gigawatt-scale datacenter projects: the new ai infrastructure paradigm

executive summary

the emergence of gigawatt-scale datacenter projects represents a fundamental transformation in digital infrastructure, driven entirely by ai and machine learning compute demands. as of october 2025, there are 11 confirmed projects with power capacity of 1 gigawatt (1,000 megawatts) or greater, collectively representing 38.5 gigawatts of capacity and over $100 billion in investment.

to put this scale in perspective: a single 1 gw datacenter consumes as much power as a mid-sized city. the largest project - data city texas at 5 gw - will consume more electricity than the entire state of vermont. these projects are not incremental expansions of existing infrastructure; they represent a complete reimagining of what datacenter scale means in the age of large language models and generative ai.

why gigawatt scale is unprecedented

historical context: before 2024, the largest datacenter campuses topped out at 200-300 mw. meta’s eagle mountain (504 mw) was considered enormous. microsoft’s largest campus clusters approached 500 mw. the entire northern virginia “data center alley” - the world’s largest concentration - contains approximately 2.5 gw total across hundreds of facilities built over 20 years.

the paradigm shift: a single gigawatt-scale project now equals or exceeds the entire historical capacity of northern virginia. the 11 projects documented here represent 15x the total capacity of data center alley, planned for deployment in just 3-5 years. this is not evolution - it is revolution.

what changed: three factors converged:

1. ai training demands: gpt-4 required 25,000+ nvidia a100 gpus. claude 3, gemini ultra, llama 3 similar or greater. next-generation models demand 100,000+ gpus in single training runs
2. power density explosion: traditional datacenters consumed 6-8 kw per rack. ai workloads require 50-100 kw per rack, with liquid-cooled systems enabling 200-360 kw
3. behind-the-meter economics: waiting 5-7 years for utility grid interconnections became untenable. on-site power generation (natural gas, nuclear) enables 18-24 month deployment

complete gigawatt projects analysis

all projects sorted by capacity

geographic distribution

gigawatt-scale projects are concentrated in states that can solve the fundamental power availability challenge. traditional datacenter markets (silicon valley, northern virginia, dallas-fort worth) are notably absent - they lack the available power capacity for projects at this scale.

by state

pennsylvania dominance: with 11.1 gw across 3 projects (29% of total), pennsylvania has emerged as the epicenter of gigawatt-scale development. the state’s advantages are decisive:

• marcellus shale: largest natural gas formation in north america, producing gas at $2-3/mmbtu
• brownfield sites: former coal plants provide existing transmission infrastructure, substations, and grid connections
• pjm interconnection: access to nation’s largest wholesale power market enables grid export opportunities
• strategic location: between new york and washington dc provides low-latency connectivity

utah dark horse: millard county (population 12,975) will host 8 gw across two competing mega-campuses - more than the entire historical capacity of northern virginia. the rural county’s advantages include $5,000/acre land costs, dry climate enabling efficient cooling, and intermountain power project infrastructure.

traditional markets absent: silicon valley, northern virginia (outside prince william), dallas-fort worth, chicago, atlanta - the traditional datacenter powerhouses - have zero gigawatt-scale projects. their grids cannot accommodate projects at this scale within viable timeframes (5-7 year interconnection queues).

power sourcing strategies

the most critical finding: gigawatt-scale projects cannot rely on traditional utility grid connections. the power infrastructure required is equivalent to a major industrial facility or small city, and utilities cannot deliver capacity at this scale within acceptable timeframes.

power sourcing breakdown

on-site generation dominance: 60% of gigawatt capacity relies on behind-the-meter power generation, completely bypassing utility interconnection queues. these projects build their own power plants co-located with datacenters, operating as industrial microgrids.

fuel source analysis

natural gas overwhelming dominance: 77% of gigawatt-scale capacity will be powered by natural gas, primarily from marcellus shale (pennsylvania, west virginia) and permian basin (texas). natural gas provides:

• rapid deployment: 18-24 months from groundbreaking to operation
• low cost: $2-3/mmbtu in marcellus and permian vs$6-8 nationally
• baseload reliability: 24/7 operation at full capacity (vs 35% solar, 45% wind capacity factors)
• hydrogen readiness: ge vernova turbines can transition to hydrogen blends and eventually 100% hydrogen

renewable minority: only 3 projects (data city texas, vermaland arizona, cloverleaf wisconsin) have significant renewable energy components. even these use natural gas for baseload with solar/wind supplementation.

hydrogen transition path: data city texas and homer city pennsylvania both plan hydrogen transitions. data city will use adjacent 2 twh salt dome storage facility producing 280,000 tons green hydrogen annually. homer city’s ge vernova turbines are hydrogen-enabled for future conversion.

notable power sourcing innovations

data city texas (5 gw): world’s first 100% behind-the-meter datacenter city. completely isolated from ercot grid. phases:

1. initial: natural gas baseload with solar/wind/battery supplementation
2. transition: increasing renewable percentage with hydrogen blending
3. final: 100% green hydrogen from adjacent hydrogen city salt dome facility

homer city pennsylvania (4.5 gw): largest natural gas power plant in united states. seven ge vernova 7ha.02 hydrogen-enabled turbines consuming 665,000 mmbtu/day from marcellus shale. existing transmission infrastructure from former coal plant enables grid export to pjm and nyiso.

joule capital utah (4 gw): caterpillar g3520k generator sets with combined cooling, heat, and power (cchp/cogeneration). 1.1 gwh grid-forming battery storage. waste heat capture for district heating. completely independent from rocky mountain power grid.

shippingport pennsylvania (3.6 gw): conversion of 2.7 gw bruce mansfield coal plant to natural gas plus 900 mw new generation. will export 1+ gw excess capacity back to pjm grid. uses 800 million cubic feet/day from marcellus and utica shales.

investment scale and economics

gigawatt-scale projects represent the largest single-facility infrastructure investments in united states history, rivaling major airports, nuclear plants, and petrochemical complexes.

confirmed investment commitments

investment range: $2-40 billion per project, with$10-15 billion typical for greenfield gigawatt-scale campus. vermaland’s $33b represents largest datacenter development in us history.

per-gigawatt costs: vary dramatically based on strategy:

• $13-15b/gw: full greenfield campus with grid connections, extensive infrastructure (prince william, vermaland)
• $5-8b/gw: on-site generation with hybrid infrastructure
• $2-3b/gw: power-focused projects with datacenter tenants (homer city)
• $1-2b/gw: brownfield conversions leveraging existing infrastructure (shippingport)

brownfield advantage: converting former coal plants (homer city, shippingport, delta gigasite connection) saves $5-10 billion per gigawatt by reusing transmission lines, substations, cooling water infrastructure, and land with industrial zoning.

economic impact comparison

to contextualize these investments:

• **$40 billion prince william digital gateway** = larger than entire washington dc metro system ($6.7b), panama canal expansion ($5.3b), and big dig boston ($15b) combined
• $33 billion vermaland la osa = exceeds entire federal infrastructure investment in arizona over past decade
• **10 billion homer city** = equal to three nuclear reactor constructions (vogtle units 3&4 = 35b total, $17.5b per reactor at initial estimate)

job creation is modest: gigawatt-scale datacenters employ 100-500 permanent workers despite multi-billion investments. project jupiter (1 gw) projects 750 jobs. automation and remote management limit employment. economic benefit comes primarily from tax revenue, power infrastructure, and ecosystem development.

technology drivers: ai compute demands

gigawatt-scale projects exist solely because of artificial intelligence and machine learning workload demands. traditional enterprise computing, cloud services, and internet applications cannot justify or utilize capacity at this scale.

ai training requirements

next-generation models under development by openai (gpt-5), anthropic (claude 4), google (gemini 2), and meta (llama 4) are rumored to require 100,000+ gpus in single training runs consuming 400-500 mw for 6-12 months. this explains stargate’s 1.2 gw abilene campus design for 100,000 gpus on single network fabric.

inference at scale: once trained, deploying models globally for inference (answering user queries) requires distributed capacity. chatgpt serves 100+ million daily active users. each query requires gpu compute. microsoft, google, meta, amazon deploying foundation models to billions of users necessitates gigawatt-scale inference infrastructure.

power density evolution

6x power density increase: traditional datacenter racks consuming 6-8 kw now require 50+ kw for ai workloads. next-generation nvidia blackwell and amd mi300 systems will demand 100-200 kw per rack with liquid cooling, 200-360 kw with immersion cooling.

cooling paradigm shift: air cooling physically cannot remove heat at 50+ kw densities. liquid cooling (direct-to-chip cold plates, rear-door heat exchangers) required for 50-120 kw. immersion cooling (servers submerged in dielectric fluid) enables 200-360 kw. gigawatt-scale projects incorporate liquid cooling infrastructure from design phase.

gpu deployment scale

stargate abilene (1.2 gw): 400,000 nvidia gpus across 8 buildings on single network fabric. largest single gpu cluster globally. designed for training models with 100+ trillion parameters.

homer city (4.5 gw): could theoretically support 1+ million gpus at full buildout, enabling simultaneous training of multiple frontier models or massive-scale inference serving.

data city texas (5 gw): 15 million square feet could house 1.5+ million gpu servers, creating world’s largest ai compute facility by far.

infrastructure challenges at gigawatt scale

deploying gigawatt-scale datacenters requires solving engineering challenges never before encountered in digital infrastructure.

electrical infrastructure

substations: 3-5 gw requires 345 kv or 500 kv transmission-class substations typically used only for major industrial facilities or grid interconnections. prince william digital gateway needs multiple new 230 kv substations built by dominion energy.

transformers: stepping down transmission voltage (230-500 kv) to distribution (12-35 kv) to facility (480v) requires hundreds of large power transformers. global transformer manufacturing capacity is limited - lead times extending 24-36 months.

switchgear: 4-5 gw requires industrial-scale switchgear and circuit protection normally used in power plants. shippingport and homer city benefit from existing coal plant electrical infrastructure.

behind-the-meter generation: building 3-4 gw natural gas power plants requires:

• gas turbines: 5-10 large turbines per gigawatt. ge vernova, siemens energy only manufacturers at scale
• gas pipelines: 500,000-800,000 mmbtu/day consumption requires dedicated pipeline laterals
• emissions controls: selective catalytic reduction, co2 capture add $500m-1b per project
• grid synchronization: even behind-the-meter projects need grid connection for black start capability

cooling infrastructure

water requirements: 1 gw using traditional evaporative cooling consumes 1-2 million gallons/day - equivalent to small city. locations like arizona and utah face water scarcity constraints driving air-cooling and liquid-cooling designs.

cooling towers: evaporative cooling for 3-5 gw requires 50-100 large industrial cooling towers covering acres of land. visual impact and environmental concerns in rural areas.

liquid cooling distribution: direct-to-chip cooling requires facility-wide chilled water distribution at 10x the flow rates of traditional systems. pumping infrastructure rivals water treatment plants.

waste heat: 3-5 gw datacenter generates equivalent heat output to large power plant. tecfusions and joule capital capture waste heat for combined heating and power (chp) systems, potentially enabling district heating for nearby communities.

fiber connectivity

backbone requirements: gigawatt-scale facilities need 100+ gbps to 1+ tbps backbone connectivity. requires multiple diverse fiber paths to major internet exchanges.

latency constraints: rural locations (millard county utah, mingo county west virginia) are 100+ miles from tier 1 fiber routes. new fiber builds required adding $100-500 million to project costs.

redundancy: n+2 or n+3 fiber path diversity required for reliability. pennsylvania projects benefit from existing new york-washington corridor fiber. texas, utah, west virginia require new builds.

water and wastewater

potable water: construction phase and ongoing operations require 500,000-2 million gallons/day depending on cooling strategy.

wastewater treatment: on-site wastewater treatment required for rural locations lacking municipal infrastructure. adds $50-200 million to project costs.

stormwater management: 1,000+ acre campuses with buildings, parking, substations require extensive stormwater detention and treatment to prevent flooding and pollution.

timeline analysis

gigawatt-scale projects face dramatically longer development timelines than traditional datacenters due to power infrastructure, permitting complexity, and construction scale.

development phase timelines

grid interconnection bottleneck: waiting 5-7 years for utility to build transmission infrastructure and substations makes grid-connected gigawatt projects economically nonviable. this is primary driver of behind-the-meter generation trend.

behind-the-meter advantage: building on-site power generation reduces time to operation from 8-12 years to 3-5 years. homer city targeting 2027 operation (4 years from announcement). joule capital targeting 2026 (2 years from zoning approval).

project status and expected online dates

first wave operational: 2026-2027 will see first gigawatt-scale facilities come online (data city phase 1, homer city, joule capital). these will be first projects in history operating at 1+ gw single-facility scale.

buildout curve: 2025-2026: 2.5 gw | 2027-2028: 12+ gw | 2029-2030: 25+ gw | 2031+: full 38.5 gw

long-term developments: prince william digital gateway (20-year buildout) and vermaland arizona (10-year buildout) represent multi-decade infrastructure visions rivaling major airport or seaport construction timelines.

environmental and community impacts

gigawatt-scale datacenters create environmental and social impacts equivalent to major industrial facilities, raising concerns in host communities.

carbon emissions

natural gas projects emissions: 4 gw natural gas facility emits 12-16 million tons co2 annually - equivalent to adding 2-3 million cars to roads. 11 gigawatt projects using natural gas will collectively emit ~120-150 million tons co2 annually when fully operational.

coal comparison: homer city’s natural gas turbines emit 60-65% less co2 per mwh than former coal plant. shippingport conversion similar. brownfield natural gas conversions represent emission reductions vs legacy infrastructure.

renewable exceptions: only vermaland arizona (66% solar target) and data city texas (100% renewable goal via hydrogen) avoid massive carbon emissions. cloverleaf wisconsin commits to 30-40% renewable.

carbon capture: adams fork energy west virginia projects include >99% co2 capture for ammonia production co-located with datacenters. captured co2 becomes feedstock rather than emission.

water consumption

evaporative cooling impact: 1 gw with traditional cooling consumes 1-2 million gallons/day. 5 gw facility would consume 5-10 million gallons daily - as much as city of 50,000-100,000 people.

water-scarce locations: utah, arizona, west texas projects face water constraints. solutions include:

• air cooling (higher operating costs but zero water)
• liquid cooling with dry coolers (minimal water)
• groundwater rights acquisition (controversial in agricultural areas)

millard county utah controversy: two 4 gw projects in county with limited groundwater recharge. local opposition growing over water rights and agricultural impact.

community impact

rural overwhelm: millard county (12,975 population) will host 8 gw consuming more power than salt lake city. mingo county west virginia (23,000 population) faces 4.8 gw development. infrastructure, housing, services cannot scale fast enough.

tax revenue windfall: gigawatt projects generate $50-200 million annual property tax revenue for rural counties. millard county's entire annual budget is$47 million. transformational revenue but questions about capacity to manage growth.

local opposition: gouldsboro pennsylvania mega-project facing intense community resistance. concerns include:

• environmental degradation
• water depletion
• noise from generators and cooling systems
• visual impact (cooling towers, substations)
• traffic during construction (thousands of workers)
• strain on housing, schools, emergency services

job creation limited: 100-500 permanent jobs despite billions invested. construction phase creates 1,000-5,000 temporary jobs but housing shortages in rural areas force long commutes.

air quality

natural gas combustion: nox, sox, particulate matter from gas turbines require emissions controls. adams fork energy west virginia projects facing air quality permit scrutiny.

diesel backup generators: gigawatt-scale sites have hundreds of megawatts of diesel backup generators. emissions during testing and outages.

regulatory compliance: all projects require air quality permits from epa and state environmental agencies. public hearings becoming contentious.

nuclear and renewable alternatives

while natural gas dominates current gigawatt-scale projects, nuclear and renewable pathways offer carbon-free alternatives.

nuclear opportunities

three mile island restart: microsoft’s 20-year power purchase agreement for 835 mw unit 1 nuclear reactor demonstrates viability. $1.6 billion restart cost =$1.9 million/mw, dramatically cheaper than new nuclear at $6-10 million/mw.

restartable nuclear fleet: across united states, 12 reactors totaling 10+ gw shut down in past decade due to economic factors, not safety issues. could be restarted for $1-3 billion per gigawatt:

• palisades michigan: 805 mw
• indian point new york: 2,000 mw
• pilgrim massachusetts: 685 mw
• oyster creek new jersey: 636 mw
• fort calhoun nebraska: 478 mw

small modular reactors (smr): next-generation nuclear technology could enable datacenter co-location:

• nuscale: 77 mw modules, 6-12 module plants = 462-924 mw
• terrapower natrium: 345 mw sodium fast reactor
• x-energy xe-100: 80 mw high-temperature gas reactor

challenges: nrc approval takes 18-24 months minimum. public opposition in some areas. spent fuel storage. high upfront capital costs. but 100% capacity factor and zero carbon emissions attractive for long-term 20-30 year datacenter operations.

renewable + storage pathways

solar + battery: utility-scale solar costs $0.90-1.20/watt installed. 4 gw solar =$3.6-4.8 billion. but 35% capacity factor requires 12+ gw solar to achieve 4 gw effective capacity. battery storage for 24/7 operation adds $2-4 billion for 10+ gwh storage. total:$10-15 billion for 4 gw solar-powered datacenter vs $3-5 billion natural gas.

wind limitations: even in excellent wind resources, capacity factors max 45-50%. 4 gw effective requires 8-9 gw wind turbines. offshore wind in great lakes or coastal areas could support midwest or east coast gigawatt datacenters but costs $3-5 million/mw installed.

geothermal potential: utah delta gigasite exploring geothermal. great basin has geothermal resources. enhanced geothermal systems (egs) could provide baseload renewable power but technology still emerging.

hydropower: google pennsylvania modernizing 670 mw holtwood and safe harbor dams. hydropower provides baseload renewable but limited availability - most viable sites already developed.

hydrogen future

green hydrogen transition: data city texas represents most ambitious vision - 100% green hydrogen from adjacent 2 twh salt dome facility producing 280,000 tons annually. phases natural gas to hydrogen over 5-10 years.

hydrogen-enabled turbines: ge vernova 7ha.02 turbines (homer city) can burn hydrogen blends today, 100% hydrogen in future. enables natural gas projects to transition as green hydrogen production scales.

production economics: green hydrogen (electrolysis powered by renewables) currently costs $4-7/kg. natural gas equivalent is$1-2/kg. scaling production and renewable energy could reduce green hydrogen to $2-3/kg by 2030, approaching competitiveness.

infrastructure advantage: marcellus and permian natural gas pipelines can be repurposed for hydrogen with modifications. creates pathway for gigawatt gas projects to become gigawatt hydrogen projects.

key insights and takeaways

paradigm shift summary

1. unprecedented scale: 11 projects totaling 38.5 gw represent 15x the historical capacity of northern virginia data center alley, planned for deployment in 3-5 years vs 20+ years

2. power-first strategy: every gigawatt project solves power availability before site selection, zoning, or datacenter design. power is the primary constraint and primary enabler

3. behind-the-meter dominance: 60% of capacity uses on-site generation, completely bypassing utility interconnection queues. building your own power plant is faster than waiting for utilities

4. natural gas overwhelming: 77% of capacity powered by natural gas from marcellus shale and permian basin. renewable energy is supplemental, not primary, except for 3 projects

5. pennsylvania leadership: 29% of all gigawatt capacity in pennsylvania due to marcellus shale, brownfield coal plant sites, and pjm grid access

6. ai-exclusive demand: these projects exist only for ai training and inference workloads. traditional enterprise computing cannot justify or utilize this scale

7. investment concentration: $100+ billion across 11 projects, with individual projects ranging$2-40 billion rivaling nuclear plants and major airports

8. brownfield advantage: converting former coal plants saves $5-10 billion per gigawatt by reusing transmission, substations, cooling, and land infrastructure

9. timeline compression: behind-the-meter projects targeting 2-4 year deployment vs 8-12 years for grid-connected, making them economically viable

10. environmental controversy: natural gas projects emit 12-20 million tons co2 annually per gigawatt, equivalent to 2-3 million cars. rural communities facing water, air quality, and infrastructure strain

11. nuclear renaissance: microsoft three mile island deal proves nuclear restart viability. 10+ gw restartable nuclear capacity could power carbon-free gigawatt datacenters

12. geographic transformation: gigawatt scale impossible in traditional datacenter markets. future is rural sites with power abundance: pennsylvania gas fields, utah deserts, texas permian basin

the bottom line

gigawatt-scale datacenter projects represent a complete reimagining of digital infrastructure driven by artificial intelligence compute demands. the scale is unprecedented - single facilities consuming as much power as major cities. the approach is revolutionary - building power plants rather than connecting to utilities. the timeline is compressed - 2-4 years vs 8-12 years historically. the geography is transformed - rural pennsylvania, utah, and texas instead of silicon valley and northern virginia.

these are not incremental improvements. they are infrastructure projects comparable to nuclear power plants, major airports, and petrochemical complexes in scale, cost, and complexity. they will reshape the american power grid, accelerate natural gas consumption (or hydrogen transition if successful), and determine which states capture the economic benefits of the ai revolution.

the 11 gigawatt projects documented here are likely just the beginning. as ai models scale to 1 trillion+ parameters requiring 500+ mw training runs and serving billions of global users, demand for gigawatt-scale infrastructure will only intensify. the question is not whether more will be built, but where - and whether they transition from natural gas to nuclear or renewable energy before carbon emissions reach untenable levels.

data sources and methodology

this analysis is based on comprehensive review of:

• gigawatt projects data file (support/datacenters/enhanced/power/gigawatt_projects.json)
• individual state project databases covering all 50 us states
• project announcements, press releases, and regulatory filings
• utility interconnection queue data (pjm, ercot, rocky mountain power)
• environmental permit applications and air quality reviews
• local government zoning and planning documents
• industry publications: data center dynamics, data center frontier, utility dive

last updated: october 17, 2025

coverage: all confirmed datacenter projects with power capacity ≥ 1,000 mw (1 gw) in united states

verification: all projects verified through multiple independent sources including sponsor announcements, utility filings, or government records