In 2024, a voltage fluctuation in Northern Virginia caused the simultaneous disconnection of 60 data centres. Emergency grid interventions prevented what could have been a cascading outage affecting millions of users and billions of dollars in economic activity. The incident was contained. But it was also a preview — a glimpse of what grid fragility looks like at AI scale, and a warning that the infrastructure assumptions underlying the AI boom are more fragile than the industry's confidence suggests.
By 2026, the energy implications of artificial intelligence have moved from a technical footnote to a geopolitical priority. Data centre electricity consumption in the United States grew from approximately 76 TWh (1.9% of total usage) in 2018 to 176 TWh (4.4%) in 2023. The International Energy Agency projects that global data centre electricity demand could exceed 1,000 TWh annually before 2030 — a figure that exceeds the total electricity consumption of many mid-sized nations. Capital expenditure on data centre infrastructure reached $770 billion in 2025, outpacing investment in upstream oil and gas for the first time in history.
The energy question is no longer about cost or sustainability in the conventional sense. It has become a question of sovereignty: who controls the power that controls the intelligence, and what happens when that power is insufficient, unreliable, or geopolitically compromised?
The Infrastructure Gap
The fundamental tension at the heart of the AI energy crisis is temporal. A hyperscale data centre can be designed, permitted, and constructed in 12 to 24 months. The transmission infrastructure required to power it — new substations, upgraded transmission lines, expanded generation capacity — typically takes a decade or more to plan, permit, and build. This asymmetry has created a structural infrastructure gap that is now visible in grid reliability data across the United States, Europe, and Asia.
Data centre capital expenditure reached $770 billion in 2025 — outpacing investment in upstream oil and gas for the first time in history.
Unlike traditional electricity loads — residential consumption, industrial processes, commercial buildings — AI data centres demand continuous, high-density power. A hyperscale AI facility may require 80 MW of continuous power, compared to the 32 MW standard for conventional data centres. This load profile is fundamentally incompatible with a grid designed around variable demand patterns and the assumption that peak loads are temporary. AI infrastructure creates a new category of baseload demand that the grid was not designed to accommodate at scale.
The consequences are already visible. Grid interconnection queues in the United States now stretch up to a decade in some regions. Utilities are imposing moratoriums on new large-load connections in areas where transmission capacity is constrained. In Europe, several major data centre projects have been delayed or relocated due to grid capacity limitations. The infrastructure gap is not a future risk — it is a present constraint on AI deployment.
The National Security Dimension
Governments have been slow to recognise the national security implications of AI energy dependency, but the recognition is now accelerating. The World Economic Forum's April 2026 analysis explicitly frames AI infrastructure as critical infrastructure — subject to the same security considerations as power grids, water systems, and financial networks. The Belfer Center at Harvard Kennedy School has documented the systemic risks created by the concentration of AI compute in a small number of geographic locations, each dependent on grid infrastructure that was not designed with adversarial resilience in mind.
The cybersecurity dimension is particularly concerning. Research indicates that a small number of manufacturers hold remote access to significant portions of global generation capacity — in some cases, over 10 GW in Europe — through the inverter-based systems that manage renewable energy generation. This creates an attack surface that is both large and poorly understood. A sophisticated adversary with access to these systems could, in principle, destabilise grid operations across multiple countries simultaneously.
Data centre capital expenditure reached $770 billion in 2025 — outpacing investment in upstream oil and gas for the first time in history.
A 2024 voltage fluctuation in Northern Virginia caused the simultaneous disconnection of 60 data centres — a preview of what grid fragility looks like at AI scale.
The EU's response has been to launch the AI.grids initiative — a flagship project that integrates digital solutions across the energy value chain to enhance grid management — alongside the broader Tech Sovereignty Package. The objective is to reduce dependence on non-EU technology providers for critical grid management functions, and to develop European alternatives that can be deployed with confidence in high-security contexts. The initiative reflects a growing consensus that energy sovereignty and digital sovereignty are not separate policy domains — they are two dimensions of the same strategic challenge.
The Nuclear Pivot
The most significant strategic response to the AI energy crisis has been the pivot to nuclear power by the major hyperscalers. As of July 2026, Amazon, Microsoft, Google, and Meta have collectively committed over 9.8 GW of nuclear capacity across 13 deals. The strategic logic is compelling: nuclear power offers a capacity factor exceeding 92% — nearly four times that of solar energy — providing the continuous, high-density baseload power that AI infrastructure requires.
However, the gap between commitment and delivery is substantial. Of the 9.8 GW committed, only 1.92 GW — Amazon's front-of-meter agreement with the Susquehanna nuclear plant — is currently delivering power to AI infrastructure. The remainder is contingent on projects that will not be operational until 2027 at the earliest, with many not expected until the early 2030s.
The Small Modular Reactor Promise
Small Modular Reactors (SMRs) — typically under 300 MW, with a compact footprint of approximately 50 acres — are central to the long-term nuclear strategy for AI infrastructure. Their modular design allows for factory fabrication and on-site assembly, theoretically enabling faster deployment and lower construction risk than conventional large-scale nuclear plants. Their compact footprint makes them suitable for co-location with data centre campuses, reducing transmission losses and improving supply security.
The current pipeline of hyperscaler SMR commitments includes:
- Microsoft: A 20-year deal for 835 MW at Three Mile Island, with power expected by late 2027.
- Google: A partnership with Kairos Power for 500 MW using Generation IV molten salt technology, with initial power targeted for 2030.
- Meta: Agreements supporting up to 6.6 GW of capacity via partnerships with TerraPower and Oklo, targeting delivery between 2032 and 2035.
- Amazon: In addition to its active Susquehanna agreement, backing a 12-unit Xe-100 deployment via X-energy, targeting operations in the early 2030s.
A 2024 voltage fluctuation in Northern Virginia caused the simultaneous disconnection of 60 data centres — a preview of what grid fragility looks like at AI scale.
The economic challenges are significant. Current first-of-a-kind (FOAK) SMR costs range from $80 to $150 per MWh, compared to the $60 to $80 per MWh target required for long-term commercial competitiveness. The industry's assembly-line efficiency model — in which the cost of subsequent units falls as manufacturing processes are refined — depends on achieving sufficient deployment volume to drive learning curve effects. Whether that volume materialises before the economics become prohibitive is the central uncertainty in the SMR investment thesis.
Fuel Supply Constraints
A critical bottleneck for advanced SMR designs is the supply of High-Assay Low-Enriched Uranium (HALEU), which is required by several Generation IV reactor designs. Domestic HALEU production capacity in the United States is currently limited, and the supply chain for this fuel is not yet established at commercial scale. The Department of Energy has initiated programmes to develop domestic HALEU production, but the timeline for achieving sufficient supply to support a large-scale SMR deployment remains uncertain.
The Renewable Integration Challenge
Nuclear is not the only response to the AI energy crisis. Renewable energy — solar, wind, and increasingly offshore wind — continues to expand rapidly, and the cost trajectory for utility-scale solar and battery storage has made renewable-plus-storage combinations increasingly competitive with conventional generation. However, the variable generation profile of renewables creates a fundamental mismatch with the continuous baseload requirements of AI infrastructure.
The industry's response has been to pursue hybrid strategies: combining renewable generation with battery storage, demand response programmes, and — increasingly — direct ownership of generation assets rather than reliance on power purchase agreements. Data centre capital expenditure reached $770 billion in 2025, and a growing share of that investment is being directed toward behind-the-meter generation — solar arrays, battery systems, and in some cases small-scale gas turbines — that provide supply security independent of grid conditions.
The EU's AI.grids initiative is exploring a different approach: using AI itself to optimise grid management, enabling higher penetration of variable renewables while maintaining reliability. The initiative funds projects that apply machine learning to demand forecasting, grid balancing, and fault detection — creating a feedback loop in which AI infrastructure both drives energy demand and contributes to the grid management capabilities required to meet it.
The Sovereignty Architecture
The concept of energy sovereignty — the ability of a nation or organisation to secure reliable, affordable energy without dependence on external actors who may be adversarial or unreliable — has been transformed by the AI energy crisis. Traditional energy sovereignty was primarily about fuel supply: reducing dependence on imported fossil fuels through domestic production or diversified sourcing. The AI era adds a new dimension: compute sovereignty, in which the ability to run AI workloads without interruption becomes a strategic capability equivalent to military readiness or financial system resilience.
Hyperscalers have committed over 9.8 GW of nuclear capacity across 13 deals, yet only 1.92 GW is currently operational — the gap between ambition and delivery defines the energy sovereignty crisis.
Nations are beginning to develop what might be called "Digital and Power Audits" — systematic assessments of the alignment between their AI compute requirements and their energy infrastructure. These audits identify gaps in generation capacity, transmission infrastructure, and grid resilience that could constrain AI deployment or create vulnerabilities to adversarial disruption. The outputs inform infrastructure investment priorities and, increasingly, national security planning.
Hyperscalers have committed over 9.8 GW of nuclear capacity across 13 deals, yet only 1.92 GW is currently operational — the gap between ambition and delivery defines the energy sovereignty crisis.
The European Commission's approach — linking the AI.grids initiative to the broader Tech Sovereignty Package — reflects a recognition that energy sovereignty and digital sovereignty are inseparable. A nation that depends on non-European cloud infrastructure for its AI capabilities, and on non-European grid management technology for its energy system, has effectively outsourced two of the most critical dimensions of its strategic autonomy. The Commission's objective is to develop European alternatives in both domains — not to achieve autarky, but to ensure that European institutions retain meaningful control over critical systems.
The Grid Modernisation Imperative
The AI energy crisis has created a powerful political constituency for grid modernisation investment that did not previously exist. Utilities, technology companies, and national governments are aligned — for different reasons — on the need to expand transmission capacity, accelerate the deployment of battery storage, and modernise grid management systems. The question is whether the regulatory and permitting frameworks that govern grid infrastructure can be reformed quickly enough to enable the required investment.
In the United States, the infrastructure gap has triggered regulatory debate about cost allocation — specifically, who should pay for the grid upgrades required to accommodate large new loads. Texas Senate Bill 6 reflects a broader trend of regulatory intervention to address concerns about the fairness of passing grid upgrade costs onto smaller consumers. The resolution of these cost allocation questions will significantly influence the pace of grid modernisation and, by extension, the pace of AI infrastructure deployment.
In Europe, the permitting timelines for new transmission infrastructure remain a significant constraint. Environmental impact assessments, public consultation requirements, and cross-border coordination processes can extend permitting timelines to a decade or more — a pace that is fundamentally incompatible with the urgency of the AI energy challenge. The Commission has proposed reforms to accelerate permitting for strategic infrastructure projects, but implementation requires coordination across 27 member states with different regulatory traditions and political priorities.
The Efficiency Imperative
Alongside the supply-side responses — nuclear, renewables, grid modernisation — there is a growing focus on demand-side efficiency. AI-optimised energy routers, advanced cooling systems (including adiabatic cooling and liquid immersion cooling), and chip-level power management improvements are reducing the energy intensity of AI workloads. The industry's power usage effectiveness (PUE) metric — the ratio of total facility energy to IT equipment energy — has improved significantly over the past decade, and continued improvement is expected as cooling technology advances.
More fundamentally, the AI research community is beginning to grapple with the energy implications of model architecture choices. Large language models trained on massive datasets with billions of parameters are energy-intensive by design. Research into more efficient architectures — sparse models, mixture-of-experts approaches, neuromorphic computing — is motivated in part by the recognition that the current trajectory of energy consumption is not sustainable at the scale required for widespread AI deployment.
The Strategic Outlook
The AI energy crisis is not a temporary bottleneck that will resolve itself as grid infrastructure catches up with demand. It is a structural challenge that reflects a fundamental mismatch between the pace of AI deployment and the pace of energy infrastructure development. Resolving it requires coordinated action across multiple dimensions: nuclear deployment, renewable integration, grid modernisation, permitting reform, and efficiency improvement.
The nations and organisations that navigate this challenge successfully will be those that treat energy sovereignty as a strategic priority equivalent to semiconductor independence or financial system resilience — and that invest accordingly. The nations and organisations that treat it as a technical problem to be solved by the market will find themselves dependent on energy infrastructure they do not control, for AI capabilities they cannot afford to lose.
The power paradox — that the technology designed to optimise everything is itself creating an optimisation problem of historic proportions — is not irresolvable. But resolving it requires a clarity of strategic intent that has, so far, been more evident in corporate boardrooms than in government ministries. That balance is beginning to shift. Whether it shifts quickly enough is the defining energy question of the AI era.
