The Data Centre Energy Crisis: How AI's Power Hunger Is Reshaping Global Energy Markets
In 2020, data centres consumed approximately 200–250 terawatt-hours of electricity annually — roughly 1% of global electricity demand — and this figure had been roughly stable for a decade despite exponential growth in data traffic, because efficiency improvements in servers, cooling, and power delivery kept pace with workload growth. That era of efficiency-driven demand stability is over. The deployment of large language model inference at scale, the buildout of GPU clusters for AI training, and the proliferation of AI-embedded applications across enterprise software are driving data centre electricity demand growth that efficiency improvements cannot offset. Goldman Sachs estimates that data centre power demand will increase 160% by 2030 relative to 2023 levels. The International Energy Agency projects data centre electricity consumption reaching 945 TWh by 2026 — more than Japan's total electricity consumption. The energy implications of this demand surge are reshaping power markets, energy policy, and corporate strategy in ways that extend far beyond the technology sector.
Why AI Training and Inference Are So Energy-Intensive
The energy intensity of AI workloads relative to traditional computing reflects the architecture of neural network processing. Training a large language model involves running billions of floating-point operations per second across thousands of GPU or TPU chips simultaneously, each drawing 300–700 watts, for weeks or months continuously. GPT-4's training run is estimated to have consumed approximately 50 GWh of electricity — equivalent to the annual electricity consumption of roughly 4,500 US households — and next-generation model training runs are estimated to require multiples of this. Inference — serving responses to the millions of queries that deployed AI models receive — is less energy-intensive per query but operates at far greater scale, with estimates suggesting that ChatGPT responses consume approximately 10x the energy of a standard Google search. As AI inference becomes embedded in search, productivity software, customer service, and industrial automation, the aggregate inference energy demand will exceed training demand, which is already outsized by historical computing standards.
The hardware that processes AI workloads amplifies the energy intensity. Nvidia's H100 GPU — the primary chip for AI training and inference — draws 700 watts under full load. A single H100 server rack containing eight GPUs draws approximately 5.6–10 kilowatts. A hyperscale AI training cluster of 10,000 GPUs — the scale that OpenAI, Google, and Anthropic are operating — draws 50–100 megawatts continuously. The cooling infrastructure required to manage the thermal output of these dense compute clusters — liquid cooling, direct liquid cooling, immersion cooling — consumes additional energy and represents a significant engineering challenge that is driving data centre design innovation at pace.
The Grid Constraint That Is Creating a Power Crisis
The scale and speed of data centre demand growth is creating grid interconnection and capacity constraints in the geographies where hyperscale operators want to build. Virginia's data centre corridor — which hosts approximately 35% of the world's data centre capacity in a cluster of counties around Ashburn — is experiencing power constraints so acute that Dominion Energy, the regional utility, has warned that it cannot accommodate new large loads on current timelines without transmission infrastructure investments that will take 5–10 years. Amazon, Microsoft, Google, and Meta are collectively seeking to connect gigawatts of new data centre capacity to grids that were not designed for this demand concentration, in timelines measured in months rather than the years that grid upgrade projects require. The same constraint is emerging in Ireland, where data centres already account for approximately 21% of national electricity consumption and the grid operator EirGrid has imposed connection moratoria in some areas. Northern Virginia, Phoenix, Dallas, and Silicon Valley are all experiencing versions of the same problem: demand growth outpacing grid expansion at a rate that is straining both physical infrastructure and regulatory processes.
The Energy Procurement Strategy Reshaping Power Markets
The hyperscale response to power constraints has been to vertically integrate into energy procurement and generation at a scale that no industrial company has previously attempted. Microsoft's 20-year power purchase agreement with Constellation Energy for Three Mile Island nuclear power — announced in September 2024 and representing the first nuclear plant restart driven by corporate power demand — is the clearest signal that data centre operators are willing to make generation-level commitments to secure long-term power supply. Google's agreements with Kairos Power for 500 MW of high-temperature gas-cooled reactor capacity, Amazon's nuclear agreements with Energy Northwest and Dominion Energy, and Meta's request for proposal for nuclear power supply collectively represent the largest corporate nuclear energy procurement drive since the 1970s. These commitments are not driven by environmental policy — they are driven by load matching: nuclear power's 24/7 output profile matches the continuous operation of data centres far better than solar and wind generation, whose intermittency creates grid balancing challenges at scale.
The capital flowing into geothermal power — which also provides 24/7 baseload generation from a carbon-free source — reflects the same demand matching logic. Fervo Energy's partnership with Google, the commissioning of the first enhanced geothermal system projects in Nevada, and the surge of geothermal project development in areas with favourable geology are directly driven by data centre operators' search for clean, always-on power. The power purchase agreements that hyperscalers are signing are also reshaping the economics of independent power projects: a 20-year PPA with Microsoft or Google provides debt financing certainty that makes projects viable that would otherwise struggle to attract capital in merchant power markets.
The Efficiency Response That Is Buying Time
The energy industry's demand projections assume continued efficiency improvement in AI hardware and data centre infrastructure, which provides some moderating force on the most extreme demand scenarios. Nvidia's Blackwell architecture achieves approximately 4x better energy efficiency per token of inference output than the H100 it replaces. Custom AI chips from Google (TPU v5), Amazon (Trainium 2), and Microsoft (Maia) are optimised for efficiency in specific workloads and operate at better energy-per-inference ratios than general-purpose GPUs. Power usage effectiveness — the ratio of total facility energy to IT equipment energy — has improved from industry averages of 1.5–2.0 in 2015 to below 1.2 at leading hyperscale facilities, with some liquid-cooled high-density clusters approaching 1.05. Algorithmic efficiency improvements — model distillation, quantisation, and speculative decoding — are reducing the compute required per inference query. These efficiency gains are real, but they are not preventing absolute energy demand growth: the Jevons paradox effect — where efficiency improvements lower the cost of AI inference and therefore drive greater consumption — is outpacing the efficiency gains in aggregate demand terms.
The Investment Opportunity Hidden in the Crisis
The data centre energy crisis is creating investment opportunities across the energy value chain that extend well beyond data centre operators themselves. Grid infrastructure companies — Quanta Services, MYR Group, Aecom — are facing multi-decade demand for the transmission and distribution upgrades that data centre load growth requires. Transformer manufacturers — whose lead times extended to 2–3 years in 2024 due to demand exceeding production capacity — are the most acute single bottleneck in grid expansion and represent a constrained supplier market with pricing power that has not existed in decades. Power electronics companies producing grid-scale inverters, switchgear, and power conditioning equipment are positioned at the intersection of data centre demand growth and renewable energy integration. The energy storage companies providing grid-scale battery systems to manage the load variation and renewable intermittency that data centre demand creates are building recurring revenue streams from the same infrastructure investment cycle. Understanding the data centre energy crisis as an energy investment thesis — not just a technology investment thesis — is the reframe that positions investors in the companies capturing the most durable value from AI's physical energy requirements.