Executive Summary
Artificial intelligence has crossed a threshold from experimental technology to mission-critical global infrastructure — and it is bringing an unprecedented electricity demand surge with it. Global data centre electricity consumption reached approximately 415 TWh in 2024, representing around 1.5% of world electricity use, growing at 12% per year since 2017 — more than four times faster than total electricity consumption growth. The International Energy Agency (IEA) projects this to nearly double to 945 TWh by 2030 in its base case.
The United States holds the largest share (45%) of global data centre electricity consumption, followed by China (25%) and Europe (15%). U.S. data centres alone consumed 183 TWh in 2024 — equivalent to the annual electricity demand of Pakistan — and this figure is projected to rise 133% to 426 TWh by 2030.
This teaser maps the full energy-AI ecosystem: the scale of the challenge, the technology innovations addressing it, and the investment landscape forming around each layer — from silicon to the grid. The opportunity for industry leaders is enormous. Global investment in data centres nearly doubled since 2022, reaching half a trillion dollars in 2024.
Key Figures at Glance

Related Topics
Section 1: The Scale of AI's Energy Demand
1.1 The Numbers in Context
AI-focused hyperscale data centres draw as much electricity as power-intensive industrial factories such as aluminium smelters, but they are far more geographically concentrated — nearly half of U.S. data centre capacity sits in five regional clusters. A single AI-focused hyperscaler annually consumes electricity equivalent to 100,000 households; the largest under construction are projected to use 20 times as much.
Numbers in Context
1.2 AI Inference: The Dominant Driver
Model training draws headlines, but inference now dominates AI energy consumption, estimated at 80–90% of total AI computing energy. Every ChatGPT query consumes almost ten times more power than a traditional Google search. As AI features embed into daily consumer and enterprise products, inference workloads are multiplying exponentially.
AI-specific servers in U.S. data centres used an estimated 53–76 TWh in 2024. Projections suggest 165–326 TWh by 2028, representing a potential 3–6× increase in just four years.
1.3 Water and Physical Infrastructure Pressure
Energy is not the only resource under pressure. U.S. data centres directly consumed approximately 17 billion gallons of water in 2023 for cooling. Hyperscale and colocation facilities accounted for 84% of this total. By 2028, hyperscale water consumption alone is projected to reach 16–33 billion gallons annually — before accounting for indirect water used in electricity generation or semiconductor manufacturing.
Section 2: Hardware Innovation — Doing More Per Watt
2.1 The Chip Revolution: NVIDIA Blackwell Architecture
The most direct lever for reducing AI's energy intensity lies in the silicon itself. NVIDIA's Blackwell GPU architecture, unveiled in March 2024 and deployed commercially in 2025, delivers up to 25× less energy consumption for large language model inference compared to its predecessor, the Hopper architecture (H100). This efficiency gain stems from six core innovations: a dual-reticle 208-billion transistor design using TSMC's 4NP process; a second-generation Transformer Engine supporting FP4 precision; fifth-generation NVLink connecting up to 576 GPUs; 8 TB of HBM3e memory; a fifth-generation NVSwitch; and RAS (Reliability, Availability, Serviceability) engine.
NVIDIA GB300 NVL72 - Key Energy Statistics
- Delivers 50× higher output for reasoning model inference vs. Hopper platform
- Fully liquid-cooled rack-scale design (mandatory for 1kW+ per GPU TDP)
- NVIDIA technologies power 23 of the top 30 systems on the Green500 energy-efficiency supercomputer list
- Post-Blackwell 'Rubin' architecture already announced — continuous efficiency roadmap
2.2 The Efficiency-Demand Paradox (Jevons Effect)
Despite hardware efficiency gains, overall AI energy demand continues to rise — a textbook example of Jevons Paradox. As AI becomes cheaper and faster per query, total queries scale rapidly. Google reports that between May 2024 and May 2025, it reduced median energy consumption per Gemini prompt by a factor of 33 and the associated carbon footprint by a factor of 44 — yet Google's total energy demand continues to climb, driven by broader AI adoption.
Section 3: Cooling Technology — From Air to Liquid-First
3.1 Why Air Cooling Is No Longer Sufficient
Modern AI chips such as the NVIDIA H100 and AMD MI300X generate heat loads exceeding 700W per GPU. A single Blackwell Ultra NVL72 rack consumes up to 140 kW — far beyond what traditional air cooling can manage. Data centre cooling has shifted from an engineering afterthought to a primary constraint on AI infrastructure deployment. By 2024, liquid-based cooling had captured 46% of the total data centre cooling market; industry analysts expect liquid cooling to dominate all new AI and hyperscale builds within two to three years.
3.2 Three Cooling Modalities in the Market
A. Direct-to-Chip (Cold Plate) Cooling
Coolant is delivered through precisely engineered cold plates directly to CPUs and GPUs. Direct-to-chip cooling commands a projected 47% share of the AI data centre liquid cooling market in 2025. Next-generation cold plates from Accelsius and CoolIT are targeting heat flux of up to 300 W/cm², while Frore is demonstrating 600 W/cm² capability. AWS commercially launched its proprietary In-Row Heat Exchanger (IRHX) liquid cooling system in 2025, claiming up to 46% reduction in mechanical energy use.
B. Single-Phase & Two-Phase Immersion Cooling
Immersion cooling submerges servers in dielectric fluid, with two-phase systems achieving heat flux of up to 1,500 W/cm² — far exceeding air cooling's theoretical limit of ~16 W/cm². In May 2025, G42 (Abu Dhabi-based AI and cloud company) partnered with LiquidStack to build an AI supercomputing centre in Masdar City using two-phase immersion cooling. Firmus's Project Southgate secured $327 million in funding for a 1,600 MW campus incorporating full immersion cooling. Chemours announced a strategic agreement with DataVolt in May 2025 to advance two-phase immersion cooling using its Opteon™ fluid, claiming up to 90% cooling energy reduction and near elimination of water use.
3.3 Colocation Operators: Equinix as the Benchmark
As a global colocation provider, Equinix has committed to expanding direct-to-chip liquid cooling support to more than 100 of its International Business Exchange (IBX) data centres across more than 45 metros worldwide — including London, Silicon Valley, Singapore, and Washington D.C. The company operates a Co-Innovation Facility (CIF) in Ashburn, Virginia to evaluate and validate next-generation cooling technologies from partners including Accelsius, JetCool, and ZutaCore. A pilot at Equinix HK1 in Hong Kong — developed with Dell Technologies and Schneider Electric — demonstrated a 50 kW in-rack direct-to-chip system delivering up to 150 kW of cooling per rack (30× conventional air cooling capacity) with a Power Usage Effectiveness (PUE) of 1.2 or lower, compared to 1.4–1.5 for traditional CRAC systems. AI workloads are driving rack power from 5–10 kW to over 100 kW; current liquid-cooled deployments at Equinix typically operate at an 80/20 liquid-to-air split.
3.4 Market Size and Key Players
The global data centre liquid cooling market was valued at $3.93 billion in 2024, projected to reach approximately $22.57 billion by 2034 at a 19.1% CAGR. The AI-specific liquid cooling sub-segment is projected to grow from $6.6 billion in 2025 to $61.8 billion by 2034. Key market participants include Vertiv (liquid cooling revenue more than doubled in Q1 2025, projected at 40% CAGR through 2028), Schneider Electric (acquired Motivair for $850M in October 2024 to strengthen cold plate capabilities), CoolIT Systems, LiquidStack, Submer, and Iceotope. Schneider Electric also holds a reference design partnership with NVIDIA for the GB200 NVL72 system.
Section 4: Energy Sourcing — The Nuclear Revival
4.1 Why 24/7 Carbon-Free Power Is Non-Negotiable
Wind and solar provide intermittent generation. AI data centres require near-zero downtime and always-on power at massive, predictable scale. This fundamental mismatch is driving hyperscalers toward nuclear power — the only established low-carbon technology that delivers baseload, dispatchable electricity around the clock. As of 2024, natural gas supplied over 40% of U.S. data centre electricity, renewables about 24%, nuclear around 20%, and coal approximately 15%.
4.2 Big Tech's Nuclear Commitments — Verified Deals
In just 12 months spanning 2024 and 2025, the world's largest AI companies signed contracts for more than 10 gigawatts of possible new nuclear capacity in the United States alone. The following deals are sourced from company announcements and partner press releases:
Sample Companies & Recent Deals
Section 5: Renewable Energy & Battery Storage
5.1 Hyperscalers as the World's Largest Clean Energy Buyers
Amazon, Meta, Google, and Microsoft collectively accounted for 49% of all global corporate clean energy power purchase agreement (PPA) volumes in 2025 — even as overall corporate clean energy PPA volumes globally fell 10% to 55.9 GW, the first decline after eight consecutive years of growth. Big Tech's dominance in this market accelerated when other corporate buyers pulled back.
Microsoft surpassed Amazon as the largest corporate clean energy buyer globally as of end-September 2025, with 34.7 GW contracted. Amazon had previously led with nearly 35 GW. U.S. data centres have now contracted more than 80 GW of clean energy — approximately 30 GW above estimates made just one year earlier, and more than half of all U.S. corporate clean energy procurement.
5.2 Battery Energy Storage: The Grid's Shock Absorber
Battery Energy Storage Systems (BESS) serve two purposes in the AI energy ecosystem: they provide on-site resilience for data centres experiencing power grid interruptions, and they help integrate the intermittent renewable generation that hyperscalers procure. The global BESS market is projected to grow from $76.7 billion in 2025 to $172.2 billion by 2030, with installed capacity rising from 200 GWh to 1,200 GWh over the same period.
A critical, practical driver: data centre developers who deploy behind-the-meter batteries alongside their facilities can navigate grid interconnection queues significantly faster — by years in some cases. A strategic BESS deployment could unlock up to 76 GW of new U.S. data centre capacity, according to modelling cited by FlexGen. U.S. utility-scale battery capacity grew 66% in 2024 alone; global BESS shipments reached 421.2 GWh in 2025, a 75.5% year-on-year increase.
Long-duration energy storage (LDES) is gaining momentum, with 2025 installations reaching 15 GWh — a 49% increase over 2024. However, LDES represents just 6% of all energy storage installations, constrained by policy gaps and financing challenges. The global AI data centre energy storage market is projected to grow from approximately $1.2 billion in 2025 to $4.1–6.0 billion by 2030, at a 28–38% CAGR.
Section 6: AI Optimising the Energy System — Closing the Loop
6.1 AI as a Grid Optimisation Tool
The relationship between AI and energy is not one-directional. AI is increasingly deployed to optimise the very grid that powers it, creating a virtuous cycle when done well. Published academic research in peer-reviewed journals documents AI's role in improving load forecasting accuracy, optimising power distribution, automating fault detection, and enabling demand-response programs. These capabilities allow greater penetration of variable renewable energy by dynamically balancing supply and demand.
6.2 Verified Industry Applications
Demand response for data centres emerged as a major theme at GridFuture 2025, hosted in Washington, DC. Innovators from Mercury Computing, Emerald AI, Bentaus, and Miratech demonstrated technologies enabling data centres to adjust power use dynamically — a development described by industry participants as the energy sector's 'holy grail.' Google partnered with PJM Interconnection (the largest U.S. grid operator) and its Alphabet subsidiary Tapestry in April 2025 to modernise U.S. grid infrastructure using AI-driven analysis.
Microsoft is investing in AI tools to design and test materials with greater accuracy, better manage water resources, and expedite the licensing process for carbon-free electricity, alongside its stated goal to add 10.5 GW of renewable energy to the electricity grid.
Section 7: Ecosystem Map & Investment Landscape
7.1 Capital Flows: The Scale of Commitment
Global data centre investment nearly doubled since 2022, reaching approximately half a trillion dollars in 2024. In the U.S., hyperscalers Google, Meta, and Amazon were estimated to spend $364 billion on data centre construction in 2025 alone. These companies are collectively the largest corporate buyers of renewable energy in the world. Big Tech companies accounted for 43% of all clean energy PPAs signed globally in 2024, driving PPA prices up by an average of 35%.
7.2 The Ecosystem — Layer by Layer
Layer 1: Silicon & Semiconductors
- NVIDIA (NVDA): Blackwell and Blackwell Ultra GPU platforms; 25× LLM inference efficiency gain; roadmap to 'Rubin' architecture
- AMD: MI300X series competing for AI workloads
- TSMC: CoWoS-L advanced packaging enabling multi-die GPU designs
- SK Hynix, Micron: HBM3E memory for high-bandwidth AI compute
Layer 2: Cooling Infrastructure
- Equinix: Expanded direct-to-chip liquid cooling to 100+ data centres across 45+ metros globally; pilot in Hong Kong with Dell and Schneider Electric delivering 150kW of cooling per rack (30× conventional air); Co-Innovation Facility (CIF) in Ashburn, VA testing next-generation immersion and rear-door heat exchanger solutions
- Vertiv: CDU market leader; liquid cooling revenue doubled in Q1 2025; 40% CAGR projected through 2028; $71B market cap
- Schneider Electric: Acquired Motivair ($850M, Oct 2024); reference design partner with NVIDIA for GB200 NVL72
- CoolIT Systems: Cold plate CDUs for multi-megawatt deployments
- LiquidStack: Two-phase immersion; G42 partnership in Masdar City
- Submer: Barcelona-based immersion cooling for AI data centres
- Chemours (Opteon™): Dielectric immersion fluids; DataVolt partnership (May 2025)
Layer 3: Power Generation & Nuclear
- Constellation Energy: Three Mile Island restart (Microsoft PPA); Clinton Clean Energy Center (Meta PPA 1.1 GW)
- Kairos Power: SMR developer; Google 500 MW agreement; NRC construction license (Hermes, Oak Ridge, TN)
- X-Energy: High-temperature gas reactors; Amazon $500M+ commitment; TRISO fuel
- Talen Energy: Susquehanna nuclear plant; Amazon PPA (1.9 GW through 2042)
- Vistra / TerraPower: Meta 2,600+ MW PPA; Natrium advanced reactor
- Oklo: 50 MW modular powerhouses; Meta 1.2 GW deal
Layer 4: Renewables & Offtake
- TotalEnergies: 1.5+ GW in B2B renewable PPAs by March 2024; Amazon (474 MW US/Europe) and Microsoft (Spain) agreements
- Brookfield Renewable: $3B+ hydro framework with Google (up to 3,000 MW)
- NextEra Energy: Nuclear and renewables partnership with Google
- RE100 Initiative: 400+ companies, 500 TWh combined annual demand, committed to 100% renewable power
Layer 5: Energy Storage
- Battery Council International: BESS market projected $76.7B → $172.2B (2025–2030)
- FlexGen: BESS solutions for data centre interconnection acceleration
- Form Energy: Iron-air batteries for 100-hour long-duration storage (Xcel Energy deployment)
- Hydrostor: Advanced compressed air energy storage; $1.8B DOE commitment
Layer 6: Onsite Fuel Cell Power
- Bloom Energy × Oracle (Project Jupiter): April 2026 — Oracle and BorderPlex Digital Assets announced up to 2.45 GW of Bloom Energy solid oxide fuel cells (SOFCs) to fully power Project Jupiter AI data centre campus in Doña Ana County, New Mexico — expected to be the largest data centre microgrid in the U.S. at completion. Master services agreement covers up to 2.8 GW total across Oracle U.S. data centres; initial 1.2 GW contracted with deployment underway. Bloom can deliver onsite power for an entire data centre within 90 days.
- Bloom Energy × Equinix: 100 MW fuel cell agreement to provide primary baseload power for Equinix data centre operations, validated as a scalable model for hyperscale colocation.
- Bloom's SOFC technology: Electrochemical conversion (not combustion) — high efficiency, negligible air pollutants (NOx, SOx), no water use; fuel-agnostic (natural gas, biogas); aligned to emerging 800V DC AI data centre standards.
Layer 7: Gas Turbines — Fast-Deployable Dispatchable Power
- GE Vernova × Crusoe: 29-unit LM2500XPRESS aeroderivative gas turbine deal (10 units Dec 2024 + 19 units June 2025) — nearly 1 GW of electricity for Crusoe AI data centres. Each 35 MW unit has 5-minute fast-start capability, 95% factory-assembled for rapid installation.
- GE Vernova × Chevron / Engine No. 1 × Microsoft: Joint venture to develop 4 GW of gas-fired power, including capacity to support Microsoft AI data centres.
- GE Vernova × NRG Energy / Kiewit: Partnership targeting up to 5.4 GW of new gas capacity for U.S. data centre corridors.
- GE Vernova × AWS: Strategic framework agreement supplying gas turbines and turnkey grid substations for AWS infrastructure.
- Siemens Energy × Eaton: June 2025 partnership to enable simultaneous construction of data centres and associated onsite power generation with grid connection.
Section 8: Challenges, Risks & What Is at Stake
8.1 Grid Infrastructure and Rate Impacts
The concentration of massive data centre loads in specific geographies is creating acute stress on legacy grid infrastructure, much of which is over 60 years old. In the PJM electricity market (Illinois to North Carolina), data centres contributed to an estimated $9.3 billion price increase in the 2025–26 capacity market. Average residential bills in western Maryland are expected to rise by $18/month and in Ohio by $16/month. A Carnegie Mellon University study estimates data centres and cryptocurrency mining could increase average U.S. electricity bills by 8% by 2030, potentially exceeding 25% in the highest-demand markets of central and northern Virginia.
8.2 Nuclear: Promise vs. Timeline Risk
The nuclear revival faces a critical distinction between proven and speculative assets. Microsoft's Three Mile Island restart uses an established reactor (targeting 2028). In contrast, no SMRs currently operate commercially in the United States. Google's Kairos Power deal targets first reactors by 2030 under an optimistic scenario; Amazon's X-Energy SMRs face similar timelines. Regulatory, cost, and supply chain risks are material. The first planned commercial SMR in the U.S. — NuScale's Carbon Free Power Project — was cancelled in 2023 due to rising costs and insufficient customer commitments.
8.3 Regulatory and Disclosure Gaps
Most major closed AI model providers do not disclose sufficient information to reliably estimate their total energy use or carbon footprint. The U.S. Clean Cloud Act of 2025 (S.1475), introduced in the Senate, seeks to amend the Clean Air Act to improve transparency. In 2024, the EIA attempted to collect mandatory data on cryptocurrency mining data centres but ceased collection following a federal court agreement — illustrating the current limits of regulatory oversight over AI-adjacent energy consumers.
Section 9: Closing Perspective — An Invitation to Go Deeper
The energy challenge facing AI is not a temporary bottleneck. It is a structural, decade-long transformation of electricity systems, capital markets, hardware design, and geopolitical competition. The IEA describes this as a 'new age of electricity' — and data centres, while not the largest single driver of global demand growth, are uniquely concentrated, politically visible, and economically high-priority.
The prize for those who solve the equation is substantial. Global investment in data centres overtook that in crude oil supply for the first time in 2025, according to the IEA's World Energy Outlook 2025. AI infrastructure spending is projected to grow from approximately $400 billion in 2025 to over $1.5 trillion annually by 2030.
Five Questions For Industry Leaders
- How will your organisation position in the emerging 24/7 carbon-free power market as nuclear timelines remain uncertain?
- Is your cooling infrastructure roadmap aligned with the mandatory shift to liquid-first design for next-generation AI racks?
- What grid interconnection strategy will you deploy to stay ahead of multi-year queue backlogs in AI-dense markets?
- How are you capturing the opportunity in AI-driven grid optimisation — not just consuming energy, but managing it?
- What disclosure and regulatory posture will your organisation adopt as government oversight of AI energy use intensifies?
Sources & References
- AI Business · Meta Signs Nuclear Deals
- AIMultiple · AI Energy Consumption Statistics
- Battery Council International · Batteries Behind AI Data Centers
- Bloom Energy × Oracle Press Release (April 13, 2026)
- Bloom Energy SEC Filing · DEF 14A (2025)
- Brookings Institution · Global Energy Demands in the AI Regulatory Landscape (2026)
- Chemours · DataVolt Partnership Press Release (May 2025)
- CIO.inc · Liquid Cooling Future of AI Data Centers (September 2025)
- Congress.gov · CRS Report R48646
- CPower Energy · GridFuture 2025 Takeaways
- Energy Industry Review · Battery Storage Capacity 2026
- EnkiAI · TotalEnergies 2025 Data Center Strategy
- Equinix · Direct-to-Chip Liquid Cooling Press Release
- Equinix Blog · AI Engine Room (November 2025)
- Equinix SEC Filing 8-K · AI Cooling Strategy
- ESG Today · Amazon, Meta, Google, Microsoft (February 2026)
- FlexGen · Solving Data Center Power Needs with BESS (April 2025)
- Frontiers in AI · Role of AI in Smart Grid — Mini Review (2025)
- FutureIoT · Equinix HK1 Liquid Cooling Pilot
- GE Vernova × Crusoe Official Press Release (July 22, 2025)
- Georgetown Journal · AI Energy Demands and Nuclear (April 2025)
- IEA · Energy and AI (2025)
- IEA · Executive Summary
- IEA · World Energy Outlook 2025
- IEA Energy & AI (2025)
- IEEE Spectrum · Big Tech Embraces Nuclear Power
- Introl Blog · Nuclear Power for AI Data Centers (Dec 2025)
- Logistics Viewpoints · Navigating AI Energy Demands (October 2025)
- McKinsey · 24/7 Clean Power Purchase Agreements
- MLQ.ai · AI Data Center Cooling Report
- MLQ.ai · Data Center Cooling
- NVIDIA · Blackwell Architecture
- NVIDIA Developer Blog · Inside Blackwell Ultra
- NVIDIA Newsroom · Blackwell Platform Announcement
- Oracle × BorderPlex × Bloom · Project Jupiter Announcement (April 27, 2026)
- Pew Research Center · Data Centre Energy (October 2025)
- S&P Global Market Intelligence · Corporate PPA Leaderboard (March 2025)
- S&P Global Sustainable1 · Hyperscaler Procurement (December 2025)
- SEC EDGAR · Brookfield/Google Hydro Framework Agreement
- SpringerOpen Energy Informatics · Transforming the Electrical Grid (2025)
- Technavio · Liquid Cooling for AI Data Centers Market
- Technavio · Liquid Cooling Market (2025)
- Tom's Hardware · Data Centre Cooling State of Play (2025)
- Trellis.net · Amazon, Google, Meta, Google Go Nuclear
- U.S. DOE · 2024 Report on U.S. Data Center Energy Use
- Wood Mackenzie / Utility Dive · LDES Deployments Rose 49% (March 2026)
- Yahoo Finance / Research Report · AI Data Center Energy Storage (2025)
- Yale Clean Energy Forum · Power Hungry Power Smart (November 2025)
