Episode Summary
Executive Summary: Shail Khan and Josh Wong discuss how utility planning and operations still rely on siloed, manual, worst-case studies that are slow, expensive, and disconnected from day-to-day grid reality. Wong argues that physics-informed AI can make utility analysis deterministic, fast, and scalable—turning studies from months into minutes and enabling continuous, closed-loop grid optimization. They also examine data centers, microgrids, and whether large loads should be treated as micro-utilities.
Main Topics: Utility planning today is siloed and manual (Priority: 5/5): Wong explains that transmission, distribution, and DER planning often occur in separate departments with separate studies, and much of the work is spent cleaning bad data before analysis even begins. Worst-case planning limits grid flexibility (Priority: 4/5): Current utility studies are often designed around a small number of worst-case hours rather than full-year behavior, which is useful for reliability but slow to adapt to modern flexibility resources like batteries and demand response. AI for physics-informed grid planning (Priority: 5/5): Think Labs aims to use AI that learns grid physics to automate power flow and interconnection studies, generate solutions, and produce deterministic outputs utilities can trust. Operations are more automated than most people realize, but still constrained (Priority: 4/5): Wong says field automation is strong, while control-room work remains heavily focused on alarm management, switching, dispatch coordination, and responding to abnormal events. Planning and operations are poorly connected (Priority: 5/5): The transcript emphasizes that study results rarely feed directly into operations, and utilities lack an analytical platform that unifies long-, medium-, and short-term decision-making. Data centers and behind-the-meter generation (Priority: 4/5): The discussion closes on whether large data-center microgrids are hard to operate and whether they should be treated not as isolated microgrids but as micro-utilities that help strengthen the broader grid.
Key Arguments: Utility planning is slowed by siloed departments, inconsistent data, and repeated restudies; this creates long interconnection timelines and high internal costs. Most current studies are built for worst-case scenarios, which preserves reliability but does not scale well to a more dynamic grid with flexible loads and storage. Physics-informed AI is better suited than general LLMs for grid work because utilities need deterministic, trustworthy outputs rather than probabilistic language generation. Think Labs claims its models can be trained on a utility's own proprietary data, avoiding cross-utility data sharing while still achieving high accuracy and fast runtimes. AI can do more than identify problems; it can generate solutions and optimize options such as line upgrades, storage placement, load flexibility, and operational switching. Operations are not just about outage response; they include switching plans, alarm triage, dispatch coordination, and emergency preparation, all of which could benefit from continuous AI analysis. Planning and operations currently lack a closed feedback loop, so utilities fail to learn systematically from daily events, historical utilization, and post-event investigations. Behind-the-meter generation may help individual data centers move faster, but it can also create new operational complexity and should be viewed as a potential benefit to the broader system, not just a private workaround.
Data Points: Utility study duration: 6-9 months - Typical time to perform a large transmission or interconnection study Cost per study: about $250,000 - Typical internal utility cost for one study SCE energization requests: up to 10,000 per month - Volume cited for distribution-side requests Manual study turnaround: 30 to 45 days - Time for each distribution energization request if handled manually AI training time: about 10 minutes - Think Labs training run for a power-flow model the size of a state Training compute cost: about $5 - Approximate compute cost per training run Model accuracy: over 99.9% - Think Labs claims accuracy across system states such as voltages and line flows Inference speed: sub-second - Claimed runtime for 87,600 power-flow scenarios Solution generation example: about 15 minutes - Time to generate more than 10,000 lines for a large utility low-growth region study Device aggregation: 2.5 million customer devices - EnergyHub sponsor mention about VPP aggregation Dispatchable capacity: 3.4 gigawatts - Capacity from aggregated thermostats, batteries, and EVs in EnergyHub VPPs Peak-period activity: May and June alone - Sponsor mention describing device shifting during peak periods
Pivotal Quotes: "the greatest place where we stuck, though. Is all the studies tell you problems, not solutions?" — Josh Wong: On why utility planning needs automation that can generate remedies, not just flag violations "Can AI help us plan the grid? To run a system study. How do generators and loads and battery storage impact the grid?" — Josh Wong: Describing Think Labs' core mission for physics-informed utility AI "I believe the grid has enough existing latent capacity to connect the majority, if not all, of the data centers today." — Josh Wong: On data center load growth and the belief that planning/operations modernization matters more than new generation alone
Implications: Utilities may be able to unlock capacity faster by modernizing analytics and workflows than by only building new wires and plants. If AI can make planning and operations continuous, deterministic, and connected, the grid could become more reliable, flexible, and responsive to data centers, DERs, and electrification.