AI Exchanges: Where Will the Power Come From?

Core Constraint and Infrastructure Reality

• AI infrastructure is defined as one of the most physical technology infrastructures ever built, requiring vast data centers filled with racks of servers and GPUs that consume power at an unprecedented scale and density.
• Power has emerged as the single most binding constraint on the delivery of AI promises, creating a critical bottleneck that sits at the nexus of hyperscalers and highly regulated utilities.
• Unlike the ephemeral perception of AI software, the hardware requirement involves a grid that is, on average, over 40 years old and requires massive new capacity.

Historical Context and Current Demand Shock

• U.S. power demand remained flat for the last 20 years despite economic and population growth, a trend driven by offshoring manufacturing and increased appliance efficiency.
• Energy policy over the last two decades prioritized decarbonization, leading to a grid increasingly reliant on intermittent sources (wind and solar) while retiring reliable 24/7 baseload plants.
• The current AI demand represents a "step change" rather than a gradual increase, catching the industry off guard as there is no existing playbook for this volume of growth.
• Electric vehicle (EV) charging was previously viewed as the potential breaker of the grid, but AI demand has since superseded it as the primary stressor.

The Supply Chain and Timing Mismatch

• There is a severe timing mismatch between the technology and utility sectors: data centers can be built in 1–2 years, whereas new power generation assets take 5–10+ years to permit and construct.
• Supply chain bottlenecks are acute; specifically, gas combustion turbines are sold out until 2030, creating a hard cap on near-term generation capacity.
• The current administration has enacted a national energy emergency declaration and issued executive orders to streamline permitting and siting for new power projects.
• Craft labor shortages are a critical constraint, with the construction of a single new nuclear plant historically requiring upwards of 10,000 workers on site.

Strategic Partnerships and Financial Dynamics

• Hyperscalers and utilities are forming unprecedented partnerships to bridge the gap, with hyperscalers utilizing their vast balance sheets to support utility capital requirements.
• Projects are now valued in the tens of billions of dollars, with power constituting roughly 10% of the total data center build-out cost.
• Utilities face an existential scaling challenge, being asked to double their current operating capacity within a decade, a pace that would have previously taken them 100 years to achieve.
• M&A activity has surged due to the scarcity of megawatts; $30 billion was spent on a large gas-fired power portfolio this year, alongside a $12 billion acquisition of similar assets.

Technological Solutions and Grid Modernization

• Nuclear power is experiencing a resurgence, with hyperscalers partnering to bring retired plants back online (e.g., Mile Island) or perform "uprates" to build additional megawatts at existing sites.
• Small Modular Reactors (SMRs) are a potential future solution, designed at roughly 350 megawatts to match the scale of medium-to-large data centers, though commercial deployment timelines remain uncertain.
• The industry is shifting toward "behind-the-meter" generation, where data centers are co-located directly next to power plants to minimize transmission line loss.
• New technologies focus on "peak shaving," making data centers flexible demand centers that can adjust load to utilize unused grid capacity during non-peak hours.
• Carbon capture and sequestration equipment is seeing increased investment to allow gas-fired plants to serve as clean baseload power, aligning with hyperscalers' emission requirements.

Contractual Structures and Market Outlook

• To mitigate the risk of overbuilding for obsolete hardware, contracts between utilities and hyperscalers now feature minimum tenors of at least 10 years to protect utility returns on capital.
• Data center chip lifecycles are estimated at 3–6 years, whereas power infrastructure assets are 30-year investments, creating a necessity for long-term, guaranteed revenue streams for utilities.
• While power prices are rising, they represent a small fraction of total data center costs and are unlikely to significantly impact hyperscaler returns on investment.
• The sector is shifting from a 20-year era of flat demand to a period of rapid innovation and investment, with policymakers and stakeholders finally prioritizing grid reliability and decarbonization simultaneously.