Startup Wants to Turn Spare Office Power Into Edge AI Data Centers
A New York startup backed by Montauk Capital is coming out of stealth with a plan to plant AI servers in the basements and mechanical rooms of commercial office buildings, using power those buildings no longer need. Perimeter Compute, led by former Lambda executive David Hall, said it has identified more than a gigawatt of spare capacity sitting idle in U.S. office towers — enough to run clusters of Nvidia GPUs without waiting years for new data center construction.
The Pitch: Buildings Already Have What Data Centers Need
Hall told Latitude Media that walking through buildings in New York and Boston changed how he sees the grid. "I walked through some buildings in New York and Boston and was astounded that they have everything a data center has: space, cooling, and power," he said. Years of energy conservation work have cut power consumption in these buildings by 30% to 60%, yet the electrical infrastructure was still sized for the peak load they started with. That gap is the product.
Perimeter plans to install the latest AI chips in Class A buildings with between half a megawatt and 20 MW of spare capacity. Even a modest 0.5 MW slice can power roughly 240 of Nvidia's newest GPUs, Hall said — enough to process billions of inference requests a day. The company pays for the GPU hardware, installation, and metered electricity, then splits compute revenue with the landlord.
The economics matter because the central problem in AI infrastructure right now is not demand — it's delivery. Sightline Climate's research found that 30% to 50% of large data centers scheduled to come online in 2026 will be delayed by power constraints, equipment shortages, and local opposition. A quarter of the 140 projects tracked globally have not even disclosed how they plan to get electricity. Delays were already the norm in 2025, when 26% of 110 projects slipped past their target dates.
A New Layer of the Grid Gets Put to Work
Perimeter is the third company this year to chase the idea of computing in places the grid already serves. Span, the smart-panel maker, is rolling out the XFRA Node — a home-sized edge data center paired with its panels and whole-home batteries — with a pilot planned in 100 newly built homes. Sunrun, the residential solar and storage giant, announced its own pilot last week, installing nodes packed with Nvidia chips in homes that already run on its systems. The home versions install far fewer chips per site than Perimeter's office deployments, but the logic is the same: use capacity that already exists instead of building new poles, wires, and substations.
The numbers behind that logic are stark. Arch Rao, Span's founder and CEO, argues the distribution network runs at only 40% to 45% utilization nominally. Building a 100 MW data center takes three to five years and costs upward of $15 million per MW, he said. Span's answer is to spread the same compute across 8,000 homes with XFRA Nodes — roughly six months of work at about $3 million per MW.
Sunrun's Bet on a GPU in Every Garage
Sunrun frames the same math in terms of generation. "There's aging poles and wires and ever increasing complexities and timelines to build a new generation," Paul Dickson, the company's president and chief revenue officer, told Latitude Media. A one-gigawatt nuclear plant can take a decade or more to build, he noted; Sunrun builds the equivalent amount of distributed capacity every year across its base of more than a million residential customers.
"Energy is the main bottleneck to compute," Dickson said. "So rather than trying to put 100,000 GPUs in one big building, let's put one GPU in someone's house and essentially build a distributed data center."
The pilot homes sit on Sunrun's solar and storage systems, sized so the panels cover household use plus the extra draw of the node. Sunrun pays the hosting fee out of what it earns from AI compute customers, and homeowners keep their existing solar contracts intact. Dickson said the pilot is meant to answer three questions: how the hardware holds up in a home, what compensation actually entices people, and which AI workloads belong at the edge at all.
The company sees the program as the compute side of a bigger energy play. It recently teamed up with Renew Home and Tesla to market a combined 16 GW of virtual power plant capacity to data centers — power those centers can tap during grid stress instead of cutting workloads. "We're approaching the energy crisis from two angles," Dickson said. "With Renew and Tesla, we're sitting on more than 16 gigawatts of power that data centers can tap into. We'll also have GPUs in homes to produce the compute ourselves."
Why Inference Fits the Edge
The workloads these distributed clusters target are inference, not training. Running a chatbot or a medical imaging model needs far less energy than building an LLM from scratch, and it is where AI companies actually make money. Put the GPUs close to users in dense cities and you also cut latency — a premium hyperscalers and AI labs will pay for, Hall said.
The signal is already there. "We are getting great signals not just from the OpenAI's and Anthropics of the world, but also individual tenants of these buildings that are in the financial services, robotics, and healthcare industries," he said.
The Local Opposition Problem
Power isn't the only obstacle slowing big campuses — communities are pushing back too. Sightline's analysts describe community resistance as a genuine driver of attrition in the development pipeline. A proposed $1 billion data center in Michigan, which a local official linked to Meta, was withdrawn in December after months of public opposition over water use, grid strain, and environmental impacts; the township followed with a six-month moratorium on new proposals. Similar moratoriums have been floated in at least ten states, including Louisiana, New York, Ohio, and Virginia.
Distributed computing sidesteps that fight entirely. A GPU node in an office basement or a home utility closet needs no rezoning, no new transmission line, and no public hearing. That's part of why the model is attracting notice from hyperscalers even as it remains unproven at scale. It is also why grid operators are watching closely: developers often announce projects in parallel and test which ones clear local hurdles first, which has flooded interconnection queues with speculative requests that may never materialize.
The Hard Parts Nobody Has Solved Yet
Perimeter does not have a signed contract yet. It is still raising capital — equipment runs about $35 million per MW — and it is negotiating with utilities about the load its servers will add. Urban distribution grids are often oversubscribed: a substation rated for 80 MW may have approved 90 MW of connections because no one runs everything at full tilt at once. GPUs run flat out all day, which can break that careful math. Hall says the company wants protections in place where utilities have oversubscribed.
Building owners carry no compute cost, and the arrangement gives landlords a new revenue line from floors that have been emptying out. Whether the contracts pencil out at scale is still an open question, and the company concedes it is early days. But the broader direction is clear: as the queue for grid connections stretches and opposition to big campuses grows, the industry is looking for compute in places it never considered — office basements, home garages, anywhere with an underused circuit. It won't replace the megacampuses that train the models, but for the day-to-day work of running them, the edge is getting serious. That is a shift worth watching for anyone tracking Cloud & Edge Computing, and it's a theme we've covered before in EU Opens Bidding for Seven AI Gigafactories in €30 Billion Push.
For more detail on the company's plans, see Latitude Media's report on Perimeter Compute, Span's home inference-node rollout, Sunrun's edge computing pilot, and Sightline Climate's data center delay research.