Episode Summary
Executive Summary: The episode argues that data-center growth need not default to off-grid gas. Jesse Jenkins and Astrid Atkinson outline a two-part model—flexible interconnection plus a portfolio of grid, battery, compute, and clean resources—that can meet new load faster, cut costly grid upgrades, and avoid ratepayer harm. In practice, only a few hours of flexibility per year may unlock major capacity.
Main Topics: The data-center power crunch and the rush to gas (Priority: 5/5): The hosts describe a grid bottleneck: surging load from AI/data centers, clogged interconnection queues, and a widespread industry reaction of building on-site gas generation, including turbines and temporary generators. Flexible interconnection as a faster grid solution (Priority: 5/5): The core proposal is to allow data centers to connect with some non-firm capacity, so they can be curtailed a small number of hours annually instead of forcing expensive wires upgrades for 100% firm service. Bring-your-own-capacity portfolio model (Priority: 5/5): Instead of on-site gas alone, the guests advocate a portfolio of resources—batteries, compute flexibility, VPPs, nearby renewables, and sometimes gas firming—to meet both transmission and capacity needs. Compute flexibility and data-center operations (Priority: 4/5): The discussion explains that not all compute must run continuously; cloud providers can shift workloads in time/space, and some data-center jobs can be tiered for curtailment, making flexibility operationally real. Distributed energy resources and VPPs as capacity (Priority: 4/5): The episode emphasizes that aggregated distributed batteries and demand response can count as accredited capacity, especially for generation adequacy, though utilities are still slower to accept them for transmission-constrained interconnection. Regulatory/process reform and utility tooling (Priority: 4/5): The guests argue the main barriers are process and software: utilities need new interconnection rules, better modeling tools, and willingness to evaluate thousands of flexible scenarios instead of rigid worst-case planning. Economic and climate implications (Priority: 5/5): Flexible interconnection can shift costs off ratepayers, speed connection by a year or more, and avoid locking in gas and coal. The guests frame this as a fork in the road for the AI buildout.
Key Arguments: The current default—building behind-the-meter gas generation—is operationally inefficient, slow, and risky, and it likely locks in high-emitting assets. Data centers do not need 100% firm grid service 100% of the time; a small amount of annual flexibility can avoid major grid upgrades and unlock much more capacity. Only a tiny fraction of hours may need curtailment in practice, so batteries and load shifting can cover most events cheaply relative to gas. A portfolio approach is superior to relying on one resource: combine flexible interconnection, batteries, compute flexibility, VPPs, renewables, and limited firming resources if needed. Utilities and data centers are both interested but hesitant; the main barrier is not technology but aligning new contracting, planning, and interconnection processes. Distributed resources should count as real capacity for generation adequacy, but utilities are less comfortable using them to solve local transmission constraints without new rules and software. Because gigawatt-scale data centers are enormous loads, getting online sooner has very large economic value, making even relatively expensive flexibility measures worthwhile. The grid should be used as the default delivery backbone; off-grid solutions are a fallback, not the optimal model for most sites.
Data Points: Planned gas plant additions in the U.S.: 85 gigawatts - Halcyon is tracking planned gas generation additions, reflecting the current buildout trend around data centers. Typical data-center flexibility need discussed in the intro: 10% to 20% of the time - Host’s framing of the idea that data centers could run on their own batteries/generators during a limited share of hours. Actual curtailment in modeling: 7, 11, 13, and 35 hours per year - For four of the six modeled sites, annual curtailment needs were extremely small. Largest annual curtailment share: 0.4% of hours of the year - At the highest end of modeled flexibility need for one site. Longest modeled curtailment event: 16 hours - One site had a single longer event, but most were much shorter. Long events frequency: 3 or 4 per year - Modeled events were infrequent, supporting battery-based ride-through. Grid capacity unlock estimate: ~30% - Back-of-the-envelope estimate of extra capacity possible with flexible interconnection, after regional variation and practical constraints. Data-center utilization relative to nameplate: ~40% on average - General statement that many data centers operate below peak nameplate most of the time. Cooling load share at peak: 150 to 250 MW for 800 MW of compute - Cooling systems can add substantial electrical load, creating both a challenge and a flexibility opportunity. Opportunity cost of one year of earlier online date: ~$7 billion per gigawatt per year - Jenkins estimates the value of bringing a gigawatt-scale data center online a year earlier.
Pivotal Quotes: "For data centers, a little flexibility goes a long way." — David Roberts: Opening framing of the episode’s thesis. "The current ability to do this varies between provider. Typically, hyperscalers who actually operate a very large fleet have more ability to do this just because they have more locations to put things." — Astrid Atkinson: On the uneven but real potential for compute flexibility across cloud operators. "There’s a fork in the road." — Jesse Jenkins: On whether the data-center buildout locks in gas or pivots to flexible, cleaner grid-connected solutions.
Implications: If adopted, flexible interconnection and portfolio-based capacity could speed data-center deployment, reduce grid upgrade costs, and steer AI growth toward cleaner power instead of a gas lock-in. It also creates a major market for batteries, VPPs, and flexible load software.