Can AI revolutionize grid operations?

Planning is not one activity. It is bulk resource planning, transmission planning, distribution planning and distributed-energy-resource planning, each siloed, and inside one large East Coast utility Wong counts roughly a dozen departments doing transmission planning separately. Every request becomes its own study, even though connecting a generator and connecting a load run against the same model of the same grid, and half or more of the effort, especially in distribution, goes into cleaning scattered data. The analysis is then tuned to a worst case, the five hours in a year that will be worst over the next decade. Better practice is moving toward hundreds of representative hours and, as a gold standard, all 8,760, which is what you need if storage or flexible load is part of the answer. Then the assumptions change and the whole thing is restudied. Wong’s old line is that planners are always right, because they can blame the assumptions, and always wrong, because the assumptions are never correct. Kann asks whether the price is systematic overbuild; Wong half agrees, calling the grid an engineering marvel for its time that failed to keep up with digitalization.

The deeper problem is what a study returns. A decades-old simulator reports congestion and voltage violations, which amounts to a red or yellow light: restrict access, or build more lines. What to do is left to the planner’s experience. That was survivable while the answer was always more substations, wires and transformers. Affordability pressure and supply chain lead times have taken that answer away, so the menu now includes curtailing generation or load, adding a battery of some size somewhere on some charge schedule, and reconfiguring switching, several of which are operations decisions a planner does not own. Against millions of candidate line segments, trial and error cannot scale. Volume makes it worse: a cluster or large-load transmission study runs six to nine months and costs roughly a quarter of a million dollars, restudies compound that, and clustering helps less than it looks because the queue keeps changing. Distribution studies are simpler and parallelizable, but the data is worse and the volume far higher.

His answer is deliberately not a large language model. ThinkLabs trains deep machine learning models at the intersection of the physics equations and the model itself, so the system learns how power flows and how loads behave under contingency. The point of the physics is trustworthiness, since the utility authorizes the interconnection and has to believe a result enough to act on it. Two properties follow. The models are deterministic, so the same question returns the same answer and they cannot hallucinate, which is the first objection a utility would raise. And they are trained per utility on that utility’s own data, with no cross-training, which answers the security objection. The payoff he emphasizes is not only speed. It is that the model becomes generative: pre-train across whole ranges of futures, from load growth to renewable penetration to data center buildout, and inference can propose the upgrades that solve a region rather than listing what is broken.

Operations is where the episode turns. Kann assumes the ceiling is lower there, since the grid already looks automated. Wong says the opposite. The automation that exists is field automation, with substations and generator sites acting on their own faster than any human could intervene, which is precisely why reliability is so good. What humans do is alarm management, rules and dispatch, plus switching plans a week or two out deciding which lines can come out of service for planned work, and that last one is quietly becoming a crisis, because a congested grid leaves little room to take anything out of service in order to upgrade it. The business case today points at planning because interconnection is the growth engine, but Wong argues the durable value has always been reliability and workflow efficiency. The missing piece is a feedback loop: nothing learns from a day’s events until a postmortem, and the records never reconcile into one account of what happened. The same gap explains why planning and operations barely speak, since study results reach operators, if at all, as a spreadsheet of fixed criteria. His diagnosis is that no platform exists that can hold all the planning horizons at once, so utilities have organized themselves around the limits of their tools.