AI & Data Center Energy

The AI buildout is driving the fastest increase in electricity demand in 20 years. Global data centres consumed ~415 TWh in 2024 — comparable to France. By 2030 this could reach 1,000–2,800 TWh depending on GPU cluster growth. The carbon footprint, water stress, and grid stability implications are substantial and largely unregulated at the international level.

415 TWh
DC electricity 2024
~1.4% of global electricity; IEA 2024
~2–3×
Expected growth by 2030
IEA central: ~1,000 TWh; hyperscale: ~1,550 TWh
~200 Mt
Operational CO₂/yr (2024)
Grid-mix weighted; excludes embodied
>$1 Trillion
Hyperscaler capex 2024–2028
Google, Microsoft, Meta, Amazon combined
0.50 L/kWh
Avg water usage (WUE 2024)
~200 km³/yr globally; water stress risk
45%
Hyperscaler share of DC load
5 companies; growing toward 60–70% by 2030
Uncertainty notice: AI electricity demand forecasts have a 3–5× range between pessimistic (efficiency breakthrough) and optimistic (unconstrained buildout) projections. Model training energy costs have historically doubled every 6–9 months; inference scaling is newer and less predictable. Water stress estimates depend on cooling technology mix (air vs water cooling) and vary widely by region.
Overview
Energy Demand
Carbon Footprint
Water & Cooling
Corporate Concentration
Timeline
Scientific Context
The grid challenge: The US alone is adding ~40 GW of new data centre load by 2030 — equivalent to building 40 nuclear power plants in 6 years. Most utilities plan to meet this with a mix of natural gas peakers and renewables. In the Hyperscale Boom scenario, data centres would consume more electricity than the entire EU grid by 2035.

Total DC Electricity by Scenario (TWh/yr)

Efficiency-Led Moderate Growth Hyperscale Boom Unconstrained

Total CO₂ Footprint (Operational + Embodied, Mt/yr)

Scenario Summary — 2030

ScenarioDC TWh 2030% Global electricity Total CO₂ (Mt)Water (km³/yr)PUE
Efficiency Led 650 TWh 1.93% 149.8 Mt 0.163 km³ 1.28
Moderate Growth 1000 TWh 2.97% 253.5 Mt 0.296 km³ 1.42
Hyperscale Boom 1550 TWh 4.6% 430.0 Mt 0.51 km³ 1.46
Unconstrained 2800 TWh 8.32% 827.3 Mt 1.0 km³ 1.54
AI vs total DC: In 2024 AI workloads (training + inference) account for ~24% of data centre electricity. By 2030, AI is projected to be 38–68% of total DC load depending on scenario, as inference demand from deployed models scales with user growth.

Total DC Electricity (TWh/yr)

AI-Specific Subset (TWh/yr)

DC Share of Global Electricity (%)

Average Power Usage Effectiveness (PUE)

PUE = total facility power / IT equipment power. 1.0 = perfect; 1.5 = 50% overhead for cooling. Hyperscalers currently achieve 1.1–1.2 at custom campuses; global average ~1.58.

Embodied carbon: GPU server manufacture accounts for ~3,500 kg CO₂e per rack server. With replacement cycles of 3–5 years (AI hardware obsolescence is rapid), hardware churn adds 10–20% to operational carbon. This is systematically excluded from corporate Scope 1+2 reporting.

Operational CO₂ (from electricity, Mt/yr)

Embodied Carbon — Hardware Churn (Mt/yr)

Total CO₂ Footprint (Operational + Embodied, Mt/yr)

Hidden water cost: Cooling data centres consumes both direct water (cooling towers, evaporative cooling) and indirect water (for thermal power plant cooling serving the grid). WRI Aqueduct 2023 shows >60% of planned US and European data centre expansions are in high or extremely high water stress areas.

Annual Water Consumption (km³/yr)

Water Stress Context

ComparatorVolume (km³/yr)
Global DC water consumption (2024 est.)~0.20
Rhine River annual flow~70
Colorado River (managed flow)~20
Nevada annual water allocation~3.5
Hyperscale Boom 2030 (this model)0.51
Unconstrained 2040 (this model)4.054

Water Efficiency Trends

Modern hyperscale facilities use liquid cooling (direct-to-chip, immersion) that can achieve WUE as low as 0.1–0.2 L/kWh vs the industry average of 0.5 L/kWh. However, older facilities (still the majority of the global installed base) use evaporative cooling towers at 0.7–1.2 L/kWh.

As AI clusters are deployed in existing co-location facilities (not just custom campuses), the average WUE is unlikely to fall as fast as pure hyperscaler numbers suggest. The efficiency-led scenario assumes aggressive adoption of liquid cooling; unconstrained assumes retrofit constraints keep WUE elevated.

Five companies, one global grid signal: Google, Microsoft, Amazon, Meta, and Apple collectively control ~45% of global data centre electricity — and are committing to 100% clean power. But their PPAs are creating a secondary market distortion: they purchase renewable certificates faster than new capacity is added, leaving other consumers on higher-carbon residual grid mix.

The PPA Paradox

Hyperscalers report near-zero Scope 2 emissions via Renewable Energy Certificates (RECs) and Power Purchase Agreements. But the physical electrons serving their data centres often come from the same gas-fired grid as everyone else.

Market-based accounting (Scope 2) vs location-based accounting can diverge by 200–300 gCO₂/kWh for US hyperscalers. The CE model uses a blended effective grid CI, not market-based.

Nuclear Renaissance — AI-Driven

Microsoft (Three Mile Island restart), Google (Kairos Power SMR deal), Amazon (X-energy), and Meta have all signed nuclear offtake agreements since 2023. AI's 24/7 baseload demand profile — unlike variable renewable offtake — makes it an ideal nuclear PPA buyer.

The CE model captures this in the Efficiency-Led and Moderate scenarios as a decline in effective grid CI for data centres specifically, diverging from the broader grid average.

Scope 3 — The Dark Number

AI model training and inference carbon is typically reported as Scope 1+2 by the cloud provider. For the enterprise customer using a cloud API, this is Scope 3 Category 1 (purchased goods) — and almost universally unreported in corporate sustainability disclosures.

Luccioni et al. 2023 estimates GPT-4-class model inference emits ~130–300 gCO₂ per user per day for power users. At 1 billion daily active users, that's 50–150 Mt CO₂e/yr from inference alone — equivalent to a mid-sized industrial economy.

Key Events in AI Energy Impact

YearEventDetail
2019 Strubell et al. — Training Energy Shock ACL paper reveals training one large NLP model emits ~284 tCO₂e — equivalent to 5 round-trip New York–London flights per parameter-update experiment. Triggers industry-wide energy accounting debate.
2022 ChatGPT Launch — Inference Demand Spike OpenAI ChatGPT reaches 100M users in 60 days. Inference workloads, previously negligible vs training, become the dominant ongoing energy load. Microsoft commits to >$10B in Azure OpenAI infrastructure.
2023 IEA Flags Data Centre Demand as Grid Risk IEA Electricity 2024 report warns data centre electricity demand may double by 2026. Ireland and Singapore impose temporary data centre moratoria. US grid operators (PJM, MISO) revise interconnection queues upward.
2024 $1 Trillion Hyperscaler Capex Announced Google, Microsoft, Meta, and Amazon collectively announce >$300B in 2024 data centre capex. Goldman Sachs estimates aggregate 5-year hyperscaler spend exceeds $1T. US utilities receive unprecedented grid connection requests.
2024 Nuclear Restart for AI Power — Three Mile Island Microsoft signs 20-year PPA with Constellation Energy to restart Three Mile Island Unit 1 (835 MW) to power AI data centres. Marks first US nuclear restart driven by data centre demand.
2025 EPRI Powering Intelligence Report Electric Power Research Institute projects US data centre demand grows from 176 TWh in 2023 to 325–580 TWh by 2030 — potentially 7.5% of US electricity. Notes water stress at 60% of planned US data centre sites.
2026 EU AI Act Data Centre Transparency Requirements EU AI Act (effective Aug 2026) requires providers of general-purpose AI models to report training energy consumption and carbon footprint. First binding energy disclosure requirement for AI globally.
2030+ Water Stress Constraint Emerges WRI Aqueduct projections show >60% of planned US and European data centre expansions are in high or extremely high water stress areas. Cooling water access may become a binding site constraint by 2030.
Model scope: This model uses published energy node trajectories from IEA, Goldman Sachs, and EPRI reports, with interpolation between scenario nodes. Embodied carbon uses a simplified fleet-scaling approach. The model does not capture geographic heterogeneity (US vs EU vs Asia carbon intensity differences) or the rapidly evolving chip efficiency curve.

Sources & References

SourceDescriptionKey Contribution
IEA 2024 Electricity 2024: Analysis and Forecast to 2026 Global DC demand ~415 TWh 2024; may double by 2026; baseline and central forecast
Goldman Sachs 2024 "AI is Driving a Surge in Power Demand" — Equity Research $1T hyperscaler capex; data centres 3–4% of global electricity by 2030 in bull case
EPRI 2024 Powering Intelligence: Analyzing Artificial Intelligence Technology's Potential Impact on the US Electric Grid US DC demand 325–580 TWh by 2030; 60% of new sites in water-stressed areas
Strubell et al. 2019 ACL — "Energy and Policy Considerations for Deep Learning in NLP" Training one NLP model: 284 tCO₂e; catalysed AI energy debate
Patterson et al. 2021 Google — "Carbon Intensity of AI in Practice" Location and timing of training matter 10–100× more than model architecture for carbon
Li et al. 2023 "Making AI Less 'Thirsty'" — Nature Machine Intelligence ChatGPT consumes ~500 mL water per 5–50 queries; global inference water estimates
Luccioni et al. 2023 NeurIPS — "Counting Carbon: A Survey of Factors Influencing the Emissions of Machine Learning" Inference carbon taxonomy; Scope 3 estimation methodology; 130–300 gCO₂/user/day
WRI Aqueduct 2023 World Resources Institute water stress mapping >60% of planned US/EU DC expansions in high water stress areas