Artificial intelligence (AI) is transforming global computing infrastructure, but its rapid expansion is colliding with a critical physical constraint: energy demand and electricity grid capacity. In 2024, data centres worldwide consumed approximately 415 terawatt‑hours (TWh) of electricity, representing about 1.5% of global electricity consumption. Global data centre electricity demand is projected to more than double to around 945 TWh by 2030, driven by AI‑intensive workloads, machine learning applications, and the rise of cloud‑based digital services.

The surge in electricity demand is concentrated in AI-optimised servers, which are growing faster than conventional data-centre hardware. According to Gartner, AI‑optimised servers in 2025 already represent a meaningful share of data centre power use and are expected to account for 44% of total data centre power consumption by 2030, with their electricity usage rising sharply over the decade. This trend is a key driver of AI infrastructure energy costs, data centre operational risks, and electricity grid stress.

The impact of rising energy demand is already visible at regional and grid levels. In major data centre clusters in the United States and Europe, utilities and grid operators are adapting to surging demand, incorporating AI workload patterns into power planning and collaborating with technology firms to manage electricity supply constraints. The accelerating load from data centres is prompting energy planners to rethink grid capacity and reliability to support future digital growth.

Where AI growth intersects with constrained energy availability, compute‑intensive AI projects risk shifting from competitive advantage to operational constraint. When electricity grids cannot keep up with AI demand, high‑performance AI workloads may face throttling, increased costs, and reduced return on investment. This is the point at which AI expansion can destroy value before it creates it.

AI energy constraint landscape: regional, operational, and regulatory exposure

AI’s rapid expansion is reshaping patterns of electricity consumption, grid planning, and policy attention. Unlike traditional IT workloads, AI-optimised infrastructure concentrates massive power demand into specific regions and accelerated growth trajectories. These structural realities are already influencing where and how AI infrastructure can be deployed.

image host

The hidden costs of ignoring energy volatility in AI infrastructure

Non‑linear growth in electricity demand and cost volatility as an execution risk

AI workloads are not scaling linearly. Instead, they trigger step-function increases in power demand, often outpacing traditional grid planning models. According to the World Economic Forum, AI training and inference cycles can spike electricity usage unpredictably, especially during model retraining or deployment surges. This mismatch between compute growth and grid readiness leads to under-provisioned capacity, forcing operators into reactive procurement strategies that compromise uptime and reliability.

Compounding this challenge is electricity price volatility that has added a new layer of financial uncertainty to AI infrastructure economics, turning energy costs from a relatively stable IT budget assumption into a first-order driver of return-on-investment calculations.

In regions with constrained grid capacity, such as the PJM Interconnection in the United States, wholesale electricity and capacity auction prices have surged to record levels as power demand from large data centres has grown faster than supply and grid upgrades can keep pace. In the latest PJM capacity auction for the 2026/2027 delivery year, prices cleared at the federal price cap (approximately $329 per megawatt-day), reflecting intensely tight conditions and a significant upward shift compared with prior cycles. These elevated clearing prices were driven in large part by electricity demand forecasts that include massive data centre loads, which have been identified as the primary reason for recent high capacity costs and accounted for a large share of total capacity auction expenditures.

The knock-on effects are already visible in the marketplace. Because capacity costs are ultimately passed through to consumers via utility rates, residential and business electricity bills have started to reflect these underlying market dynamics, with some customers seeing noticeable increases in recent years as utilities recover higher supply and infrastructure costs. In such volatile markets, operating costs for data centres can fluctuate unpredictably, squeezing margins and compressing expected returns.

Capital intensity of grid expansion and infrastructure investment

Meeting AI‑driven energy demand is not simply operational; it requires substantial capital investment in the electricity system itself. Transmission upgrades, new generation capacity, substation enhancements, and long‑lead interconnection facilities must keep pace with rapid compute deployment. Goldman Sachs research estimates that approximately $720 billion in grid investment through 2030 may be required to support this demand, with permitting and construction timelines often stretching multiple years. When AI infrastructure rollouts outpace grid readiness, organisations face project delays, increased financing costs, and the strategic risk of launching compute capacity that cannot be fully utilised due to energy constraints.

Rising operational complexity and cost intensification

AI workloads also drive significantly higher internal facility demand than traditional computing, fundamentally reshaping data centre infrastructure. Average rack power density has more than doubled in just the past two years — rising from around 8 kW to about 17 kW per rack and is projected to climb toward 30 kW per rack by 2027 as AI workloads scale. In contrast to legacy servers, modern AI clusters with high-performance GPUs and accelerators can push individual rack loads far beyond these averages — training large models such as ChatGPT can exceed 80 kW per rack, and next-generation systems like Nvidia’s GB200 scale towards 120 kW and higher in dense configurations.

These soaring power densities necessitate advanced cooling solutions (including liquid and direct-to-chip cooling), enhanced electrical distribution and redundancy, and specialised facility designs to handle concentrated heat and electrical loads efficiently. As a result, operating expenses and capital expenditures rise sharply — not only from higher energy consumption but also from investments in thermal management, backup systems, larger switchgear and power feeds, and purpose-built infrastructure. Even in regions with relatively low electricity prices, these structural and systems upgrades drive up the total cost of ownership for AI-optimised data centres.

Energy availability as an execution constraint

Grid interconnection and capacity availability have emerged as critical execution constraints for AI infrastructure, posing material risks to project timelines and strategic plans. In many key markets, utilities and grid operators are struggling to accommodate surging data centre load growth, with interconnection queue times ballooning to four–ten years in regions such as Northern Virginia and averaging roughly eight years across major U.S. grids, compared with typical timelines of less than two years in the early 2000s.

Transmission system upgrades and high-voltage capacity expansion – prerequisites for powering large AI-optimised facilities – can take five to ten years or more due to permitting, engineering, and construction processes that lag far behind data centre deployment cycles. As a result, projects that reach commercial operation often require roughly five years from interconnection request to completion, more than double the historical norm, while AI compute infrastructure is frequently planned on much shorter horizons.

This mismatch between “time to compute” and “time to power” forces organisations to delay rollouts, scale back planned deployments, or relocate projects to alternative locations with more available capacity — all of which undermine strategic timelines, inflate costs, and weaken competitive positioning in the race to deploy AI at scale.

Regulatory and policy execution risk

As AI-driven electricity demand becomes more prominent on power systems, energy regulators and policymakers are actively rethinking regulatory frameworks, cost allocation mechanisms, and reliability standards to address the implications of these large loads. In several jurisdictions, regulators and utilities are revising tariff designs and creating new rate classes to ensure that infrastructure costs associated with serving high-demand customers are borne by those customers rather than socialised across all ratepayers. This includes proposals for separate rate categories with long-term contracts, minimum demand charges, and upfront grid upgrade contributions for large data centres to prevent cost shifts to households and small businesses.

Legislatures and grid operators are also embedding new compliance requirements into interconnection agreements such as flexibility obligations, demand response capabilities, and the ability to curtail load during grid stress events to safeguard reliability as AI workloads strain existing infrastructure. Meanwhile, debates continue over how best to allocate regional transmission and distribution upgrade costs, with policymakers weighing fairness, long-term infrastructure planning, and the potential for stranded assets if capacity commitments are not realised.

As electricity market structures evolve, data center operators may face new tariffs, stricter permitting conditions, tariff redesigns, and compliance obligations that increase both direct costs and deployment friction, further compressing the economic value of AI investments and elevating regulatory risk as a key execution factor.

Strategic considerations for managing AI energy volatility

Practical strategies for mitigating electricity price risk, securing supply reliability, and de-risking AI infrastructure investments.

Diversified energy procurement and long-term hedging

  • Power purchase agreements as price stabilisation tools

Long-term power purchase agreements (PPAs) have become essential tools for managing electricity cost volatility and securing dedicated capacity for AI infrastructure. In 2024, data centre operators led the technology sector’s clean energy procurement, contracting more than 17 GW of new renewable energy capacity through long-term PPAs globally — a record share of corporate deals. These agreements provide price stability, hedge against wholesale market fluctuations, and support corporate sustainability goals.

PPAs generally take two forms: physical PPAs, which deliver power to the grid region where data centres operate, and virtual PPAs (VPPAs), which provide financial price hedges linked to renewable energy generation and allow geographic flexibility.

In addition to traditional PPAs, leading technology firms are exploring more sophisticated risk management approaches. Some companies are seeking authorisation from the Federal Energy Regulatory Commission (FERC) to participate directly in wholesale electricity markets, enabling enhanced hedging capabilities more typical of large industrial energy consumers.

  • Nuclear energy for baseload reliability

Nuclear power is increasingly part of hyperscale data centre energy strategies, offering reliable, carbon-free baseload electricity that complements intermittent renewables and mitigates energy price and supply volatility. Leading technology companies are securing long-term nuclear power agreements and supporting advanced reactor development to meet the large and continuous electricity demands of AI infrastructure.

  1. Microsoft-Constellation Energy Three Mile Island Restart: Microsoft entered a 20-year agreement with Constellation Energy to purchase electricity from the restarted Three Mile Island Unit 1 nuclear reactor in Pennsylvania, expected online in the late 2020s. This arrangement illustrates how reactor restarts can provide dependable baseload power for data centre operations.

  2. Google-Kairos power small modular reactors: Google’s partnership with Kairos Power and the Tennessee Valley Authority includes a first SMR delivering ~50 MW of clean energy via TVA’s grid, with broader plans to procure up to ~500 MW from advanced reactors by 2035, supporting data centres in Tennessee and Alabama.

  3. Meta Multi-Partner nuclear strategy: Meta has signed long-term agreements to buy power from Vistra’s existing nuclear plants and is supporting development of SMRs with Oklo and TerraPower. These collective commitments could supply several gigawatts of nuclear capacity through 2035, enhancing energy reliability for AI workloads.

  4. Other corporate nuclear engagements: AWS and other cloud providers are advancing investments and agreements with SMR developers and utilities that aim to expand nuclear capacity in support of future AI data centres, though many of these projects target the early 2030s for commercial operation.

Behind-the-meter generation and accelerated time-to-power

Extended grid interconnection timelines—often spanning multiple years due to queue backlogs, regulatory complexity, and network constraints—are driving data centre operators to deploy behind-the-meter (BTM) generation to achieve faster access to reliable power. Operators are turning to local generation technologies to bypass interconnection delays and accelerate time-to-power, especially where AI workloads demand near-continuous electricity and utility delivery timelines lag project build schedules. According to industry data, a significant share of future data centres expect to rely on on-site power, with some forecasts projecting that as many as 27% of facilities will be fully powered by on-site generation by 2030, up sharply from a small share a year earlier.

  • Natural gas turbines are among the most widely adopted BTM technologies because of their relatively fast deployment and dispatchable output, serving as a bridge to primary power when grid connections are delayed. Industry analysts note that modular gas turbine platforms are being contracted at multi-gigawatt scale to serve major data centre campuses.

  • Fuel cell systems represent another emerging solution, offering on-site baseload generation with lower emissions and faster build timelines than some traditional power plants. Research indicates fuel cells could help meet a meaningful portion of incremental data centre power demand through 2030.

  • Hybrid configurations and microgrids—which integrate generation, storage, and control systems—are also expanding, offering resilience and grid-independent operation. Microgrid capacity in the U.S. is growing quickly, driven by demand for reliable local power from tech companies and utilities alike.

Together, these BTM solutions allow operators to reduce reliance on overburdened utility grids, accelerate deployment, and enhance operational resilience, even as they require careful commodity price risk management, emissions planning, and regulatory navigation.

Grid flexibility and demand response programmes

As data centre load growth pushes existing power systems toward capacity limits, demand flexibility has emerged as a strategic enabler for accelerating interconnection timelines, reducing infrastructure costs, and supporting grid reliability during stress events. Even modest adjustments to peak consumption—such as curtailing load for a fraction of annual uptime—could unlock significant hosting capacity for new loads, potentially enabling up to 76 GW of additional demand without requiring major grid upgrades.

  • Across key markets, regulators and grid operators are embedding flexibility into interconnection and reliability processes. In Texas, Senate Bill 6 requires large load customers to demonstrate remote curtailment capabilities or alternative energy arrangements during grid emergencies, and ERCOT now has the authority to curtail non‑critical loads to prevent broader outages.

  • In the PJM region, interconnection reforms and accelerated stakeholder mechanisms are being pursued to manage growing data centre load forecasts—estimated at tens of gigawatts through 2030—while maintaining system adequacy and reliability.

  • Federal actions are also underway to revise interconnection procedures, including proposals to shorten study timelines when flexibility and co‑located generation commitments are offered by large customers.

  • Industry and research collaborations, such as EPRI’s DCFlex initiative, are exploring how data centres can integrate with grid services and provide flexibility alongside traditional grid resources.

These developments reflect a shift toward integrating data centre demand response and flexibility capabilities as core components of grid planning and reliability frameworks, enabling faster market access for AI infrastructure while reducing the need for expensive transmission or generation buildouts.

Advanced cooling technologies and energy efficiency

AI‑driven workloads and high‑performance computing demand significantly higher rack power densities than traditional data centre environments, with many hyperscale systems operating well beyond the 30–50 kW per rack range where air cooling becomes inadequate. Liquid cooling technologies—such as direct‑to‑chip cold plates and immersion systems—are emerging as mainstream solutions capable of supporting high heat fluxes and improving overall energy performance.

  • Liquid cooling improves thermal management through more efficient heat transfer and reduced auxiliary energy use (e.g., fans and chillers), contributing to lower Power Usage Effectiveness (PUE) and reduced operational costs compared to air cooling. Some installations report substantial energy and space efficiencies in high‑density environments.

  • The liquid cooling market is expanding rapidly. Global forecasts project the data centre liquid cooling industry will grow significantly over the next decade, driven by the rise in AI workloads, cloud computing, and edge applications. Immersion cooling—where components are submerged in non‑conductive fluids—also shows strong growth potential and is increasingly adopted across regions and deployment types.

  • Direct liquid cooling and immersion systems are being developed and deployed by major data centre operators and technology vendors to support higher rack densities, reduce energy consumption, and enhance sustainability.

Although liquid cooling typically carries higher upfront costs and greater integration complexity than air cooling, the operational savings, increased compute density support, and PUE improvements often justify the investment for AI-intensive data centres. Hybrid and modular solutions also facilitate adoption in legacy facilities.

Workload optimisation and algorithmic efficiency

Strategic workload optimisation and hardware‑software co‑optimisation are critical levers for reducing energy intensity in AI infrastructure without compromising performance. By aligning compute operations with energy availability, improving model efficiency, and maximising hardware utilisation, operators can significantly reduce both energy costs and carbon emissions.

  • Demand‑aware and carbon‑adaptive placement: Platforms that schedule workloads based on grid carbon intensity or energy availability can reduce emissions and energy costs. Google’s carbon‑aware computing approach dynamically shifts non‑urgent compute to periods and regions with low carbon intensity, achieving significant reductions in carbon output when workloads are flexibly placed.

  • Model efficiency: Although specific company‑level model performance claims vary, the broader industry trend favours models optimised for energy efficiency — including reduced precision formats, smaller specialised models, and inference‑optimised architectures — which reduce the energy required per computation relative to larger, generic models.

  • Hardware power trends: Modern servers equipped with AI accelerators draw more power than older systems, with typical dual‑socket servers consuming 600–750 W in operation, reflecting the energy intensity of AI workloads. At the same time, improvements in idle power proportionality and power management have enhanced overall efficiency compared with earlier generations.

  • Infrastructure and virtualisation optimisation: Predictive autoscaling, energy‑aware VM allocation, and workload consolidation frameworks can improve server utilisation and reduce unnecessary energy consumption by aligning resource provisioning with real demand patterns. Academic research demonstrates that intelligent workload forecasting and dynamic resource allocation can reduce energy use while maintaining performance.

By combining strategic workload placement, model optimisation, hardware efficiency improvements, and smarter resource management, data centre operators can significantly lower energy intensity and improve performance‑per‑watt across AI infrastructure.

Geographic diversification and site selection strategy

Strategic data centre siting now hinges critically on energy availability, grid capacity, and interconnection timelines alongside traditional considerations such as land cost and network connectivity. Operators increasingly prioritise markets that can provide reliable and cost‑effective electricity, shorter interconnection queue times, temperate climates that reduce cooling burdens, and supportive regulatory environments.

  • Grid constraints in established hubs — particularly Northern Virginia, where interconnection wait times can reach five to seven years or more — are prompting developers to explore alternative regions with more scalable power access. While Northern Virginia remains critical due to its connectivity and ecosystem, power limitations and low vacancy rates are significant challenges.

  • International diversification is reshaping expansion strategies. Nordic countries attract investment with abundant renewable energy (especially hydropower) and cool climates that improve energy efficiency, while emerging markets in Asia‑Pacific and the Middle East offer new capacity amid rapid AI and cloud adoption.

  • Ireland’s data centre sector has faced notable grid pressure and planning moratoriums due to electricity shortage concerns, although recent policy revisions aim to balance new capacity with renewable generation and storage requirements.

Geographic diversification enhances resilience against localised grid stress, natural disasters, and regional policy shifts but also raises complexities in multi-jurisdictional regulation and may increase network latency for some latency-sensitive applications.

Proactive regulatory and policy engagement

As AI‑driven electricity demand reshapes power systems, engaging proactively with regulators, utilities, and policymakers has become critical for managing execution risk and influencing favourable policy outcomes.

  • Tariff design and cost allocation: Regulatory proceedings increasingly address how costs associated with large data centre interconnections and grid upgrades should be allocated. Analyses indicate that legacy practices in some regions socialise billions in high‑voltage grid costs across all ratepayers, prompting calls for reforms that ensure costs follow the users driving demand, including bespoke contracts, minimum‑take provisions, and contribution‑in‑aid‑of‑construction mechanisms.

  • Interconnection reform: Federal and regional reforms are underway to streamline interconnection timelines and clarify obligations for large loads. At the federal level, the U.S. DOE has directed FERC to initiate rulemaking to expedite interconnection for large loads and standardise procedures often taking years, with proposed reforms targeting accelerated 60‑day study timelines and joint load‑generation interconnection requests. Concurrently, FERC has ordered PJM to revise its tariff to provide clearer rules for co‑located loads, and regional operators such as the Southwest Power Pool are implementing fast‑track interconnection processes for high‑impact loads.

  • Flexibility and reliability standards: Emerging regulatory frameworks are incorporating flexibility requirements for large loads and reliability considerations. For example, Texas’s Senate Bill 6 mandates curtailment or backup generation capabilities for large customers during grid stress, while NERC is evaluating reliability standards and technical protocols to address the reliability impact of large and hybrid loads on the bulk power system.

  • Clean energy and climate policy: Active participation in discussions around clean energy procurement policies, renewable portfolio standards, and grid modernisation frameworks allows operators to align sustainability commitments with regulatory incentives. This engagement spans debates about transmission planning, cost allocation reforms, and integration of large, low‑carbon generation sources that support continuous, reliable power for AI infrastructure.

By engaging early and collaboratively in these regulatory proceedings, operators can reduce risk, shape equitable policy outcomes, and demonstrate their commitment to responsible infrastructure development in rapidly evolving electricity markets.

Conclusion

Effective energy risk management for AI infrastructure requires strategic foresight, flexibility, and coordination across multiple dimensions. By proactively combining technological innovation, operational efficiency, and regulatory engagement, organisations can turn energy volatility from a constraint into a managed risk. Without this integration, AI expansion risks destroying value through delayed deployment, inflated power costs, stranded capital, and regulatory friction long before productivity gains materialise.

An integrated energy strategy enables reliable, resilient, and cost-effective AI operations, positioning companies to navigate tightening power markets and evolving policy environments while preserving long-term economic returns. In the AI era, energy is not merely an input cost—it is a gating factor that determines whether growth compounds value or erodes it before it can be realised.

References

International Energy Agency (IEA). (2025). Energy and AI: Analysis.
Gartner. (2025). Gartner Says Electricity Demand for Data Centers to Grow 16 Percent in 2025 and Double by 2030. [Press Release].
Gartner. (2024). Gartner Predicts Power Shortages Will Restrict 40% of AI Data Centers by 2027. [Press Release].
Goldman Sachs Research. (2025). Accelerating Power Demand from Data Centers Is Poised to Boost New Energy Technologies.
PJM Interconnection. (2025). PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal.
Constellation Energy. (2024). Constellation to Launch Crane Clean Energy Center Restoring Jobs and Carbon-Free Power to The Grid. [Press Release].
Google. (2024). New nuclear clean energy agreement with Kairos Power.
Meta. (2026). Meta Announces Nuclear Energy Projects, Unlocking Up to 6.6 GW of New Capacity. [Press Release].