Will AI deliver the promise of a digital grid?

The honest starting point is that the previous wave mostly did not work. Utilities spent years installing smart meters, now past 70% penetration in the US, along with information technology, operational technology and systems to manage distributed resources, funded through vendor cycles and rate cases. Groarke says the promises of that first wave of digital infrastructure largely did not prove out, and offers as evidence that the cost of the infrastructure needed to deliver power is nearly equal to the cost of generating it. Kann makes the point sharper: if the objective was to reduce transmission and distribution costs, it failed. Groarke agrees without qualification and adds that operating expenses are running up about 14% a year.

What makes this wave different is not a change in the utility but a change in what arrived on top of it. That failed investment left behind something valuable, which is years of multimodal training data covering electrical, customer, weather, gas, visual and text sources. Meanwhile, in Groarke’s phrase, AI was doing press-ups in the background: algorithms got better, handled more tasks and became more efficient, and the cost of building an application collapsed once vendors could assemble solutions from open-source frameworks. That inverts the economics of the last cycle, which was dominated by large platforms, communications infrastructure and hardware. The sector now wants low-cost solutions with high returns, and for the first time software of that shape is available to point at data it already owns.

But the ceiling on how far that goes is physical and institutional, not a question of ambition. Modeling grid physics accurately is genuinely difficult for an algorithm, because anything touching power flow has to respect Ohm’s law and Kirchhoff’s law while making accurate real-time decisions, and for many such use cases the data does not exist at the millisecond granularity required. Around that sit rate-basing, cybersecurity, utility data architectures Groarke calls among the most complex diagrams you will ever see, sales cycles long enough that startups give up, half the workforce retiring within ten years, and competition for capital from transmission buildout. The consequence shows up directly in which use cases are live. Wildfire and vegetation management works because it is discrete, uses genuinely new data from cameras, drones, lidar and satellites, does not touch operations, is not especially political and can be rolled out quickly. Customer propensity modeling works differently: the smart meter data has been sitting there for years, and what changed is the ability to disaggregate a meter signal well enough to detect an electric vehicle charging and act on it, which is running at utilities including Duke and Southern California Edison. Substation asset management is where Groarke sees the most mature combination of information technology, operational technology and AI, and the largest dollars, since extending transformer life and avoiding downtime is worth a great deal. The one case that does touch power flow is transmission capacity optimization, replacing static line ratings with near-real-time sensing, which he calls right in AI’s wheelhouse precisely because it is non-intrusive, though he notes the regulatory direction is still emerging.

Who builds this is somewhat different from last time, though less than a venture investor would hope. Groarke’s team looked at several hundred deployments and found that startups tied to them took about $1.5 billion across 80 funding rounds since 2021, concentrated at the grid edge in distributed energy integration and electric vehicle charging management, with some enterprise work applying language models to regulatory documents. Kann tests the figure immediately, pointing out that venture funding shows startups are being backed, not that they will win, and Groarke concedes the point. Deeper into core operations it remains the incumbents making incremental product improvements and acquiring startups along the way. Then comes the most important sentence in the episode, and it deflates the premise: Groarke says he cannot claim an enormous new market is emerging because of AI, that AI is riding the wave of priorities the sector already had, and that complete reinvention would require policy and business model change rather than better algorithms. Offered three futures by Kann, he picks the middle one: incremental change that eventually produces a very automated system, supported by proven technology and regulatory infrastructure. Full real-time optimization he considers potentially possible but distant, since today’s data is AMI at minutes to hours, SCADA at seconds to minutes and phasor measurement units at milliseconds, most analysis happens after the event, and getting there would need something close to complete nodal visibility.