Episode Summary
Executive Summary: The episode features Shail Khan interviewing data center energy expert John Coomey about whether AI will cause explosive electricity demand growth. Coomey argues the hype is overstated: historical forecasts have often been wrong, efficiency gains and infrastructure limits can blunt load growth, and demand may not scale infinitely. He expects near-term growth, but doubts long-term projections beyond a few years.
Main Topics: Historical errors in computing electricity forecasts (Priority: 5/5): Coomey reviews past overestimates from the dot-com era, arguing they made him skeptical of current claims about AI-driven power demand. AI demand versus efficiency as the two key variables (Priority: 5/5): He frames future data center electricity use as a balance between rising service demand and improving efficiency in delivering compute. Why current projections are so uncertain (Priority: 4/5): The discussion highlights proprietary data, time lags in reporting, and inconsistent institutional estimates as reasons historical and forward-looking numbers are unreliable. The conventional wisdom around infinite AI demand (Priority: 5/5): Khan describes the prevailing industry view that AI demand is effectively unlimited and will overwhelm grid constraints, while Coomey questions that assumption. Efficiency innovations and possible future bends in the curve (Priority: 4/5): Coomey argues that AI infrastructure could eventually see major efficiency gains through better hardware, software, algorithms, and special-purpose systems, similar to earlier computing transitions. Infrastructure and utility planning constraints (Priority: 4/5): The conversation covers real-world bottlenecks such as transmission, generators, transformers, labor, and utility rate design that shape how much load can actually materialize. Near-term growth, but skepticism about long-term explosive expansion (Priority: 5/5): Coomey concedes data centers may keep growing for a few years, but rejects confident claims about 2035-scale demand and says utilities should plan cautiously.
Key Arguments: Past electricity-use forecasts for computing were dramatically overstated, which should make analysts cautious about current AI load projections. Data center electricity use is driven by two uncertain factors: service demand growth and efficiency improvements. Many forecasts implicitly assume NVIDIA-like hardware sales trajectories and treat them as destiny, despite major uncertainty. Jevons-paradox style arguments may apply, but Coomey thinks rebound effects are usually modest and do not prove infinite demand. AI models still face unresolved reliability and hallucination issues that could limit commercial adoption and therefore electricity demand. The industry has many paths to better efficiency beyond Moore’s Law, including hardware design, software optimization, algorithms, and special-purpose computing. Infrastructure constraints in the physical world mean even very wealthy AI companies cannot scale instantly; power and grid buildout take time. Utility rate design and take-or-pay contracts can reduce risk and prevent existing customers from subsidizing speculative data center builds. A reasonable expectation is continued growth over the next few years, but not necessarily the explosive, unconstrained trajectory often described in public discussion.
Data Points: Historical share of electricity used by computing in 2000: 3% - Coomey says analysts found actual computing electricity use was 3%, not the 13% being claimed at the time. Claimed computing electricity use in 2000: 13% - An overestimate circulating during the dot-com era. Internet electricity use projection: 50% of all electricity in 10 years - A widely cited but erroneous late-1990s/2000-era forecast. Data center electricity growth, 2000-2005: doubled - Coomey says data center electricity use roughly doubled from a small base during this period. Data center share of electricity by around 2010: about 1% - After efficiency gains and industry response, data centers reached about 1% of electricity use. Compute output growth, 2010-2018: 6-fold increase - Service demand for data center compute rose sharply over this period. Electricity use growth, 2010-2018: 6% - Despite major growth in compute, electricity use rose only modestly. IEA global data center electricity estimate range for 2022: 220 to 340 TWh/year - Coomey cites a wide uncertainty range even for historical global data center electricity use. IEA estimate revision: 50% difference - The IEA’s January 2024 and October 2024 studies differed by 50% on the same historical year. VPP customer devices: 2.5 million - Energy Hub aggregates thermostats, batteries, and EVs into a virtual power plant. VPP dispatchable capacity: 3.4 GW - Energy Hub’s customer devices provide this amount of flexible grid capacity. Equivalent grid capacity: more than three nuclear reactors - Used to illustrate the scale of virtual power plant flexibility. U.S. data center share of electricity in 2020-2021: about 2% - Coomey cites an LBL report for the U.S. historical estimate. U.S. data center share of electricity in 2023: 4.4% - The same LBL bottom-up calculation shows significant growth by 2023. U.S. total electricity generation trend: 2023 lower than 2022 - Khan notes aggregate generation did not show explosive growth even as data center use rose.
Pivotal Quotes: "they're just saying, well, NVIDIA has this projection for the next few years of how many AI nodes they're going to sell. And they just use that and say, that's what's happening." — John Coomey: On how many projections simply extrapolate GPU vendor growth into power demand forecasts. "I think anyone who claims to know what's going to happen in 2035, the data center electricity use, I think is just making it up." — John Coomey: On the limits of long-term forecasting for AI/data center load. "if you make a thing cheaper with sort of endless potential demand for the thing, people will just use more of it in the end." — Shail Khan: Khan summarizes the Jevons-paradox/conventional wisdom view of AI efficiency and demand.
Implications: Listeners should expect AI-related electricity demand to grow, but not assume endless linear expansion. Forecasts are highly uncertain, efficiency can move fast, and utilities should use cautious, contract-based planning to avoid stranded costs.