DailySandIn March 2026, Microsoft disclosed in an SEC filing that it had signed power purchase agreements totaling 67 terawatt-hours annually — more electricity than the entire country of Portugal consumes in a year. The filing was buried in a footnote. Almost no one noticed.
This is the energy reality of the AI era. The large language models powering ChatGPT, Claude, Gemini, and their successors are not just software. They are physical infrastructure, and that infrastructure runs on electricity at a scale that was, until recently, the exclusive domain of industrial manufacturing nations.
The power demand problem is now the central constraint on AI deployment — and its ramifications extend into commodity markets, geopolitics, and supply chains that most AI observers have never considered.
The Numbers Behind AI Power Demand
A single training run for a large frontier model consumes roughly 50 to 100 gigawatt-hours (GWh) of electricity — equivalent to the annual electricity consumption of approximately 5,000 to 10,000 average U.S. households. GPT-4's training run was estimated at 50 GWh. The training runs for subsequent frontier models have been substantially larger.
But training, despite its dramatic energy figures, is not the dominant power consumer. Inference — serving model responses to users — is. A single ChatGPT query consumes approximately 10 watt-hours, compared to roughly 0.3 watt-hours for a Google search. With billions of queries per day across deployed AI systems, the continuous power draw for inference dwarfs the episodic cost of training.
The International Energy Agency's 2025 AI Energy Report projected that data centers globally would consume between 650 and 1,000 terawatt-hours of electricity annually by 2030 — up from approximately 400 TWh in 2024. The high-end scenario represents a quantity of electricity roughly equal to Japan's total annual electricity consumption.
Where the Power Comes From — and Why That Creates Problems
AI data centers require firm power — electricity that is available 24 hours a day, 7 days a week, regardless of weather conditions. Solar and wind power, while inexpensive when generating, are intermittent. A solar farm in Arizona produces zero power at night; a wind farm in Iowa produces zero power when the air is calm.
This creates a structural problem for AI operators who have made commitments to operate on 100% renewable energy. The practical resolution, for now, is a combination of:
- Renewable energy purchase agreements for a portion of grid power
- Battery storage (primarily lithium-ion at the facility level, primarily grid-scale LFP for longer duration)
- Natural gas backup for when renewables are insufficient
- Nuclear power purchase agreements — a growing trend
The nuclear pivot deserves attention. Microsoft signed a 20-year power purchase agreement with Constellation Energy in 2023 to restart Unit 1 of Three Mile Island — the reactor that produced no power for 20 years after the 1979 accident. The deal was motivated explicitly by the need for firm, carbon-free power for AI data centers.
Google, Amazon, and Meta have all signed contracts with nuclear energy developers for small modular reactor (SMR) output scheduled for the late 2020s and 2030s. These commitments are extraordinary: they represent the tech industry underwriting the next generation of U.S. nuclear power development, driven directly by AI power requirements.
The Grid Connection Challenge
The single biggest bottleneck for AI data center expansion in the United States and Europe is not power generation — it is grid connection.
Connecting a large data center to the transmission grid requires a new substation, new high-voltage transmission lines, and approval from multiple regulatory bodies. In the United States, the average wait time for a grid interconnection approval — the utility's formal commitment to connect a new load — has grown from 18 months in 2018 to approximately 5 years in 2025. The interconnection queue for the PJM grid (covering the Mid-Atlantic and Midwest) contained over 3,000 projects totaling more than 1,600 gigawatts of requested capacity as of early 2026.
This backlog means that even a fully funded, permitted, and constructed AI data center may wait 3–5 years to receive grid power at the desired capacity. Hyperscalers have responded by:
- Co-locating with power plants: Microsoft's Three Mile Island agreement is partly motivated by the ability to take power directly from the plant rather than through the grid interconnection queue
- Building on-site generation: Natural gas turbines and fuel cells placed directly at the data center facility, avoiding transmission grid entirely
- Acquiring retired industrial sites: Former steel mills and chemical plants that already have high-voltage grid connections built for industrial loads
The Critical Minerals Dimension
The power infrastructure required to supply AI data centers at scale creates its own mineral demand cascade.
Copper is the largest volume mineral requirement. The transformers that step down transmission voltage to data center operating voltage, the busbars that distribute power within the facility, and the cabling throughout the building all require large quantities of copper. A 500MW data center campus — roughly the scale of Microsoft's Iowa facilities — requires an estimated 15–25 million pounds of copper in its power infrastructure alone, before accounting for servers and networking equipment.
Lithium, cobalt, and nickel go into the UPS battery systems that protect computing equipment from grid instability. As power draw per facility increases, UPS battery capacity grows proportionally. The transition toward LFP (lithium iron phosphate) chemistry in large stationary storage systems reduces cobalt demand per unit, but total cobalt demand for data center UPS systems continues to grow with the overall scale of deployment.
Rare earth magnets appear throughout the power distribution system — in transformer core materials (certain ferrite and amorphous metal transformer cores use rare earth doping), in the motors driving cooling systems, and in any power generation equipment co-located at the facility (wind turbines and certain gas turbine generators both use NdFeB permanent magnets in their generators).
Silicon carbide is the enabling material for the next generation of power conversion equipment at data centers. SiC inverters and converters operate at higher efficiency than silicon-based alternatives at the voltages and currents involved in large-scale power distribution. Wolfspeed, the leading SiC wafer manufacturer, cited AI data center demand as a primary growth driver in its 2025 investor presentations.
The Investment Implications
The power infrastructure buildout for AI creates investable exposure across multiple parts of the supply chain, distinct from the semiconductor layer that receives the majority of market attention.
Electrical equipment manufacturers — transformer producers (ABB, Eaton, Siemens Energy), switchgear manufacturers, and power cable producers — have multi-year backlogs as data center and grid infrastructure spending accelerates. ABB's electrification segment reported order growth exceeding 15% in 2025, explicitly citing AI data center demand.
Nuclear energy — the restart of existing plants (Constellation Energy) and development of new capacity (NuScale, TerraPower, X-Energy, Kairos) has been directly catalyzed by AI power demand. Uranium demand implications are a downstream consequence.
Grid infrastructure — companies building transmission infrastructure, substations, and grid interconnection equipment (Quanta Services, MYR Group, Aecom) benefit from both AI-driven demand growth and the broader grid modernization trend.
Copper miners — as discussed, the copper demand case is among the most concrete and well-underwritten of the AI infrastructure mineral stories.
The common thread across all of these is that they are infrastructure-layer bets — they benefit from AI buildout regardless of which model or company wins the AI race. They are also slower-moving and more capital-intensive than software plays, which is why they trade at lower multiples and receive less coverage from technology analysts.
The Timeline That Matters
The power constraint is not going to be resolved quickly. Grid interconnection timelines of 3–5 years, nuclear plant restart timelines of 2–4 years, and mineral extraction timelines of 7–10 years for new capacity all operate on longer horizons than the quarterly earnings cycles that dominate AI stock coverage.
The implication for AI development timelines is that model scaling will eventually be constrained not by a lack of ideas or capital but by a lack of electrons. The companies that locked in power purchase agreements, grid connections, and co-location deals in 2023 and 2024 have a structural advantage over those attempting to secure the same resources in 2026 and beyond — because the queue is now years long.
This is the hidden infrastructure race beneath the AI benchmark race. And it is made of copper, cobalt, rare earths, and silicon carbide.