Nine Industries That Have to Work Together to Power AI

AI growth depends on both compute and electricity or data centers sit in the middle of a much larger infrastructure ecosystem. Getting new data centers online to power AI requires coordination across a vast supply chain that spans chip making in Taiwan, building gas turbines in South Carolina and drilling for gas in Pennsylvania.

Developers are required to watch each step of the supply chain and look for emerging bottlenecks.

The AI energy tech stack:

• AI platforms and hyperscalers: The demand side of the stack. Compute buildouts from AWS, Microsoft, Google and OpenAI are the primary driver of new electricity demand across every layer below.
• Compute and infrastructure: Chip and server capacity is gated by fab output rather than demand. TSMC, SK Hynix and Nvidia allocation schedules determine how quickly new racks can be deployed, independent of how much power is available.
• Data center developers and operators: Land acquisition and shell construction generally proceed faster than power procurement. In most markets, securing an interconnection agreement is now the longer lead item in a project timeline.
• On-site power: Turbine orders from providers like GE Vernova and Siemens Energy carry multi-year lead times. Developers are increasingly building on-site generation to bypass utility interconnection delays rather than waiting for grid capacity.
• Utilities: Interconnection queues in PJM, ERCOT and MISO run several years for both new generation and large new loads. Utilities also face their own capital constraints in upgrading distribution and substation infrastructure to serve data center customers.
• Independent power producers: IPPs are contracting directly with hyperscalers through long-term power purchase agreements, similar to the Fervo-Google framework. Permitting and financing timelines, particularly for non-gas generation, remain the primary constraint on how quickly new supply can come online.
• Grid infrastructure and transmission: New transmission lines routinely take five to ten years to permit and build. In the first quarter of this year alone, 198 gigawatts of large load applied for ERCOT interconnection, with 86 gigawatts under review, roughly equal to ERCOT's entire current peak load.
• Power equipment makers: Vertiv, Eaton, Schneider Electric and Siemens Energy are reporting order backlogs stretching multiple years. Manufacturing capacity for transformers and switchgear has not kept pace with simultaneous demand from data centers, grid modernization, and electrification.
• Fuel supply: Gas-fired generation depends on pipeline capacity that is increasingly contested by LNG export demand. Domestic fuel supply for new gas plants is no longer guaranteed simply because a plant has been permitted.

When bottlenecks emerge, startups are often created to address the issues. CoreWeave was founded in 2017 to address a growing crypto market but pivoted to AI-focused data centers.