The startup that taught the electric grid to whisper its secrets - in time for the AI boom to hear them.
It is not a person. It is not, strictly speaking, a forecaster. It is a generative-AI engine running on a server in California, paging through quadrillions of grid permutations - storms, load spikes, contingency outages, the small hourly betrayals of summer afternoons - and looking for something specific. A pocket. A pocket of capacity that nobody asked the grid for, because nobody knew it was there.
"The grid we already have is bigger than the grid we think we have. The math just hasn't caught up."
That model belongs to GridCARE, a 27-person company in Redwood City that has - in the span of about eighteen months - convinced utilities, hyperscalers, John Doerr, and a Google founding board member that the bottleneck for artificial intelligence is not GPUs. It's wires. It's transformers. It's a queue at the substation that, in many U.S. markets, now stretches past 2032.
GridCARE's bet is that most of that queue is unnecessary. Not because the grid is infinite, but because it is mismeasured. Their software finds the slack between the rules and the reality, then walks a utility, a developer, and a regulator through it together. The pitch fits on a napkin: cut time-to-power from five-to-seven years to six-to-twelve months. The science fits on a server.
Conventional interconnection queues for new data-center load can stretch most of a decade. GridCARE's approach compresses the same delivery window by an order of magnitude. The bars below are illustrative, drawn from the company's public claims.
The flagship: a generative-AI model that simulates congestion, weather, outages and demand variability across quadrillions of states to locate flexible capacity in time and space.
A workflow - software plus expertise - that walks AI builders and utilities through siting, interconnection, and timing for gigascale clusters.
Joint planning and flexible-interconnection workflows already in motion with National Grid and Portland General Electric.
The team blends chip-design veterans, Stanford energy researchers, and one of the country's most decorated energy thinkers. Three of the four have published academic papers about the grid. One of them sold his last company to Schneider Electric.
AI demand is outpacing the energy system.Amit Narayan, Co-Founder & CEO
More than 100 gigawatts of data-center capacity is already on the grid - it's just invisible.GridCARE, per TechCrunch coverage
The model finished its run. A pocket of capacity that nobody asked for is now sitting on a project plan in a utility office, flagged for an AI cluster that was, last quarter, on a 2031 list. A planner who used to spend her week reading spreadsheets is reading a map. A founder who used to budget years for power is budgeting months. None of this is loud. None of it has changed the weather. The wires hum at the same frequency they always did.
But the clock that decided how fast intelligence could grow in the United States - that clock just got a new second hand. And it belongs to a 27-person company in Redwood City that thinks the most useful thing you can do with a generative model, right now, is teach it to read a grid.