The Three-Minute Grid

A new customer asks an electric utility for power. Perhaps it is a factory. Perhaps it is a charging depot. These days it may be a data center, the sort of building that turns electricity into artificial intelligence and heat. Before the utility can say yes, an engineer has to determine what that new appetite will do to the network. At Southern California Edison, the work behind one such request typically consumed about six hours of preparation, calculation and reporting. The request itself could spend 30 to 35 days in the queue.

Then ThinkLabs AI put more than 100 circuits through a different routine. Its models trained in minutes per circuit. A full year of hourly power-flow data ran in under three minutes. A proposed remedy and engineering report arrived in under 90 seconds.

That contrast is the company's whole argument in miniature. The electric grid is changing at machine speed; much of the analysis that governs it still travels at institutional speed. ThinkLabs wants to close the gap.

The stubborn thing about electrons

Ordinary generative AI is persuasive because language is forgiving. A sentence can be rephrased. Electricity is less sociable. Voltage, current and impedance follow physical relationships whether the software has understood them or merely produced a confident answer.

ThinkLabs' answer is to let physics act as the teacher. Utility network data becomes a conventional engineering model. Power-flow solvers then generate large sets of synthetic examples: different loads, generators, storage devices, switches, outages and weather-shaped conditions. Machine-learning models train on those examples and are fine-tuned as the real network changes.

This is the useful middle ground. Traditional simulations are trusted but can be computationally expensive and brittle when utility records are untidy. General AI is quick but, on its own, has no instinct for Kirchhoff's laws. ThinkLabs is trying to retain the guardrails of the first and borrow the pace of the second.

“The grid is the most critical infrastructure on earth right now, and it's being asked to do something it was never designed to do.”Josh Wong, founder and CEO

The second first company

Josh Wong came to this problem twice. He began his career at Toronto Hydro and later founded Opus One Solutions, whose software helped utilities plan and manage distributed energy. GE acquired Opus One in 2021. Wong went on to run grid orchestration at what became GE Vernova, then left to found ThinkLabs. GE Vernova launched the new venture in May 2024, called it the first startup to emerge from the company and supplied its initial backing.

The return trip matters. Wong had already watched a grid absorb solar panels, batteries and electric vehicles. Inside GE, he could see both the reach of incumbent utility systems and the stubbornness of the workflow around them. His conclusion was not that the old control systems should be thrown away. It was that they needed a faster intelligence layer.

ThinkLabs Copilot therefore sits alongside the software a utility already trusts - advanced distribution management, distributed-energy management and energy-management systems. It estimates the state of the network, spots congestion or voltage violations, tests contingencies and proposes actions such as changing network topology, using storage or calling on flexible load. The incumbent system remains the communications and control backbone. The operator remains accountable.

What failed first was the calendar

The enemy is not a dramatic blackout in the demo room. It is delay. Utilities often study a handful of worst-case snapshots because a full sweep of possible conditions costs too much time and engineering attention. Data must be cleaned. Models must be prepared. Reports must be assembled. By the time a study lands, the queue has grown and the network may already look different.

The financial cost is equally plain. Wong has said a single legacy study can consume tens to hundreds of thousands of dollars in engineering time. ThinkLabs has not published product pricing. One early investor estimated potential annual utility contracts from $1 million to $10 million and described a tiered software subscription tied to the number of end customers served. Treat that as an investor's expectation, not a rate card. The purchase makes sense only when repeated studies, avoided upgrades or faster connections are worth more than the software and integration effort.

A proof point, not permission to coast

The SCE result is unusually specific, which makes it valuable. It is also a defined collaboration, not proof that an autonomous agent is running California's grid. The first phase modeled a subsystem and automated an energization workflow. Moving from that success to continuous operational use means security reviews, integration with live systems, performance monitoring, change management and regulatory confidence.

Nor does physics-informed mean data-independent. A model can tolerate imperfect records better than a brittle optimization routine and still be limited by missing topology, bad sensor readings or an operating condition outside its training envelope. Recommendations must be checked against protection schemes, field constraints and local operating practice. In a utility, 99.8 percent accuracy can be impressive and still leave the important question: what lives in the remaining fraction?

This is where ThinkLabs' engineering-heavy culture helps. Its leadership includes veterans of GE Vernova, Powertech Labs, Eaton, Itron, Rivian and Toyota Financial Services. The team speaks the dialect of wide-area monitoring, model validation and regulated procurement. It joined the Total Grid Orchestration Alliance alongside utilities and vendors, and it works in an ecosystem that includes Microsoft, NVIDIA and EPRI. Selling into a cautious industry requires technology; being understood by that industry is a separate product.

The recipe hiding in the substation

There is a playbook here for anyone applying AI to a physical system. Choose a narrow workflow whose delay is measurable. Use a trusted simulator to manufacture examples that the real world rarely supplies. Put hard constraints around the model. Integrate with the installed system instead of demanding its removal. Keep a skilled human at the decision point. Then publish a field result with numbers an engineer can dispute.

The approach travels best where the governing physics is known, simulation is credible, actions repeat and the cost of waiting is high: energy networks, industrial processes, transportation or water systems. It travels poorly when the underlying model is weak, the environment changes in ways training never captured, or the organization cannot maintain its data. It also struggles when a one-off decision is cheaper to make by hand than to integrate, validate and govern in software.

ThinkLabs raised a $5 million seed round at launch and another $28 million in March 2026. Energy Impact Partners led the later round; NVIDIA's NVentures and Edison International joined, along with returning investors including GE Vernova. The money is meant to expand delivery, products and partner deployments. That is sensible. The next test is less photogenic than a benchmark: can a young vendor install this machinery repeatedly inside old, consequential institutions?

AI created part of the demand now crowding the grid. ThinkLabs offers a pleasing reversal - use the same family of tools to reveal capacity, clear interconnection queues and make better use of the wires already hanging above us. The irony is neat. The engineering, thankfully, is not supposed to be.