There is a straightforward problem behind this project. Getting a new large data center connected to the U.S. electricity grid now takes an average of four years. In the most congested markets, like Northern Virginia, the wait stretches to seven. Texas saw interconnection requests jump 700 percent in a single year. The grid was never designed for this.
So the question became: what power already exists, right now, that nobody is using?
The answer turned out to be the electrical substations already scattered across the country. There are 55,000 of them in the U.S., and most carry a few megawatts of unused capacity on any given day. Nvidia, real estate firm Prologis, builder InfraPartners, and the nonprofit power research group EPRI are now building a network of small data centers positioned right next to those substations, tapping that idle power. Five pilot sites are planned across the U.S. by the end of 2026. If one location gets overloaded, the computing work shifts automatically to another one with spare room.
This matters because of where AI is going, not just where it is today. For the past few years, most AI work has been about training, which means building a new AI model from scratch. That requires a massive, concentrated facility, thousands of connected computers working in lockstep, and months of continuous operation. You cannot split that across 25 small buildings.
But training is slowing down as the dominant workload. The fast-growing part of the market is now inference, which simply means using an already-trained model to answer a question, generate a document, flag a fraud case, or read a medical scan. Inference can be done in smaller facilities, can be routed dynamically to wherever power is available, and does not require that tight hardware connection. Inference workloads are projected to grow at 35 percent per year through 2030, outpacing training. Deloitte estimates inference will account for roughly two-thirds of all AI computing by the end of 2026. IDC expects that by 2027, most large companies will be sourcing AI inference from distributed locations rather than central cloud data centers.
That is exactly what this pilot is designed to serve. And it carries a second advantage: speed. Building next to an existing substation means the fiber cables are already there, the land is often industrial, and the permitting path is shorter than a greenfield site. The whole model is built around the idea that you can get to operational faster than the traditional queue allows.
There is a broader political and economic context here that affects businesses far beyond the tech sector. U.S. electricity prices rose 11.5 percent in 2025 alone. Utilities requested more than 29 billion dollars in rate increases in just the first half of that year, double the amount from the year before. The cost of building new grid infrastructure to serve data centers is being passed, at least partially, to regular electricity consumers. Wholesale electricity costs near major data center clusters have risen as much as 267 percent compared to five years ago. That is showing up in business operating costs across every industry.
A model where AI compute plugs into spare existing capacity, rather than demanding entirely new power plants and transmission lines, puts less pressure on that system. It does not solve the problem, but it changes the direction of it.
Nvidia's strategic interest here is also worth noting plainly. More distributed AI inference means more of their hardware deployed in more locations. The company is not doing this out of public spirit. But the infrastructure outcome and Nvidia's commercial interest happen to align in a way that could accelerate something useful.
The deeper signal is this: the geography of AI is about to spread out. The next wave of AI infrastructure will not all look like a campus in the Virginia suburbs. It will look like a modest building next to a power substation in a mid-sized city, quietly handling the questions that millions of people and business systems are sending to AI tools every day. For businesses planning their own AI adoption, that means the compute they rely on may soon be physically closer and potentially more resilient than it is today.