The Latitude stage: Solving the AI load growth puzzle

Episode Summary

Executive Summary: The episode examines how AI-driven data center growth is reshaping U.S. electricity demand, utility planning, and grid investment. Rob Gramlich and Anuja Ratnayake argue the demand surge is real, but queue data overstates what will actually be built. They discuss transmission, supply-chain constraints, gas versus renewables, data-center flexibility, and the risk of stranded assets and rising retail rates.

Main Topics: AI-driven load growth and the end of flat power demand (Priority: 5/5): The speakers frame AI infrastructure as a major new source of electricity demand, with growth large enough to alter utility planning and potentially support the broader economy. Separating real demand from speculative interconnection queues (Priority: 5/5): They distinguish between genuine future load needs and inflated queue volumes created by developers applying to multiple utilities and projects at different stages of maturity. Transmission as the primary near-term bottleneck (Priority: 5/5): Both speakers identify transmission as the most acute constraint, emphasizing that it can unlock large amounts of capacity more efficiently than incremental reactive builds. Gas, renewables, and resource adequacy (Priority: 4/5): The conversation weighs the role of gas for firm capacity against the continued need for wind, solar, and storage, stressing that the grid will need a diverse portfolio. Data centers as flexible grid resources (Priority: 4/5): EPRI’s DC Flex concept is presented as a way to treat data centers as partially dispatchable loads, using workload, cooling, and backup systems to reduce interconnection friction. Supply-chain stress, labor shortages, and rate impacts (Priority: 4/5): The speakers warn that transformers, breakers, pipelines, and specialized labor are all constrained, which could raise costs and drive retail rate increases. Risk of stranded assets and poor planning (Priority: 4/5): Both discuss the danger of building the wrong infrastructure if planners rely on outdated assumptions instead of proactive, flexible, and financially committed load planning.

Key Arguments: AI load growth is real and substantial, but the exact buildout pace is uncertain because many announced projects are speculative or duplicated across multiple interconnection queues. Utilities cannot treat every queue entry as firm demand; they need discounting frameworks based on project maturity, land acquisition, engineering progress, and financial commitment. Transmission is the most powerful lever for meeting new load because it delivers scale and economies of scale more efficiently than smaller, reactive local upgrades. The near-term system still needs all available supply options: gas for firm capacity and reliability, plus wind, solar, and storage for cost and emissions benefits. Data centers do not have to be perfectly inflexible; if they can respond to peak conditions, transmission constraints, and contingency events, they can interconnect faster and at lower cost. Co-location alone is not a full answer because grid reliability and diversity of resources still provide the strongest system-wide reliability and cost advantages. The power sector is entering a period of tighter supply chains and competition for equipment and labor, making rate increases and stranded assets more likely if planning is not proactive. The most important planning shift is to move from reactive, incremental builds to proactive regional planning with clearer financial commitments and better load forecasting.

Data Points: New U.S. load by end of decade: about 120 gigawatts - Rob Gramlich’s estimate of new load growth across the U.S. power system Share of new load from data centers: about half - Gramlich’s estimate that roughly half of new load will be data centers Share of data-center load that is AI-driven: about half - Within the data-center portion, roughly half is expected to be AI-related Potential share of new U.S. load from data centers by end of decade: half - Host framing early in the episode, citing forecast discussions Typical utility planning horizon vs. data center expectations: 7-10+ years vs. 2-3 years - Anuja Ratnayake contrasts utility interconnection timelines with data-center deployment targets Peak periods used for flexibility planning: maximum of 15 days or 300 hours per year - Ratnayake cites the limited annual hours when flexibility can relieve grid stress Voltage comparison for transmission economics: 765 kV vs. 230 kV delivers about 4x more power for the same dollar - Gramlich argues higher-voltage transmission is far more efficient Large-load facility comparison: 1 GW facility is about 10x Tesla Gigafactory; about 500 Costco wholesale facilities; about the size of Charlotte, NC or Columbus, OH - Ratnayake uses comparisons to illustrate the scale of hyperscale data centers Data center buildout scale: more than a dozen announced 1+ GW facilities - Ratnayake highlights the unusually large and clustered nature of announced projects Load diversity concern: up to 60% of a utility’s peak load from a single industry - Ratnayake describes the risk for local planners as data centers dominate growth Conference timing: June recording; Transition AI 2026 on April 13-14 in San Francisco - Contextual event and recording references in the episode

Pivotal Quotes: "The era of flat power demand is over." — Rob Gramlich / Grid Strategies report title: Referenced as the turning point that signaled the urgency of load growth planning "The second question is the one that keeps the utility industry up at night." — Anuja Ratnayake: On distinguishing real AI demand from speculative interconnection-queue entries "Transmission is the best opportunity to expand capacity significantly." — Rob Gramlich: On the most effective near-term solution to large-scale load growth

Implications: Utilities should plan for real AI demand while discounting speculative queues, prioritize transmission, and build flexibility into data-center interconnections. Failure to do so risks higher rates, stranded assets, and delayed grid expansion.