installed base reached approximately 27 GW by end of 2024, growing 68% in a single year, with 20+ GW more projected for 2026.^[4] Storage is genuinely transforming the grid's ability to integrate variable renewables. What it cannot do, at any currently available scale or cost, is fully substitute for firm generation when multi-day or seasonal weather events reduce solar and wind output simultaneously, which they do every year, in every climate.

The reliability standard AI data centers must meet, typically 99.999% uptime, meaning less than five minutes of unplanned downtime per year, requires power during every weather scenario, including scenarios that last longer than any currently commercial battery system. Lithium-ion batteries at 4–8 hours of storage address the standard overnight gap after solar output ceases, buffer short wind lulls, and support grid arbitrage. They do not address multi-day events.

Form Energy's iron-air battery (100-hour storage) is the most promising long-duration technology, targeting commercial deployment from 2028 onward at approximately $20/kWh. A 100-hour battery covers multi-day weather events that lithium-ion cannot. But a sustained winter storm produces 168+ hours of reduced solar and wind output. A Dunkelflaute (the German term for simultaneous low-sun and low-wind periods occurring in winter) can last 10–14 days (240–336 hours) across Central Europe. For a data center requiring 99.999% uptime, 100-hour storage reduces the number of hours per year requiring firm backup from approximately 6,000 (with no storage) to perhaps 200–500 hours, meaningful progress that does not eliminate the firm power requirement.

The useful planning framework is capacity credit: long-duration storage dramatically reduces the fraction of installed capacity that must come from always-on firm sources, shifting the system from requiring 100% firm backup to perhaps 5–15% firm backup for worst-case weather scenarios. This changes portfolio economics significantly without eliminating the nuclear, geothermal, or gas backup requirement entirely.

Two Separate Problems, One Hidden Constraint on Technology Choice

Data centers face two distinct water consumption challenges that are routinely conflated. Understanding them separately is essential for technology selection and site planning, because the same site can face very different water constraints depending on which of the two problems dominates.

Power generation water use

The electricity generation process itself consumes water, primarily for cooling steam cycles at the power plant. This is the generation-side water problem. Nuclear power uses approximately 2.1 liters per kWh if wet-cooled, the highest consumption of any always-on clean source. Wet-cooled utility CCGTs use approximately 1.2 L/kWh. On-site simple-cycle gas turbines are typically air-cooled and use near-zero water. Bloom SOFC fuel cells use zero operational water because there is no steam cycle or cooling tower. EGS geothermal uses roughly 0.1–0.3 L/kWh in a closed loop where the same water recirculates continuously. Solar and wind use effectively zero water at the point of generation.

This creates a counterintuitive ranking: nuclear, the cleanest always-on power source by CO₂ emissions, is also the thirstiest if wet-cooled. In water-stressed regions such as Nevada, Utah, New Mexico, and Arizona, where EGS geothermal is being developed and where SMRs might be sited, water availability is a genuine constraint that may determine which clean technologies can actually be deployed, regardless of which ones are theoretically preferred.

Oracle's choice of Bloom fuel cells over gas turbines for Project Jupiter in Doña Ana County, New Mexico, is the clearest documented case of water being a decisive technology selection factor. Bloom uses zero water. A wet-cooled CCGT utility plant would use approximately 1.2 L/kWh. In one of the most water-stressed counties in the US, that difference is operationally significant and appears to have been a primary driver of the fuel cell selection.

Server cooling water use

Separately from power generation, data centers consume large quantities of water to cool the servers themselves, typically through evaporative cooling towers. This accounts for the majority of on-site direct water consumption. LBNL estimated direct US data center water use at approximately 17 billion gallons in 2023, with 2028 scenarios that could roughly double to quadruple depending on demand growth, cooling design, and location.^[15]

Critically, this server-cooling water consumption is independent of the power source. A nuclear-powered data center and a gas-powered data center with identical server infrastructure consume the same water for server cooling. The growing adoption of liquid cooling (including direct liquid cooling, glycol loops, and immersion), driven by the thermodynamic necessity of handling NVIDIA Blackwell and successor GPU racks at 120–140 kW and above, reduces this on-site water consumption by eliminating evaporative cooling towers. A fully liquid-cooled AI campus can use dramatically less water than a conventionally air-cooled facility, regardless of what powers it.

Eight Dimensions Underweighted in Standard AI Energy Analysis

Standard AI energy narratives tend to collapse training and inference into a single undifferentiated load, conflate contractual clean-energy claims with physical power, understate water and ratepayer distributional effects, and treat national electricity supply as unconstrained by local grids, permits, and community acceptance. Eight specific dimensions receive systematically less attention than their analytical importance warrants.

1. Training versus inference: two different power problems

AI power demand is not a single homogeneous load. Training, which involves building a large model from scratch or fine-tuning it on proprietary data, runs for weeks on thousands of GPUs, drawing enormous sustained power in concentrated bursts at specific campuses. A large frontier training run can consume as much energy as several thousand homes use in a year, concentrated in a single facility over weeks. Inference, meaning running a trained model to answer queries, is continuous, distributed, and lower power per operation but multiplied by billions of daily interactions. The power profiles are qualitatively different. Training demands concentrated firm power in specific clusters for finite periods. Inference demands distributed capacity close to users globally with 24/7 reliability requirements. Agentic AI workloads, where AI autonomously completes multi-step tasks over extended periods, are emerging as a third category with intermediate power characteristics. Grid planning for AI requires separating these load types, not just tracking aggregate TWh.

2. Jevons paradox: efficiency may increase total consumption

When AI becomes more energy-efficient per operation, the cost of AI output falls, which increases demand for AI output, which may increase total energy consumption even as consumption per operation drops. This is Jevons paradox, named after a nineteenth-century economist who observed that more efficient coal engines led to greater total coal consumption. DeepSeek's January 2026 demonstration of dramatically more efficient model training caused a brief correction in energy stocks on the assumption that cheaper inference would reduce power demand. Most analysts concluded within weeks that cheaper inference would accelerate adoption faster than it would reduce per-unit consumption, potentially increasing total demand. This dynamic is the primary source of downside risk to the lower bound of demand forecasts: if efficiency gains are large but adoption growth is larger, TWh keep climbing regardless of compute-per-watt improvements.

3. Ratepayer and affordability risk

When a 500 MW data center campus connects to a regional grid, it typically triggers costs for new transmission infrastructure, upgraded substations, additional generation capacity, and interconnection studies. These costs are generally allocated across all ratepayers in the region through capacity and transmission charges, not exclusively to the data center that caused them. PJM's own analysis found that removing all data center demand from its forecasts would reduce capacity payments by $9.33 billion, a 64% reduction. Electricity costs have risen 42% since 2019 in some markets where data center concentration is highest.^[13] Democratic senators wrote to seven major technology companies in December 2025 to investigate the effects of their operations on consumer energy bills. The White House's March 2026 Ratepayer Protection Pledge represents the first formal policy response to this dynamic. The social legitimacy of AI's energy consumption depends significantly on whether ordinary households perceive that they are subsidising it.

4. Transmission and distribution: not just generation

The bottleneck in the AI energy race is not only "where does the electricity come from?" It is equally "how does electricity get from the generator to the building?" Transformers, substations, switchgear, high-voltage transmission lines, and distribution network upgrades are all required for every new large load connection. Transformer demand has increased 119% from 2019 to 2025, but manufacturing capacity has not kept pace, and lead times have stretched to 2–4 years.^[10] Every electron reaching a data center passes through a transformer. Gas pipelines face parallel capacity constraints for behind-the-meter gas generation: Oracle's Project Jupiter at 2.45 GW requires a transmission-grade gas pipeline from the Permian Basin supplying approximately 122 million MMBtu/year, a pipeline that must be permitted, financed, and constructed in parallel with the fuel cell deployment.^[7] Physical delivery infrastructure is frequently the longer constraint than the generation source itself.

5. Natural gas infrastructure lock-in

If gas turbines and gas-fed fuel cells bridge the 2025–2032 gap between current demand and available clean firm power, which the evidence strongly suggests they will, the AI industry is simultaneously creating large and durable natural gas infrastructure dependencies. Gas pipelines to serve behind-the-meter turbines and fuel cells require permitting, construction, and operational commitments measured in decades. Methane leakage from gas infrastructure, particularly from gathering lines, compressor stations, and distribution systems, can add 50–100% to the effective greenhouse gas impact of natural gas generation on a 20-year timescale, depending on the leakage rate assumed. Fuel price exposure accumulates across all gas-dependent installations: when European gas prices doubled in 2021–2022, the operating cost of gas-fed electricity generation doubled accordingly. This commodity risk, absent from nuclear, geothermal, or renewables once built, is a structural feature of the gas bridge strategy that rarely appears prominently in capital project assessments.

6. Geography: not a national story

The AI energy story is not national. It is a collection of highly localized situations with completely different characteristics. Northern Virginia has the world's most concentrated data center cluster, a gas-heavy grid, severe interconnection congestion, and organized community opposition. Iowa has abundant wind energy making data center operations there physically cleaner than anywhere else in the continental US on a grid-mix basis; Meta Altoona and Google Council Bluffs may be the most sustainably powered large AI infrastructure in the world on a physical, not merely contractual, basis. Nevada and Utah have EGS geothermal potential and existing renewable infrastructure. New Mexico has sun, EGS potential, and severe water stress. Texas has abundant gas, wind, solar, and the independent ERCOT grid allowing faster interconnection than PJM or MISO. Oregon's Columbia River hydropower makes it one of the cleanest existing grid regions for data centers. None of these situations generalizes to the others. Every infrastructure decision is shaped by local grid mix, water availability, geology, permitting environment, utility relationships, and community politics.

7. Community opposition as a structural constraint

Local opposition over noise, water use, electricity prices, land use, visual impact, backup generator emissions, and grid impacts has evolved from a communications challenge into a material operational and financial constraint. The scale is significant: as of Q1 2026, 833 active opposition groups operated across 49 states, more than double the number six months earlier.^[14] In Q1 2026 alone, at least 75 projects worth approximately $130 billion were blocked or delayed, matching the full-year total for 2025 in a single quarter. At least 69 local government jurisdictions enacted bans or moratoriums as of May 2026. Seattle passed a one-year pause affecting five proposed projects. The PW Digital Gateway, a $24.7 billion project, was legally cancelled due to procedural opposition from heritage conservation groups, not for energy reasons. Community opposition is structurally analogous to the "social licence to operate" challenges that have constrained mining and fossil fuel infrastructure for decades. It is arriving in AI infrastructure at scale for the first time in 2025–2026, and its tools (environmental review, procedural challenge, elected official accountability, and national advocacy organization involvement) are increasingly effective.^[14]

8. The geopolitical dimension

The AI energy race has a geopolitical layer that standard analyses underweight. China is building nuclear, renewables, transmission, and energy manufacturing capacity at a pace that changes the global cost comparison. Chinese data centers operate on a grid that is coal-heavy today but is adding renewable and nuclear capacity faster than any other country. Export controls on advanced GPUs may redirect Chinese AI development toward domestic accelerators; the energy consequences depend on chip efficiency, software stack maturity, and workload mix, not simply on the country of origin. The strategic point is broader: the country or company that achieves firm, scalable, low-carbon power for AI infrastructure first gains more than a sustainability profile. It gains a lower long-run operating cost structure for training, inference, and future AI workloads that may compound into structural competitive advantage over decades.

What the Evidence Actually Says

The AI energy story is simultaneously more urgent, more complex, and more resolvable than most public commentary suggests. It is more urgent because the demand numbers are large, the inflection is already underway, and the physical infrastructure being built now will operate for 20–25 years regardless of what clean alternatives emerge after 2030. It is more complex because no single energy source solves the problem: each technology brings trade-offs across cost, reliability, water, carbon, timeline, geography, and community acceptance that make a universal answer impossible. And it is more resolvable than pessimistic readings suggest, because several credible clean pathways, EGS geothermal, nuclear fleet restarts, advanced fuel cells, long-duration storage, have moved from theoretical to early-commercial within the last two years.

Understanding the current energy mix requires separating what companies report from what electrons are physically doing. Technology companies are not misrepresenting their positions when they claim high renewable matching, but those claims describe financial instruments and contractual offsets, not the generation mix dispatched to serve their servers at any given hour. The electrons physically serving AI data centers today are roughly 55% gas-dependent on a global basis. That context matters for evaluating decarbonization timelines, but it does not undermine the genuine progress that PPA markets and corporate clean-energy commitments represent. Long-term PPAs fund new clean capacity; 24/7 CFE matching programs create incentives to develop always-on clean sources. The gap between current physical reality and stated goals is the problem these programs are designed to close, and the trajectory of EGS geothermal, nuclear restarts, and advanced fuel cells suggests real progress is accumulating.

The near-term constraints are real and distinct, and treating them as a single "power problem" obscures which applies to which project. Grid interconnection queues, community and regulatory opposition, equipment supply bottlenecks, and gas infrastructure development timelines each affect different segments of the market differently. Major hyperscalers with capital, established utility relationships, and the ability to build behind-the-meter generation are stressed but continuing to build. The 30–50% delay figures cited in industry reporting reflect mid-market disruption more than hyperscaler disruption, a distinction that matters for understanding which projects are genuinely stopped versus delayed versus proceeding on revised timelines.

Firm always-on power is the requirement AI infrastructure cannot compromise on, and the sources that provide it fall into three tiers by readiness. The first tier, gas turbines, gas fuel cells, existing nuclear PPAs, is deployed and scaling. The second tier, EGS geothermal, nuclear SMRs, long-duration battery storage, is transitioning from early commercial to scaling, with Cape Station Phase I as the most important near-term proof point for the EGS category. The third tier, green hydrogen fuel cells, advanced SMR designs, offshore wind as firm power, remains post-2031 under base-case assumptions. The gap between the first tier and the second is the defining tension of the next decade.

Fuel cells are scaling fast and deserve serious attention, but the decarbonization pathway depends on fuels that do not yet exist at the volumes required. The hardware is compatible with biogas and green hydrogen, but US biogas is roughly 2.5% of total gas supply and green hydrogen costs $4–6/kg against a DOE target of $1/kg that has not been achieved anywhere commercially, and the policy environment worsened materially in 2025 with hub program defunding and IEA forecast cuts. At multi-gigawatt scale, the logistics of assembling thousands of 325 kW boxes and supplying them via pipeline also introduce operational complexity that Project Jupiter will test for the first time. The fuel cell case is not weak; it is conditional, on Project Jupiter validating the scale, and on alternative fuel supply emerging on a timeline that current evidence does not yet support.

Solar and onshore wind remain the lowest-cost new-build generation on an unsubsidized basis, but that comparison describes standalone generation. When storage is added to firm variable output for 24/7 AI loads, the all-in cost rises to levels competitive with rather than cheaper than gas. Their honest role in AI energy is as portfolio offsets, carbon accounting instruments, and daytime contributors to a mixed strategy, not primary firm power. Battery storage meaningfully reduces the firm backup requirement but cannot eliminate it: 4–8 hour lithium-ion systems cover overnight gaps; 100-hour iron-air systems, if commercially validated after 2028, would cover multi-day weather events. Neither addresses the multi-week low-wind, low-sun periods that firm baseload handles without storage at all.

Conventional geothermal is a reminder that the best clean firm power is often the option already fully subscribed. The Geysers and the Salton Sea produce 24/7 carbon-free baseload at costs competitive with gas, but those resources are contracted and geographically fixed. EGS replicates the physics without the geographic constraint, and Cape Station's well performance, 16 MW per production well, a decade ahead of NREL projections, suggests the physics is replicable faster than expected. Whether the economics follow at scale is the next question.

Nuclear's role is as a contractual benchmark more than a physical supply story for most data center operators. The PPAs being signed, Three Mile Island Unit 1 for Microsoft, Susquehanna for Amazon, Perry and Davis-Besse for Meta, are financial instruments that create incentives to restart or sustain plants that would otherwise close, which is genuine and measurable value. Whether the SMR pipeline converts from signed contracts to operating reactors before 2035 is the most consequential single uncertainty in the long-run clean energy mix. NuScale's cancellation and Vogtle's overruns are the cautionary precedent; TerraPower's Kemmerer project is the most tangible counter-evidence.

Community opposition and ratepayer impact are not peripheral. They are structural constraints that will shape which projects are built and where for the next decade, as surely as interconnection queue timelines. The social legitimacy of AI's energy consumption, whether ordinary households perceive it as imposing costs on them, is not an engineering question, but its resolution will determine whether the buildout continues at announced pace or encounters friction that compresses it. The White House Ratepayer Protection Pledge is a first formal acknowledgement; enforceability remains to be tested.

Under any plausible base-case trajectory, gas remains the dominant physical energy source for AI data centers through at least 2035. The infrastructure decisions being made between now and 2031 will determine whether 2041 looks like managed transition or structural lock-in. The difference depends less on whether clean alternatives exist, they do or credibly will, and more on whether their deployment timelines compress enough to intercept gas infrastructure before its economic life is too established to displace without policy intervention. That race is genuinely open. The next five years are when the trajectory gets set.