WASHINGTON — The future of artificial intelligence infrastructure will rely on clusters of smaller, interconnected data centers rather than massive, centralized facilities, according to a leading industry executive speaking at Data Center World this week.
.jpg)
Pete Sacco, founder and CEO of PTS Data Center Solutions, said the shift is being driven by the rapid rise of AI inferencing — the process of responding to user queries in real time — which requires computing resources to be located closer to end users.
“Clusters of small data centers requiring only 5-20 MW of power are the future of the AI boom,” Sacco said during an April 21 presentation.
Traditionally, hyperscale data centers have dominated the market, often exceeding hundreds of megawatts in capacity. However, Sacco noted that multiple challenges — including lengthy grid interconnection delays and increasing community opposition — are forcing developers to reconsider that model.
A key factor behind the transition is the evolving nature of AI workloads. While large, centralized facilities have been effective for training models, inferencing demands far lower latency and faster response times.
“It can take a far-away data center something like 10 milliseconds to answer a query — far too long for inference computing,” Sacco said. “A lag of one millisecond is closer to what’s expected.”
To meet these requirements, data centers must be physically located closer to users. Sacco emphasized that this shift will fundamentally reshape how facilities are designed and deployed.
“You can’t have 500-MW data centers sitting in the New Mexican desert and be able to deliver real-time inferencing within millisecond scale,” he said.
Inferencing is expected to surpass training as the dominant form of AI computing as early as next year, accounting for more than 55% of total demand, according to Sacco. “That requires a different infrastructure model,” he added.
Sacco is advancing this concept through a new venture, Gray Wolf Data Centers, which aims to develop distributed networks of smaller facilities. Instead of a single large site, the model envisions dozens — or even hundreds — of smaller data centers operating together.
“Instead of having a 1,200-MW data center — which there are no more places for — I can build 120 10-MW data centers in a region [and] glue them all together,” he said.
Sacco compared the approach to a franchise-style system: “The idea is to be the Starbucks of the data center industry. Some of them will be company-owned. The majority will be franchised-owned.”
These facilities would function as part of a distributed autonomous organization (DAO), allowing decentralized decision-making and shared operational frameworks across the network.
Energy strategy is also evolving alongside infrastructure. Sacco said each data center could rely on localized power solutions, including grid connections, microgrids powered by solar energy and battery storage, and potentially hydrogen-based systems.
.jpg)
“There also will ‘probably be a proliferation of hydrogen’ used to generate electricity,” he said, adding that small nuclear reactors could become viable within the next decade.
“The days of the centralized utility are gone,” Sacco said.
As a proof of concept, Gray Wolf is developing its first facility in Connecticut, a region with high energy costs but strong demand for low-latency computing. The project includes plans to convert carbon-based waste into energy, potentially lowering electricity costs while supporting additional facilities in the region.
“In essence, I can produce electricity at sub 10 cents a kWh,” Sacco said. “I can compete against utilities… I can sell it for 20 cents a kWh, and I look like a hero because it’s 29 cents a kWh normally.”
The emerging decentralized model signals a significant shift for the construction and development of data center infrastructure, as builders adapt to new performance requirements and evolving energy strategies tied to the next phase of AI growth.
Originally reported by Robert Freedman, Lead Editor in Facilities Dive.
.png)
.png)
