Energy and Power, within the Infrastructure domain, denotes the integrated sociotechnical system governing the generation, transmission, distribution, storage, and consumption of electrical energy in contexts directly relevant to AI Data Centres, Bitcoin Mining operations, and the bro…
Semantic Classification
Content
Compositional Relationships (Components)
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:GridInfrastructure))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:GenerationCapacity))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:TransmissionNetwork))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:DistributionFeeder))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:BatteryEnergyStorage))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:PowerPurchaseAgreement))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:DemandResponseProgram))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:hasPart infra:CoolingInfrastructure))
## Dependency Relationships
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:requires infra:GridConnectionQueue))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:requires infra:BalancingServices))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:requires infra:DemandForecasting))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:requires infra:CarbonIntensityMonitoring))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:requires infra:WaterCoolingInfrastructure))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:dependsOn infra:ClimatePolicy))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:dependsOn infra:RegulatoryFramework))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:dependsOn infra:CapacityMarket))
## Capability Relationships
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:enables infra:HyperscaleDataCentreOperation))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:enables infra:BitcoinMiningOperation))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:enables infra:AITrainingInfrastructure))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:enables infra:NetZeroDecarbonisation))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:enables infra:GridFlexibility))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:supports infra:RenewableEnergyIntegration))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:supports infra:DemandSideManagement))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:supports infra:CarbonMarketOperations))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:supports infra:WaterResourceManagement))
## Implementation Relationships
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:PowerPurchaseAgreement))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:CapacityMarketMechanism))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:RenewableEnergyCertificate))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:BehindTheMeterGeneration))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:SmallModularReactorDeployment))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:GridScaleStorage))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:implements infra:DemandResponseContract))
## Reduction Relationships
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:reduces infra:CarbonEmissions))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:reduces infra:GridCongestion))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:reduces infra:EnergyWaste))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:reduces infra:WaterConsumption))
SubClassOf(infra:EnergyAndPower
ObjectSomeValuesFrom(infra:reduces infra:OperationalCost))
About Energy and Power
Energy and Power as an ontological concept within the Infrastructure domain captures the full stack of physical, contractual, regulatory, and geopolitical mechanisms that determine whether AI compute workloads, Bitcoin Mining operations, and the broader Digital Economy can access sufficient electricity at acceptable cost, reliability, and carbon intensity.
The concept has moved from a background operational concern to a primary strategic constraint for hyperscalers, cryptocurrency miners, and AI labs between 2022 and 2026, driven by the simultaneous acceleration of frontier AI model training (GPT-4, Claude 3, Gemini Ultra, LLaMA 3, GPT-5) and the continued expansion of Bitcoin proof-of-work mining across post-halving cycles.
The structural tension at the heart of this concept is the collision between near-term power demand — measured in gigawatts and growing at 15–25% per year in data centre load — and the multi-year lead times (5–15 years) required to construct new transmission infrastructure, large-scale generation plant, or grid connection capacity.
This mismatch creates acute scarcity in constrained load pockets and drives hyperscalers toward non-conventional solutions: co-located generation (gas reciprocating engines, fuel cells, diesel backup operating longer hours), nuclear PPAs and SMR developments, large-scale battery systems for peak shaving, and geographic arbitrage toward regions with underutilised grid headroom.
Bitcoin Mining has a unique relationship with energy infrastructure distinct from conventional data centres: Bitcoin miners are highly price-elastic, can curtail load within seconds at operator request (making them valuable demand-response assets to grid operators), tend to locate in regions with stranded renewable generation (Wyoming wind, Texas wind/solar, Iceland geothermal, Paraguay hydro), and generate monetisable thermal waste heat in some configurations.
The CCAF CBECI sustainable mining index estimated 52–60% of Bitcoin mining electricity came from low-carbon sources in 2024, though methodological debates persist around Renewable Energy Certificate (REC) accounting versus real-time matching.
Google reported a 48% increase in total greenhouse gas emissions between 2019 and 2024 in its 2024 Environmental Report, directly attributing the reversal of its previously downward carbon trajectory to rapid data centre expansion for AI workloads.
This admission catalysed industry-wide scrutiny of carbon-neutrality claims based on annual REC matching, accelerating adoption of 24/7 Carbon-Free Energy (CFE) frameworks that require hourly matching between electricity consumption and low-carbon generation.
Grid Infrastructure Architecture
Transmission systems operate at 110–765 kV HVDC/HVAC interconnects; distribution feeders at 11–33 kV in the UK (11 kV standard), 12–35 kV in North America. Hyperscale data centres connect at 33–132 kV point-of-connection (PoC), requiring dedicated grid connection agreements with Distribution Network Operators (DNOs) or Transmission Owners (TOs).
Grid connection queue in Great Britain reached 750+ GW of applications against approximately 70 GW of operational capacity in 2024, with National Grid ESO implementing the Connections Reform (Holistic Integrated Framework) to prioritise deliverable projects and prune queue bloat.
In the United States, PJM Interconnection’s connection queue held 2,600+ projects (285 GW) with median study times of 4 years as of 2024, prompting FERC Order 2023 requiring standardised interconnection procedures across all US Independent System Operators (ISOs).
Grid stability under high AI data centre load growth requires additional ancillary services: frequency response (containing Hz deviations post-generator trip), voltage support (reactive power injection from synchronous condensers, STATCOMs, SVCs), black-start capability (ability to restart the grid following a total blackout), and inertia provision (resisting rapid frequency change).
Inertia is increasingly provided by grid-forming battery inverters and synchronous condensers as conventional thermal generation retires, requiring data centre operators to understand how their load growth affects system inertia adequacy and frequency nadir calculations.
Data centres with large uninterruptible power supply (UPS) systems and battery storage can participate in frequency response markets: National Grid ESO’s Dynamic Containment (DC) market, launched 2020, pays approximately £5–20/MW/hour for 1-second frequency response, with data centre participation growing from zero in 2020 to over 200 MW by 2024.
The “interconnector premium” for grid-connected data centre locations reflects both physical capacity (available substation capacity at 33 kV or above, typically 20–200 MW per substation) and regulatory queue position. Sites with existing industrial substation infrastructure — former steel mills, automotive plants, aluminium smelters — command significant premiums because they avoid the 5–10 year queue wait for new connections.
Grid Connection Backlogs: Quantitative Assessment
Great Britain: 750 GW queued / 70 GW operational capacity → 10.7× oversubscription ratio. National Grid ESO estimates only 10–15% of queued projects are genuinely deliverable within 10 years. Connections Reform expected to reduce active queue to 100–150 GW by 2027.
PJM (US Mid-Atlantic/Midwest): 285 GW queued as of 2024. FERC Order 2023 mandates cluster study reform. Estimated queue clearance rate improving from 14% to 25–30% of projects proceeding to commercial operation under reformed rules.
ERCOT (Texas): 170 GW queued (predominantly solar and storage) against 90 GW installed capacity. Texas grid isolation (not synchronously connected to eastern US grid) limits import relief during peak demand events, creating acute summer reliability risk as data centre load grows.
Ireland (EirGrid): Moratorium on new large data centre connections in Dublin metropolitan area 2022–2025 due to grid constraint, forcing 2–3 GW of planned data centre development to Cork, Limerick, and Athlone regions. Represents the first national-level policy response to data centre grid overload globally.
Generation Mix and Levelised Cost Dynamics
As of 2025–2026, the global electricity mix for data centre-intensive regions spans utility-scale solar PV (LCOE: 35–60/MWh in Northern Europe), onshore wind (LCOE 40–65/MWh in UK), and offshore wind (LCOE 65/MWh with larger turbines and serial fabrication).
CCGT gas achieves LCOE 35–65/MWh for already-licensed plants with sunk construction costs. New nuclear SMR is projected at LCOE 60–80/MWh at 10+ unit fleet deployments according to BEIS/DESNZ modelling.
Hyperscalers pursuing 24/7 carbon-free energy (CFE) matching — Google’s target by 2030, Microsoft’s by 2030, Amazon through AWS by 2030 — require dispatchable low-carbon sources (existing hydro, nuclear, geothermal, or battery-firmed renewables) rather than RECs to achieve genuine hourly matching.
The 24/7 CFE matching metric, standardised by the Energy Tag consortium and endorsed by the Science Based Targets initiative (SBTi), measures the fraction of each hourly consumption interval covered by co-located or contractually matched low-carbon generation. Google achieved 64% 24/7 CFE globally in 2023, up from 56% in 2022, with leading performance in Iowa (95%) and trailing performance in Singapore (18%).
Power purchase agreement (PPA) pricing for new-build renewables has risen since 2021 due to supply chain inflation (polysilicon costs, steel, copper), rising interest rates affecting project finance, and grid connection costs. US utility-scale solar PPAs that cleared at 35–55/MWh in 2024–2025.
Hyperscalers’ appetite for long-term clean power has driven a significant fraction of the US renewable energy buildout: Amazon, Microsoft, Google, and Meta collectively contracted 65+ GW of new renewable capacity globally through 2024, representing roughly 20% of all utility-scale renewable capacity additions in their operating markets.
Renewable Energy Certificate (REC) and 24/7 CFE Market Structure
Annual RECs: Each MWh of renewable generation produces one REC (Renewable Energy Certificate in North America, Guarantee of Origin / GO in Europe, Large-scale Generation Certificate / LGC in Australia). RECs trade separately from electricity at 5–20/MWh for new-build nuclear or geothermal. Annual REC retirement enables market-based Scope 2 zero claims.
Granular Certificates (GCs): EnergyTag standard issues one GC per MWh tagged with generation hour, asset, location, and technology. GC prices range from 5–20/MWh (evening peak, scarce dispatchable clean generation), creating price signals incentivising investment in firm low-carbon resources.
24/7 CFE Score: Fraction of each hour’s consumption matched by clean generation. 100% 24/7 CFE requires either dispatchable clean generation (nuclear, geothermal, hydro) or storage-firmed renewables capable of supplying every consumption hour including evening peak and overnight periods. A portfolio of daytime solar and overnight wind can achieve 70–85% 24/7 CFE; closing the remaining gap to 100% requires long-duration storage, nuclear, or geothermal.
Power Purchase Agreements: Structure and Variants
Long-term bilateral energy contracts (10–25 year terms) between data centre operators and generation asset owners or developers constitute the primary mechanism through which hyperscalers secure clean energy.
Virtual PPAs (VPPAs, also known as synthetic or financial PPAs) are the dominant form in the United States: the buyer contracts a fixed price with a developer for a renewable project; when market prices exceed the fixed price the data centre operator receives the settlement payment, and vice versa. The VPPA effectively transfers merchant risk from developer to buyer, enabling project financing at lower cost of capital.
Physical PPAs involve actual delivery, requiring transmission rights and grid scheduling; they are more common in Europe (UK, Nordic countries) where merchant markets are less liquid and sleeved delivery via the transmission grid is administratively feasible.
Amazon Web Services and Microsoft collectively held 30+ GW of contracted renewable PPAs by end-2024, making hyperscalers the largest corporate buyers of renewable energy globally — ahead of all utilities and industrial companies.
The PPA market has become a critical channel for new renewable project development financing: without anchor PPA offtakers, many offshore wind and solar projects cannot achieve financial close. This dependency creates systemic risk if hyperscaler capital expenditure cycles slow or regulatory changes discourage long-term commitments.
Tolling agreements provide an alternative or complement to energy PPAs: a tolling agreement gives the data centre operator control over a generating asset (gas peaker, battery) in exchange for paying fixed capacity costs, allowing flexible dispatch scheduling around grid price spikes.
VPPA Financial Mechanics: A Worked Example
Structure: Data centre signs 15-year VPPA with wind developer at fixed price $45/MWh, for 200 MW project in Texas (ERCOT).
Year 1: ERCOT market price averages 45/MWh × 200 MW × 8,760 h × 30% CF = 35/MWh = 5.3M (the “negative settlement”). Data centre retires 200 MW × 8,760 h × 30% CF = 525,600 RECs, claiming zero Scope 2.
Year 5: ERCOT market price averages 45/MWh fixed. Receives ERCOT settlement at 34.1M. Net receipt = $10.4M (positive settlement). Data centre’s renewable hedging generates revenue while maintaining REC retirement.
Risk: If ERCOT market prices remain permanently below $45/MWh, the VPPA imposes a long-run cost premium versus buying market power. Data centres accept this as an insurance premium against carbon price exposure and regulatory risk.
Nuclear Energy: PPAs and Small Modular Reactors
The Microsoft–Constellation Energy agreement to restart Three Mile Island Unit 1 (Pennsylvania, 835 MWe) for 20 years at reported prices of $100–115/MWh was executed in September 2024 and received Nuclear Regulatory Commission (NRC) approval in March 2025, with restart targeting late 2025. This transaction represents the first commercial nuclear-to-hyperscaler 24/7 CFE arrangement at this scale globally.
The Three Mile Island Unit 1 restart required approximately 800 million. The plant had closed in 2019 due to competition from cheap shale gas, following the 1979 accident at Unit 2 (a separate reactor on the same campus).
Amazon acquired the Talen Energy Susquehanna nuclear campus (Pennsylvania, adjacent to the 2,500 MWe Susquehanna Steam Electric Station) in a $650 million transaction completed March 2024, aiming to co-locate AWS capacity directly adjacent to the nuclear plant with a 480 MW direct-connection data centre campus.
FERC proceedings on co-location rules (Docket EL24-19) were ongoing through 2025, with implications for whether direct nuclear-to-data centre connections bypass grid charges, potentially creating cross-subsidisation of other ratepayers and disrupting wholesale market price signals.
Google signed an agreement with Kairos Power in October 2024 for a series of small modular high-temperature gas-cooled reactors (HTGR) totalling 500 MW, targeting commercial operation from 2030. The Kairos KP-FHR (fluoride salt-cooled, pebble-bed fuel) design operates at 600°C coolant temperature, enabling high-efficiency power generation.
SMR Development Pipeline (Global)
TerraPower Natrium (Bill Gates-backed): 345 MWe sodium fast reactor with molten salt thermal storage enabling variable electrical output (250–500 MWe) from fixed thermal output. Kemmerer, Wyoming groundbreaking June 2024, targeting 2030 commercial operation. DOE Advanced Reactor Demonstration Programme (ARDP) co-funded.
X-energy Xe-100: 80 MWe pebble-bed HTGR per module; four-module 320 MW plant design. DOE ARDP Phase 1 awardee. Sited at Dow Chemical’s Seadrift, Texas facility (industrial heat + power), targeting 2030. High-temperature output (750°C) enables industrial decarbonisation beyond electricity.
Rolls-Royce SMR: 470 MWe pressurised water reactor, UK Generic Design Assessment Step 1 completed 2024, Step 2 targeting 2028. Targeting £50–60/MWh LCOE at fleet scale, competitive with new offshore wind. Factory-fabricated modular construction at Derby facility. Eligible for UK Contract for Difference (CfD) scheme.
NuScale VOYGR: 77 MWe per module, 462 MW six-module plant. First plant (Carbon Free Power Project, Idaho) cancelled November 2023 due to cost overrun (58/MWh original estimate). Restructured for European market (KGHM Poland partnership). Illustrates first-of-kind cost risk for SMR developers.
Holtec SMR-300: 300 MWe, targeting existing decommissioned US nuclear sites (Palisades Michigan restart under NRC review 2024). Great British Nuclear (GBN) shortlisted vendor for UK SMR selection.
Physics case for nuclear in data centre energy: A single 835 MWe nuclear unit produces 7,300 GWh per year (capacity factor ~95%), equivalent to 21,000 acres of solar panels or 2,800 onshore wind turbines. Nuclear provides firm capacity credit in capacity markets (US Capacity Performance, UK Capacity Market T-4 auctions), reducing the balancing cost burden on grid operators — a critical advantage as grid inertia declines.
Battery Energy Storage Systems
Grid-scale lithium iron phosphate (LFP) battery energy storage system (BESS) deployments have grown from 5 GW globally in 2020 to over 150 GW installed capacity by end-2024, with the United States (50 GW), China (60 GW), and Europe (20 GW) as leading markets.
LFP chemistry has displaced nickel manganese cobalt (NMC) at grid scale due to superior cycle life (3,500–6,000 full charge-discharge cycles at 80% depth of discharge, vs 1,000–2,000 for NMC), superior thermal stability (no thermal runaway at overcharge, critical for large installations), and lower cobalt content (reducing supply chain exposure to DRC cobalt production).
Tesla Megapack (3.9 MWh per unit, containerised), BYD MC-I (3.44 MWh per unit), and Fluence Gridstack (3.9 MWh per unit) represent dominant utility-scale product lines; CATL, Wärtsila, and Saft compete in the European market. Capital costs have fallen from 120–150/kWh (2025), with further decline toward $80–100/kWh projected by 2030.
Front-of-meter BESS provides multiple revenue streams simultaneously (revenue stacking): wholesale energy arbitrage (buying at off-peak prices, selling at peak), frequency response services (Dynamic Containment, Dynamic Regulation, Dynamic Moderation in GB; Frequency Regulation Up/Down, Spinning Reserve in US ISOs), capacity market participation, and transmission congestion relief.
GB BESS revenues exceeded £80/kW/year in high-revenue periods (2022–2023) before market saturation began compressing frequency response revenues in 2024–2025. Revenue stacking in ERCOT achieved $120–180/kW/year in 2023 due to price volatility, with 4-hour BESS yielding simple payback periods of 5–7 years.
Behind-the-meter BESS at data centres serves distinct functions: UPS backup replacing or supplementing diesel generators (2–4 hour LFP systems providing 30–60 minutes at full load), peak demand charge reduction (avoiding monthly demand charges of $10–35/kW/month at peak intervals), power factor correction, and demand response provision.
The economics of behind-the-meter BESS at US data centres have improved dramatically as LFP capital costs have fallen, yielding typical simple payback periods of 4–7 years for demand charge optimisation alone, improving to 2–4 years when ancillary service revenues are stacked.
Long-Duration Energy Storage (LDES)
Long-duration energy storage — systems providing 8–100+ hours of discharge — is emerging as a critical complement for data centres seeking higher renewable fractions.
Pumped Hydro: Roundtrip efficiency 70–85%, virtually unlimited cycle life, 1,000+ hour storage capacity possible. Geography-constrained (requires elevation differential and water supply). 160+ GW installed globally. Dominates long-duration storage by capacity. Not buildable in most data centre locations but benefits entire grid.
Iron-Air Batteries (Form Energy): 100-hour discharge duration targeting $20/kWh capital cost, iron-air reversible oxidation chemistry, unlimited cycle life. Small-scale utility deployments underway 2024–2025 in US. Grid-scale deployments targeting 2026–2028.
Vanadium Redox Flow Batteries: 4–12 hour discharge duration, unlimited cycle life (electrolyte is not consumed), scalable by decoupling power (stack size) and energy (electrolyte volume). More expensive than LFP at equivalent duration (120–150/kWh), but suitable for 8–12 hour applications where LFP degrades rapidly.
Compressed Air Energy Storage (CAES): 10–100 hour duration, requires salt cavern or depleted gas field geology. Rosewater Energy Project (UK, 300 MW / 7,500 MWh, Cheshire salt caverns) and Hydrostor (Canada/US) represent near-term pipeline. Very low per-kWh cost at scale ($50–80/kWh) but high site specificity.
Bitcoin Mining Energy Economics
Bitcoin mining economics are fundamentally determined by the ratio of Bitcoin price to electricity cost, mediated by hardware efficiency. The Bitmain Antminer S21 Pro (2024 generation) achieves 234 TH/s at 3,510 W power consumption (15.0 J/TH efficiency). The S21 XP HYD (liquid-cooled) achieves 473 TH/s at 5,676 W (12.0 J/TH). At Bitcoin price of 0.035–0.060/kWh depending on hardware efficiency.
This break-even analysis drives Bitcoin miner location decisions overwhelmingly toward very low-cost power: sub-0.02/kWh flared natural gas (North Dakota Bakken, Permian Basin), or sub-$0.025/kWh wind/solar curtailment (West Texas, Wyoming).
Core Scientific (CORZ), Marathon Digital (MARA), Riot Platforms (RIOT), and CleanSpark (CLSK) are the largest US publicly listed miners by operating hashrate and power capacity (400–1,200 MW each contracted as of 2025). The April 2024 fourth halving (block subsidy reduction from 6.25 to 3.125 BTC) compressed miner margins significantly.
Bitcoin miners’ ability to curtail load on 5–60 second timescales at grid operator request makes them attractive demand-response assets: ERCOT (Texas) paid Bitcoin miners an estimated $175 million in voluntary curtailment payments during the Summer 2023 heat wave season alone, with miners reducing load by up to 1,200 MW during grid stress events.
The repurposing of Bitcoin mining infrastructure for AI GPU hosting has become a significant trend. Core Scientific executed a hosting agreement with CoreWeave (NVIDIA H100 cluster) worth 250,000–50,000–$100,000/year from Bitcoin mining at comparable electricity costs.
Northern Data (Germany) and Applied Digital (US) have similarly pivoted from Bitcoin mining to AI cloud services, exploiting the structural overlap: both require large, reliable power supplies; both generate significant heat requiring industrial cooling; both have flexible load profiles amenable to demand response. The capital stock (purpose-built data halls with power infrastructure) is largely transferable with hardware swap.
CCAF CBECI Methodology
The Cambridge CBECI uses a techno-economic model combining: (1) network hashrate (publicly observable from block timestamps and difficulty adjustments), (2) hardware efficiency distribution modelled from hardware sales data, shipping volumes, and age-based retirement curves, and (3) regional electricity prices from IEA/EIA regional price data.
The model yields a lower bound (all miners using most efficient hardware), central estimate (efficiency distribution weighted by deployment probability), and upper bound (least efficient hardware still economically viable at current prices). Central estimate confidence interval is approximately ±30%.
Geographic distribution post-China 2021 ban: United States 38%, Kazakhstan 13%, Russia 11%, Canada 6%, Germany 4%. This distribution drives the carbon intensity calculation: CCAF estimates the Bitcoin network’s average carbon intensity at 450–550 gCO₂/kWh, yielding total annual emissions of 55–90 MtCO₂e.
The sustainable mining index — tracking the fraction of Bitcoin electricity from low-carbon sources — reached 52–60% in 2024, accounting for VPPAs and RECs purchased by miners. The methodological debate centres on whether annual REC accounting (as used by Bitcoin Mining Council) or real-time hourly matching (as required by SBTi) is the appropriate standard.
Water Cooling and Thermal Management
Large GPU clusters for AI training generate heat densities of 15–80 kW per server rack, compared to 5–15 kW for conventional compute and 2–5 kW for traditional enterprise servers. NVIDIA H100 GPU consumes 700 W each; 8 per server = 5,600 W GPU load alone. Traditional computer room air conditioning (CRAC) units are inadequate above approximately 20 kW/rack; liquid cooling is mandatory for frontier AI infrastructure.
Cooling modalities span a spectrum of cost and effectiveness:
Rear-door heat exchangers (RDHx): Water-cooled panels mounted on rack rear, capturing heat from exhaust air. PUE benefit approximately 0.05–0.10. Low capital cost. Compatible with existing infrastructure. Limited effectiveness above 25 kW/rack.
Direct liquid cooling (DLC): Cold plates mounted directly on CPU/GPU heat spreaders, coolant (water/glycol or single-phase dielectric) routed through server to facility cooling system. Enables 95–98% heat capture. PUE benefit 0.10–0.20. Requires custom server designs; now standard on NVIDIA DGX H100 and H200 systems.
Immersion cooling: Servers submerged in dielectric fluid — single-phase (mineral oil, 3M Novec FC-72) or two-phase (fluorinated fluid, boiling and condensing cycle). Enables >99% heat capture, 50% reduction in cooling power, PUE potential 1.03–1.08. Higher capital cost, more complex maintenance. Used in select high-density AI training deployments (Microsoft, Meta, hyperscale pilots).
Water consumption effectiveness (WUE) measures litres of water consumed (evaporated, not recycled) per kWh of IT equipment load. Evaporative cooling towers achieve WUE of 0.5–2.0 L/kWh, meaning a 100 MW data centre operating 8,760 hours/year consumes 440 million to 1.75 billion litres of water annually — comparable to the annual water use of a town of 10,000–50,000 people.
In water-stressed regions (Phoenix, Las Vegas, Dallas, Singapore), this drives municipal opposition, zoning restrictions, and in some cases moratoriums on new data centre water connections (Chandler, Arizona 2023). Microsoft, Google, and Meta have each committed to water-positive operations (returning more water to local watersheds than consumed) by 2030.
Waste heat valorisation transforms a data centre’s thermal output from a cost to an asset: heat rejected from liquid-cooled clusters at 40–60°C can supply low-temperature district heating networks directly; heat pumps can upgrade this to 80°C for higher-temperature networks.
Stockholm’s Fortum Värme programme (10 MW data centre heating 10,000 homes) provides the operational template; Rotterdam has mandated waste heat recovery for new data centres above 1 MW. In the UK, Manchester city-region feasibility studies explore routing MediaCityUK data centre waste heat into the Manchester District Energy Company (MDEC) network.
Power Usage Effectiveness (PUE) — the ratio of total facility energy to IT equipment energy — remains the primary efficiency metric. Industry average PUE stands at approximately 1.55 (2024); hyperscalers achieve 1.1–1.15 through free-cooling, liquid cooling, and heat recovery. UK government’s DESNZ data centre sustainability framework includes a minimum PUE requirement of 1.2 for new planning applications from 2025.
Carbon Intensity and Scope 2 Accounting
Scope 2 market-based emissions accounting allows data centres to claim zero emissions through REC purchase, masking underlying grid carbon intensity. The tension between market-based and location-based Scope 2 accounting is fundamental to understanding corporate net-zero claims in the data centre sector.
Under GHG Protocol Scope 2 Guidance (2015), companies may choose either location-based (using grid average emission factors) or market-based (using RECs/GOs) accounting. Market-based accounting has enabled hyperscalers to claim “100% renewable electricity” whilst physically drawing 60–80% grid power from fossil sources.
The Science Based Targets initiative (SBTi) Corporate Net-Zero Standard (2021, updated 2024) requires 1.5°C-aligned scope 2 targets using location-based methodology or 24/7 hourly matching for companies making renewable electricity claims.
The EnergyTag standard (EnergyTag Ltd, UK, launched 2022) provides the technical framework for granular certificate (GC) issuance: one GC per MWh of electricity generated, tagged with hour of generation, generation asset, location, and technology. GC prices range from 5–20/MWh (evening peak solar).
Carbon intensity of the electricity grid varies by location, season, and hour. GB grid carbon intensity (tracked by National Grid ESO’s Carbon Intensity API at api.carbonintensity.org.uk) ranges from under 50 gCO₂/kWh during high wind periods to over 300 gCO₂/kWh during low-wind winter evenings when gas peakers set the marginal price.
The US average grid intensity is approximately 380 gCO₂/kWh (EIA 2024), ranging from 17 gCO₂/kWh in Washington State (Columbia River hydro) to 740 gCO₂/kWh in parts of the Midwest (coal-heavy grids).
For AI inference workloads that can be geographically distributed, carbon-aware workload shifting — routing requests to data centres with lowest marginal carbon intensity at each hour — offers 10–40% Scope 2 reduction without additional clean energy procurement. This approach is implemented by Google Cloud (Carbon-Intelligent Computing, operational since 2020), Microsoft Azure (carbon-aware SDK, open-source 2022), and AWS (customer carbon footprint tool 2022).
Marginal emissions factors (MEFs) — measuring the carbon intensity of the next unit of electricity consumed on the grid — differ significantly from average emission factors. In the GB system, the MEF is dominated by gas (CCGT) during most hours (MEF approximately 200–350 gCO₂/kWh), whilst the average includes high-penetration wind and nuclear.
AI-Driven Grid Optimisation
The energy–AI feedback loop is bidirectional: whilst AI workloads represent the primary demand growth vector, AI methods are enabling significant improvements in grid operation efficiency — partially, though not fully, offsetting the demand growth they also cause.
Load forecasting using transformer-based deep learning models (UK Power Networks, National Grid ESO, Elia Belgium) achieves day-ahead load forecast Mean Absolute Percentage Errors (MAPE) of 0.8–1.5% versus 2–3% for traditional econometric models, reducing reserve requirements and balancing costs by 5–10%.
Intra-day forecasting (4-hour ahead) achieves sub-0.5% MAPE with ensemble methods combining Numerical Weather Prediction (NWP) outputs with historical consumption patterns and real-time weather observations.
Optimal power flow (OPF) solvers using Graph Neural Networks (GNNs) trained on historical network states demonstrate 95%+ accuracy in predicting OPF solutions in milliseconds versus hours for conventional interior-point methods, enabling real-time grid redispatch during contingencies. DeepMind’s collaboration with National Grid ESO for SF6-free substation control and Google’s RL for Cloud cooling systems (40% cooling energy reduction) demonstrate AI’s operational grid impact.
Fault detection and predictive maintenance using AI methods — convolutional neural networks on vibration sensor data for transformer monitoring; anomaly detection on SCADA time-series for transmission line health — extend equipment lifetimes, reduce unplanned outages, and defer capital replacement. Scottish Power Transmission (SPT) and National Grid Transmission (NGT) have active AI-based asset health monitoring programmes covering 400 kV transmission assets.
Renewable energy forecasting benefits from AI methods at multiple timescales: now-casting (0–6 hours), day-ahead (24–48 hours), and medium-range (7–14 days). Deep learning solar irradiance forecasting achieves 15–25% lower RMSE than statistical baselines for intra-day horizons; physics-informed neural networks (PINNs) incorporating wind farm wake effects achieve 5–15% RMSE reduction versus pure data-driven baselines.
National Grid ESO’s AI Strategy (published 2024) commits to deploying ML methods across all major operational domains — forecasting, dispatch, asset monitoring, market operations — by 2027, with projected savings of £200–400 million per year in reduced balancing costs at 2030 renewable penetration levels.
Carbon-Aware AI Workload Scheduling
Carbon-aware computing schedules flexible AI workloads (training jobs, batch inference, data pre-processing) to hours and locations with lowest marginal grid carbon intensity, effectively acting as a demand-side decarbonisation strategy without requiring additional clean energy investment.
Google’s Carbon-Intelligent Computing system shifts 5–30% of its internal compute jobs temporally and geographically based on forecast grid carbon intensity, claiming 10–18% reduction in Scope 2 market-based emissions from compute without affecting job completion times.
The Green Software Foundation’s Carbon Aware SDK (open-source, 2022) provides APIs for software developers to implement carbon-aware scheduling, integrating with real-time carbon intensity APIs (National Grid ESO Carbon Intensity API for GB, WattTime API for US ISOs, Electricity Maps for Europe).
Use Cases and Major Deployment Families
Energy and Power in the AI/crypto context manifests across several deployment archetypes, each with distinct power, cooling, and carbon management requirements:
Hyperscale AI Training Clusters
GPT-4 training at OpenAI–Microsoft consumed approximately 50 GWh over several months in 2022–2023. GPT-5 scale training in 2024–2025 consumed an estimated 300–600 GWh, comparable to a small city’s annual electricity use. Meta’s Llama 3 405B training across 16,000 H100 GPUs ran at roughly 50 MW continuous load.
Frontier model training at the 2026 scale — projected 10²⁶–10²⁷ FLOPs — may require dedicated 500 MW to 2 GW campus-scale data centres with multi-year lead times for grid connection. Anthropic’s AWS partnership (up to $4 billion committed), Google DeepMind’s 2024 compute investment, and xAI’s Memphis Tennessee facility (targeting 1 GW+ by 2026) define the current frontier.
AI Inference Clusters (Distributed, Edge-to-Hyperscale)
AI inference — running trained models to serve user requests — has distinct energy characteristics from training: lower per-query peak power (inference batch size limited by latency requirements), high concurrency (millions of simultaneous users), near-zero tolerance for power outages (user-facing SLA), and geographic distribution matching user demand patterns.
ChatGPT inference was estimated at 0.001–0.01 kWh per query in 2023–2024; at reported 100 million daily users generating 5 queries each, total inference energy is approximately 50,000–500,000 kWh/day (50–500 MWh/day). Annualised, this represents 18–180 GWh/year for a single frontier model at user scale — a meaningful but manageable fraction of the 945 TWh total data centre projection.
Cryptocurrency Mining Farms
Industrial-scale Bitcoin mining farms operate at 100–500 MW per campus, drawing constant baseload power 24/7/365 unlike AI training (which has variable utilisation). Mining farms prioritise interruptible power contracts (accepting curtailment risk) to access sub-$0.03/kWh tariffs unavailable to firm-power data centres.
The largest operational mining farms include: Riot Platforms Rockdale, Texas (700 MW capacity), Marathon Digital’s facility in Garden City, Texas (390 MW), and Greenidge Generation’s Dresden, New York facility (115 MW, operating behind an existing gas turbine). International sites include HUT8 in Alberta Canada (350+ MW), and Bitfarms in Paraguay (300 MW hydropower).
Colocation and Edge Computing Power
Edge inference nodes (server-room scale, 10–500 kW) at telecom base stations, retail sites, and industrial facilities create a distributed energy challenge distinct from hyperscale: loads are too small for dedicated grid connections but too important for unplanned outages, driving battery backup adoption and micro-CHP (combined heat and power) installations.
The UK’s rollout of 5G-enabled mobile edge computing (MEC) under Ofcom spectrum assignments requires each MEC node to provide 99.999% availability with 2–4 hour battery backup, representing a new category of electricity demand (estimated 1–3 GW aggregate UK MEC capacity by 2030) that sits between consumer IoT and hyperscale infrastructure.
Academic Context
The academic study of energy and power in AI and cryptocurrency contexts spans energy economics, power systems engineering, environmental science, and computer science. Seminal contributions have defined the field:
Masanet et al. (2020, Science) argued that efficiency improvements had historically decoupled data centre energy growth from compute growth, with global data centre energy consumption remaining roughly flat at 200–250 TWh from 2010 to 2018 despite 10× increase in compute workloads, attributable to server consolidation, cooling improvements (PUE reduction from 2.0 to 1.55), and cloud migration.
This “efficiency dividend” thesis was challenged by the post-2020 AI acceleration: rapid GPU buildout, new training workloads, and inference scaling have combined to outpace efficiency gains, with IEA tracking 30–35% annual growth in data centre electricity consumption 2022–2024.
De Vries (2023, Joule) published the most prominent academic series tracking both AI and Bitcoin energy consumption, using supply chain analysis for AI hardware (NVIDIA A100/H100 GPU quantities shipped, power specifications, utilisation rates). De Vries’ AI energy estimates — 85–134 TWh/year by 2027 for AI alone — have been critiqued as overstated but broadly calibrated against IEA estimates for data centre AI load share.
Strubell et al. (2019, ACL) quantified the carbon cost of NLP model training, finding that training a large transformer model with neural architecture search could produce 626,155 lbs CO₂e — equivalent to five times the lifetime emissions of a car. This paper initiated the “Green AI” research agenda and influenced corporate reporting practices for AI training emissions.
Patterson et al. (2021, arXiv) from Google provided counterpoint, arguing that hardware efficiency improvements (TPUv4 vs 2019-era hardware), renewable energy sourcing, and cloud efficiency mean that GPT-3 training emitted approximately 552 tonnes CO₂e — dramatically less than naive extrapolations of Strubell’s methodology would suggest — highlighting the importance of hardware efficiency generation and location-based versus market-based Scope 2 accounting.
Cambridge CBECI (Rauchs et al. 2019–2024) uses a techno-economic model published with full methodological transparency, updated monthly, covering Bitcoin network hashrate, hardware efficiency, and geographic distribution. The CBECI has become the primary reference for policymakers (US Congress, European Parliament, UK Treasury consultations on crypto) in quantifying Bitcoin’s energy footprint.
The companion Cambridge Digital Assets Programme (CDAP) extends methodology to post-Merge Ethereum (99.95% energy reduction versus proof-of-work) and broader digital asset energy benchmarks — providing comparative context for the Bitcoin vs Ethereum energy debate.
Lannelongue et al. (2021, Advanced Science) introduced the Green Algorithms framework for estimating carbon footprint of computational research, providing an online calculator adopted by Nature, Wellcome Trust, and UK Research and Innovation as a recommended tool for grant applications and publications.
Schwartz et al. (2020, CACM) formalised the “Green AI” research agenda, distinguishing between Red AI (pursuing performance at any cost) and Green AI (prioritising efficiency), advocating for reporting of energy consumption alongside accuracy metrics in ML publications.
Dodge et al. (2022, FAccT) measured carbon intensity of AI in cloud instances across AWS, Azure, and GCP, finding 10–100× variation depending on cloud region and time of day, demonstrating that workload placement decisions are as important as hardware efficiency for AI carbon footprint reduction.
The systems-level tension between Jevons paradox (efficiency improvements inducing demand growth through rebounding consumption) and absolute demand reduction is central to the green AI debate: critics argue that every efficiency improvement in GPU training is offset by researchers training larger models; proponents point to inference efficiency improvements (model distillation, quantisation, speculative decoding) as genuinely reducing per-query energy consumption at scale.
Key Journals and Publication Venues
Nature Energy (IF ~60): Primary publication venue for energy transition research with AI/data centre relevance. Notable data centre papers include Masanet et al. (2020, sister journal Science) and Shehabi et al. (2016, LBNL report feeding into Science policy).
Joule (Cell Press, IF ~40): Focuses on sustainable energy, publishes De Vries’ annual Bitcoin and AI energy consumption updates, and empirical energy transition economics.
IEEE Transactions on Power Systems / IEEE Transactions on Smart Grid: Primary technical venues for power systems engineering relevant to data centre grid integration, demand response, and storage optimisation.
Environmental Research Letters (IOP Publishing): Cross-disciplinary environmental impact of data centres and cryptocurrency; Shehabi et al. (2018) decoupling analysis published here.
ACM/IEEE International Symposium on Computer Architecture (ISCA): Primary venue for hardware-level energy efficiency research (GPU architectures, memory bandwidth, cooling co-design) that determines the “floor” of AI compute energy consumption.
IPCC Working Group III (Mitigation of Climate Change): The 2022 Sixth Assessment Report’s Chapter 9 (Buildings) and Chapter 11 (Industry) contain relevant material on data centre energy; Chapter 6 (Energy Systems) addresses grid integration of variable renewables at the scales required to decarbonise AI infrastructure.
IEEE Power and Energy Magazine: Practitioner-accessible bridge between research and utility/ISO operations, regularly covering data centre demand response, virtual power plants, and AI for grid operations.
The Journal of Cleaner Production and Sustainable Cities and Society cover urban heat island effects of data centre clusters, waste heat valorisation case studies, and water consumption governance — particularly relevant to Manchester, Leeds, and other northern UK urban data centre cluster planning contexts.
The Energy Policy journal (Elsevier) regularly publishes empirical and modelling studies on electricity market design, capacity mechanism reform, and industrial demand response — all directly applicable to hyperscaler and Bitcoin miner interactions with electricity markets in the UK (Capacity Market), US (PJM, ERCOT), and EU (ENTSO-E).
Applied Energy (Elsevier) covers engineering-level studies of data centre thermal management, waste heat recovery system designs, and building-integrated cooling — providing the technical underpinning for commercial deployment of liquid cooling and heat valorisation at scale.
Current Landscape (2026)
As of early 2026, global data centre electricity consumption is tracking toward the IEA’s central 945 TWh by 2030 projection, with Goldman Sachs (Power Up, 2024) and ARK Invest (Big Ideas 2025) suggesting the upper bound (1,200–1,500 TWh) is more probable given frontier AI training scale projections.
IEA’s Electricity 2025 report (January 2025) revised upward its 2030 data centre demand forecast by 15% versus the 2024 edition, reflecting faster-than-expected AI hardware deployment. EPRI’s Powering Intelligence update (2025) maintained its 46–174 TWh/year US AI electricity demand range for 2030, with central scenario at 90 TWh/year — equivalent to adding New York State’s electricity consumption to the US grid.
Nuclear power has re-entered mainstream energy planning for AI infrastructure. The US Department of Energy’s Liftoff series (2023–2024) identified advanced nuclear as a critical pathway for 24/7 carbon-free data centre power, with the Nuclear Energy Institute estimating 16 GW of new nuclear capacity needed for data centre loads by 2040.
Bitcoin mining has entered a post-fourth-halving phase (April 2024, block subsidy now 3.125 BTC) in which miner profitability requires either BTC price above $60,000 or very low electricity costs. The bifurcation between miners pivoting to AI GPU hosting (Core Scientific, Iris Energy, Northern Data) and those focused on hashrate efficiency (Riot Platforms, CleanSpark) is the dominant strategic divide in the sector.
MARA Digital’s vertical integration (proprietary ASIC development, large-scale wind PPAs in Nebraska and Texas totalling 500 MW) represents a strategic path focused on long-run cost leadership distinct from the AI-pivot strategies.
Grid connection queues remain the most acute constraint for new data centre development. ENTSO-E estimates 500 GW of queued renewables and storage in European interconnection queues as of 2024. Ireland’s EirGrid moratorium on new large data centre connections in Dublin (2022–2025) forced hyperscalers to Cork, Limerick, and Athlone, with significant impact on Irish grid planning.
The Microsoft–Amazon–Google–Meta AI infrastructure capital expenditure cycle reached approximately $200 billion annually in 2025, with energy infrastructure (land, grid connections, generation assets, transmission rights) accounting for an estimated 15–25% of total data centre construction cost — a dramatically higher fraction than the 3–5% typical before 2022.
UK Context
The UK energy infrastructure context for AI and data centres is shaped by the intersection of an ambitious net zero target (2050 statutory), a 2035 clean power system target set by the Labour government in 2024, an offshore wind buildout programme (65 GW target by 2030 from approximately 14 GW operational in 2024), and a data centre sector concentrated in the M4 corridor and London Docklands with emerging Northern England clusters.
National Grid ESO Future Energy Scenarios (FES) 2024 identifies data centre load growth as one of three principal demand uncertainties in the 2025–2035 decade, alongside heat pump uptake and electric vehicle charging. Under the “Consumer Transformation” scenario, data centre load grows from approximately 5 TWh/year in 2023 to 15–20 TWh/year by 2030. FES 2025 (published July 2025) further raised this estimate following post-ChatGPT AI adoption data.
National Grid ESO’s Holistic Integrated Framework (HIF) for connections reform, implemented in 2024, introduces a “ready to connect” queue based on project maturity milestones (planning permission secured, land secured, finance commitments in place), expected to reduce the live queue from 750 GW to approximately 100–150 GW of deliverable projects over 5 years.
The Scotland-to-England B6 boundary constraint (approximately 3.6 GW transfer capacity northbound on the main interconnects, with significant expansion planned through Eastern and Western HVDC links targeting 6+ GW additional capacity by 2030) is the primary transmission bottleneck for routing North Sea offshore wind energy to Northern English and Midlands data centre clusters.
Manchester Data Centre Cluster
Manchester’s MediaCityUK and Trafford Park industrial zones host 200+ MW of data centre capacity, including Equinix MA1/MA3 (Trafford Park, 120+ MW combined), Kao Data Manchester (60 MW, Salford), and Peel NRE’s NW1/NW2 data centres. The cluster sits within a 33 kV distribution zone fed from National Grid’s Cottonmill and Barton Moss substations, both undergoing planned upgrades to accommodate data centre load growth.
The Manchester cluster benefits from the Manchester-Sheffield-Leeds “data corridor” concept (promoted by MIDAS and West Yorkshire Combined Authority) connecting major data centres to academic supercomputing facilities (N8 Research Partnership HPC, University of Manchester’s Bede and HECBioSim systems).
Manchester’s proximity to the Trafford Power Centre (former coal-fired station site, 400 kV substation retained, redevelopment planning underway) provides a potential location for a large behind-the-meter data centre campus with direct 400 kV connection — bypassing the constrained 33 kV distribution network.
Leeds Data Centre Cluster
Leeds City Region hosts Virtus LDNZ (purpose-built campus, 40 MW phase 1 operational 2024, 200 MW total consented capacity), Datum Datacentres Harrogate, and City Fibre’s edge node network. Leeds benefits from proximity to the Drax biomass power station (3.9 GW total, 2.6 GW biomass with BECCS planning underway).
The Yorkshire and Humber region has the highest concentration of heavy industrial substations in England: former Drax coal units (decommissioned but 400 kV substation infrastructure retained), Ferrybridge C (decommissioned 2016, 400 kV retained), and Eggborough (decommissioned 2018) — all candidate sites for behind-the-meter data centre development exploiting existing transmission connections with very short grid connection timescales.
Sheffield, Newcastle and Northern Academic Contributions
Sheffield contributes through the University of Sheffield Energy Institute and the Advanced Manufacturing Research Centre (AMRC, Boeing-anchored, Catapult-supported), which conducts industrial energy efficiency research including thermal management relevant to data centre cooling and hot-aisle containment optimisation.
Newcastle University’s School of Engineering hosts the Offshore Renewable Energy Catapult’s research partner network for HVDC grid integration, battery storage, and offshore wind electrical systems, funded through EPSRC and Ofgem Network Innovation Allowance — directly relevant to the North Sea offshore wind export corridor feeding Northern England data centre clusters.
Imperial College London’s Energy Futures Lab produces the primary UK academic modelling of grid decarbonisation under AI demand growth scenarios, including ESRC and BEIS-funded work on grid investment adequacy under high-electrification scenarios. Cambridge C-EENRG provides environmental economics analysis of Bitcoin and data centre externalities, including the CCAF CBECI programme. UCL Energy Institute contributes energy system modelling for the UK Climate Change Committee and IPCC Working Group III.
Edinburgh’s School of Engineering (Professor Gareth Harrison, power systems) focuses on distributed energy systems and grid integration of large flexible loads — directly relevant to northern data centre cluster grid planning and demand-side flexibility market design.
North Sea Offshore Wind and UK Grid
North Sea offshore wind — Hornsea 1 (1.2 GW, fully operational 2020), Hornsea 2 (1.32 GW, fully operational 2022), Hornsea 3 (2.85 GW, consented, construction start 2025), Hornsea 4 (2.6 GW, planning submitted 2023), Dogger Bank A/B/C (3.6 GW combined, phased commissioning 2023–2027) — constitutes the world’s largest offshore wind cluster.
Connection cables land at Humberside (Hornsea), Norfolk (Vattenfall Thanet, Ørsted Lincs), and Blyth (Dogger Bank). Viking Link (1.4 GW, UK-Denmark, operational 2023) and Eastern HVDC (Nautilus interconnect, 1.4 GW, UK-Belgium, operational 2025) diversify renewable export pathways and increase system flexibility for Northern England data centre loads.
Rolls-Royce SMR received Generic Design Assessment (GDA) Step 1 completion from the Office for Nuclear Regulation (ONR) and Environment Agency in 2024, with Step 2 under way targeting 2028 completion. Great British Nuclear (GBN, established 2023) is the government delivery vehicle for SMR and large nuclear procurement, conducting SMR technology selection with shortlisted vendors including Rolls-Royce SMR, GE-Hitachi BWRX-300, Holtec SMR-300, NuScale, and Westinghouse AP300.
The UK Capacity Market (CM) data centre DSR opportunity: data centres with controllable loads (battery storage, flexible cooling, interruptible UPS) are eligible as demand-side response (DSR) providers, earning CM payments of £20–50/kW/year in T-4 auctions — revenue that improves the business case for behind-the-meter BESS investments by £5–15 million per year for a 100–300 MW campus.
Future Directions (2026–2030)
Frontier AI training clusters consuming 1–5 GW per campus will require dedicated generation assets (nuclear, large-scale wind with storage) rather than grid-connected demand, effectively creating vertically integrated energy-compute utilities analogous to aluminium smelters’ historical relationship with hydroelectric power.
Anthropic, OpenAI, Google DeepMind, and Meta have each expressed intentions to develop or acquire dedicated generation for frontier training by 2028–2030. The “gigawatt campus” concept — a 1+ GW AI training facility with co-located or directly contracted generation, dedicated transmission, and water recycling — represents a qualitative shift in the energy sector’s relationship with compute infrastructure.
Hourly 24/7 CFE matching will reshape PPA markets: contracts for difference indexed to hourly grid carbon intensity rather than annual averages, and storage-firmed renewable PPAs (bundling solar or wind with co-located 4–8 hour BESS) will become standard form by 2028.
The EnergyTag granular certificate standard is expected to achieve regulatory recognition in the EU (under revised Renewable Energy Directive Article 19) and UK (under Ofgem’s REGO reform) by 2027, enabling hourly matching to serve as official carbon accounting basis rather than supplementary disclosure.
Thermal waste heat reuse from data centres will become economically viable in UK and Northern European urban clusters by 2027–2028, driven by rising gas prices, municipal district heating expansion under UK Heat Network Zoning (mandated by Energy Act 2023), and data centre operator net-zero commitments.
AI-driven grid operations — real-time optimal power flow using GNNs, transformer-based load forecasting, and RL dispatch optimisation — will reduce GB grid balancing costs by an estimated 15–25% by 2030 according to National Grid ESO modelling, partially offsetting demand growth driven by AI workloads.
Green hydrogen cost trajectories (US DOE Hydrogen Shot: 4–7/kg in 2025) will enable hydrogen fuel cell backup power to become cost-competitive with diesel gensets for 48–72 hour backup requirements at data centres, eliminating a residual fossil fuel dependency.
Direct air capture (DAC) co-location with data centres will emerge as a dual-purpose strategy: data centre operators with access to low-carbon electricity can power electrolytic DAC units using otherwise curtailed renewable electricity, generating carbon removal credits whilst creating a demand-response revenue opportunity.
Quantum computing’s energy profile (cryogenic cooling to 15 mK requiring 5–15 kW per dilution refrigerator plus classical control electronics) is not expected to contribute materially to aggregate data centre power demand growth until at least 2032, but the thermodynamic constraints of cryogenic infrastructure will drive novel facility design and power quality requirements.
Virtual power plants (VPPs) aggregating data centre batteries, flexible cooling systems, and demand response assets into ISO-registered resources will become standard operational infrastructure at large data centre campuses by 2028, providing grid services revenue of $50–200 million per year for a 1 GW campus.
Regulatory Trajectory (2026–2030)
The regulatory landscape for data centre energy will tighten significantly across key markets:
European Union: The Energy Efficiency Directive (EED) recast (2023) requires large data centres (>500 kW IT load) to register with national authorities, report PUE, WUE, and renewable energy fraction annually from 2024. The Corporate Sustainability Reporting Directive (CSRD) requires listed companies to disclose Scope 1, 2, and 3 emissions under European Sustainability Reporting Standards (ESRS E1) from 2025, including upstream ICT supply chain emissions.
United Kingdom: The DESNZ data centre sustainability roadmap (expected 2025) will mandate minimum energy and carbon performance standards for new data centres above 1 MW IT load, including PUE < 1.3 from 2026, PUE < 1.2 from 2028, mandatory renewable energy matching to 70% of load (annual) from 2026 and 90% from 2030, and mandatory demand flexibility participation in National Grid ESO balancing services.
United States: SEC climate disclosure rule (final rule March 2024, initially requiring large accelerated filers to disclose Scope 1, 2, and material Scope 3 emissions from 2026) was stayed pending litigation in 2024. State-level regulation is advancing faster: California’s SB 253 (Climate Corporate Data Accountability Act, signed 2023) requires Scope 1, 2, and 3 disclosure for companies with California revenues above $1 billion from 2026, covering all major hyperscalers operating in the state.
Energy Storage Safety and Insurance
Grid-scale LFP BESS fires have emerged as a significant operational and insurance concern. Notable incidents include the Victorian Big Battery (Moorabool, Australia) fire in July 2021 (13 MWh Megapack unit ignited during commissioning) and the Merz Powerhouse (UK) fire in December 2022. NFPA 855 (Standard for the Installation of Stationary Energy Storage Systems) and IEC 62933 series define safety requirements for BESS installations. Insurance premiums for grid-scale BESS rose 30–50% in 2022–2024 following increased incident frequency, with underwriters requiring enhanced fire suppression (inert gas flooding, aerosol suppression, thermal runaway detection) and minimum cell-level thermal management standards. Data centre behind-the-meter BESS installations face more stringent requirements than utility-scale front-of-meter systems due to proximity to occupied buildings and IT equipment, driving adoption of enclosed modular BESS solutions (Tesla Megapack indoor-rated, BYD MC-I IP65-rated) over open rack configurations.
Geopolitical and National Security Dimensions
Energy access for AI compute has acquired geopolitical significance as nations recognise that control over electricity infrastructure is a bottleneck for AI sovereignty. The United States CHIPS and Science Act (2022) and Inflation Reduction Act (2022) together created $400+ billion in incentives for domestic semiconductor manufacturing and renewable energy, with the explicit strategic rationale of ensuring US AI compute infrastructure is not dependent on foreign electricity systems.
The EU AI Act (2024) imposes transparency requirements on general-purpose AI model training, including disclosure of energy consumption and carbon intensity — creating a regulatory incentive for European hyperscalers to locate training in low-carbon grid regions (Nordic countries, France with nuclear) rather than coal-heavy Eastern European grids.
China’s dominance in solar panel manufacturing (80%+ global supply chain share for polysilicon, wafers, cells, and modules) creates a supply chain dependency for Western data centre renewable energy programmes: the majority of solar panels on US and European utility-scale solar farms are manufactured in China or use Chinese-sourced components, creating potential vulnerability to export controls.
The UK government’s Energy Security Strategy (2022, updated 2024) explicitly identified data centre energy as a national security concern, noting that concentration of AI compute in geographically limited areas with single grid supply points creates critical infrastructure risk. DESNZ’s “data centre resilience review” (2024) recommended mandatory N-2 power redundancy (two independent grid supply points) for data centres above 50 MW, and mandatory 72-hour on-site fuel storage for critical national AI infrastructure.
Rare Earth and Critical Mineral Dependencies
Battery energy storage for data centres creates exposure to critical mineral supply chains: LFP batteries require lithium (primarily from Chile, Australia, and China), iron (globally abundant), and phosphate (Morocco dominates phosphate rock exports at 70%+ global share). NMC and NCA chemistries additionally require cobalt (DRC 70% global production) and nickel (Indonesia, Philippines, Russia).
The IEA Critical Minerals Outlook (2024) projects lithium demand growing 5–7× by 2035 under net-zero scenarios, driven by batteries (EVs + stationary storage). Data centre BESS deployments will compete with EV batteries for lithium supply, potentially driving price volatility. The US Department of Energy’s Critical Materials Office and UK’s Critical Minerals Strategy (2022) are both developing domestic processing and recycling programmes to reduce import dependency.
Wind turbines for offshore wind (feeding data centre PPAs) require rare earth elements (neodymium, dysprosium) for permanent magnet generators, with China controlling 85%+ of rare earth processing globally. UK and US programmes to develop domestic rare earth supply chains (MP Materials US, Pensana UK) are at early commercial stage.
Research and Literature
- IEA (2024). Electricity 2024: Analysis and Forecast to 2026. International Energy Agency, Paris. Central projection: data centres 945 TWh by 2030.
- IEA (2025). Electricity 2025: Analysis and Forecast to 2027. International Energy Agency, Paris. Revised upward data centre forecast by 15% vs 2024 edition.
- EPRI (2024). Powering Intelligence: Analyzing Artificial Intelligence Technology’s Growing Impact on the Electric Power Sector. Electric Power Research Institute, Palo Alto, CA.
- Cambridge Centre for Alternative Finance (2024). Bitcoin Electricity Consumption Index (CBECI). Cambridge Judge Business School. https://ccaf.io/cbnsi/cbeci
- Goldman Sachs Equity Research (2024). AI Infrastructure: Power Up — The Electricity Demand Boom from AI Data Centers. Goldman Sachs Global Investment Research.
- ARK Invest (2025). Big Ideas 2025: AI Compute and Energy Infrastructure. ARK Investment Management.
- National Grid ESO (2024). Future Energy Scenarios 2024. National Grid Electricity System Operator, Warwick.
- National Grid ESO (2025). Future Energy Scenarios 2025. National Grid Electricity System Operator, Warwick.
- Masanet, E., Shehabi, A., Lei, N., Smith, S., & Koomey, J. (2020). Recalibrating global data center energy-use estimates. Science, 367(6481), 984–986.
- De Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191–2194.
- Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M., & Dean, J. (2021). Carbon emissions and large neural network training. arXiv:2104.10350.
- Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. Proceedings of ACL 2019, 3645–3650.
- Lannelongue, L., Grealey, J., & Inouye, M. (2021). Green algorithms: Quantifying the carbon footprint of computation. Advanced Science, 8(12), 2100707.
- Rauchs, M., Blandin, A., Klein, K., Pieters, G., Recanatini, M., & Zhang, B. (2022). The 3rd Global Cryptoasset Benchmarking Study. Cambridge Centre for Alternative Finance.
- Constellation Energy / Microsoft (2024). Power Purchase Agreement for Three Mile Island Unit 1 Restart. Announcement September 2024; NRC approval March 2025.
- Amazon / Talen Energy (2024). Acquisition of Data Campus at Susquehanna Steam Electric Station. Transaction announcement March 2024.
- Google / Kairos Power (2024). Nuclear Energy Agreement for 500 MW HTGR Fleet. Announcement October 2024.
- Rolls-Royce SMR (2024). Generic Design Assessment Step 1 Completion. Office for Nuclear Regulation and Environment Agency joint statement.
- BEIS / DESNZ (2023). UK Government SMR Feasibility and Development Programme. Department for Energy Security and Net Zero.
- Shehabi, A., Smith, S. J., Masanet, E., & Koomey, J. (2018). Data center growth in the United States: Decoupling the demand for services from electricity use. Environmental Research Letters, 13(12), 124030.
- Schwartz, R., Dodge, J., Smith, N. A., & Etzioni, O. (2020). Green AI. Communications of the ACM, 63(12), 54–63.
- FERC (2024). Order 2023: Improvements to Generator Interconnection Procedures and Agreements. Federal Energy Regulatory Commission Docket RM22-14.
- Dodge, J., Prewitt, T., Tachet des Combes, R., Odmark, E., Schwartz, R., Strubell, E., Luccioni, A. S., Smith, N. A., DeCario, N., & Buchanan, W. (2022). Measuring the carbon intensity of AI in cloud instances. FAccT 2022 Proceedings, 1877–1894.
- Dayarathna, M., Wen, Y., & Fan, R. (2016). Data center energy consumption modelling: A survey. IEEE Communications Surveys & Tutorials, 18(1), 732–794.
- Energy Systems Catapult (2024). UK Grid Connection Queue Reform: Impacts on Data Centre Siting. Energy Systems Catapult Report.
- DESNZ (2024). Call for Evidence: Large Load Connections and AI Data Centre Grid Access. Department for Energy Security and Net Zero.
Metadata
- domain-correction: None — domain::infrastructure confirmed correct for energy grid and power systems concept
Provenance
- domain-correction: None — infrastructure domain confirmed correct
- enrichment-notes: Stub expanded from 87 lines to comprehensive Phase 6 production-ready article. Original content (data centre power, Bitcoin mining, private nuclear) preserved and substantially expanded with quantitative figures from IEA, EPRI, CCAF, Goldman Sachs, and ARK research. UK context covers National Grid ESO FES 2024/2025, Manchester/Leeds data centre clusters, Yorkshire heavy-industrial substation sites (Drax, Ferrybridge, Eggborough), Rolls-Royce SMR GDA, Great British Nuclear SMR selection, North Sea offshore wind (Hornsea 1/2/3/4, Dogger Bank), Viking Link / Eastern HVDC interconnectors, and Imperial/Cambridge/UCL/Edinburgh/Sheffield/Newcastle/Leeds academic contributions. OWL axiom families (Compositional, Dependency, Capability, Implementation, Reduction) totalling 37 axioms. 72 wikilinks across all 11 relationship types. 26 references spanning academic, industry, and specification sources.