Perimeter Compute, a new startup emerging from stealth, is repurposing the excess electrical infrastructure already built into commercial office buildings to deploy GPU clusters for artificial intelligence workloads. The company has identified more than 1 gigawatt of spare power capacity across office buildings in major U.S. cities, representing a novel solution to the acute power shortage constraining AI data center expansion.

Why Are Data Centers Running Out of Power?

The U.S. power grid is buckling under AI's explosive growth. An estimated 50% of planned U.S. data center construction projects are now delayed or cancelled specifically because of power shortages and grid infrastructure limitations. The root cause is staggering: interconnection queues for new power projects in the U.S. have swelled to over 2,100 gigawatts, a volume that exceeds the entire installed capacity of the nation's grid, creating multi-year delays for new data center connections.

Global data center electricity consumption reached 485 terawatt-hours in 2025, with AI-focused facilities accounting for a disproportionate and expanding share, and is projected to reach 950 terawatt-hours by 2030. This surge is driving data center power demand from 76 gigawatts in 2026 to 134 gigawatts by 2030.

How Can Office Buildings Solve the Power Problem?

Perimeter Compute's insight is deceptively simple: modern office buildings have been over-engineered for peak electrical loads that no longer materialize. After years of energy conservation efforts, many buildings have reduced their power consumption by 30% to 60%, yet the electrical infrastructure remains designed for the original peak load.

"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," said David Hall, CEO of Perimeter Compute.

David Hall, CEO at Perimeter Compute

The company plans to install the latest AI chips in the basements and mechanical rooms of Class A commercial office buildings with between 0.5 megawatts and 20 megawatts of spare capacity. Even half a megawatt of power is enough to serve 240 of NVIDIA's latest GPUs, which can process billions of user requests per day.

What Makes Edge AI Inference Valuable?

Perimeter is targeting the AI inference market, which is the day-to-day use of chatbots and other programs that requires significantly less energy than the training of large language models. This focus matters because GPUs placed in urban centers can process AI workloads faster than those that need to transmit data long distances, allowing companies to fetch premium prices from hyperscalers.

The actual application of AI is what generates revenue for companies like Amazon, Anthropic, Google, and OpenAI. So inference compute is in high demand. Hall noted that Perimeter is receiving strong interest not just from major AI labs but also from individual tenants in these buildings working in financial services, robotics, and healthcare.

Steps to Deploy Edge AI in Commercial Buildings

• Identify Excess Capacity: Survey Class A office buildings in major metropolitan areas to locate facilities with 0.5 to 20 megawatts of unused electrical infrastructure after energy conservation efforts.
• Coordinate with Utilities: Engage local utilities early to discuss expected increased loads and ensure grid stability, as some urban distribution grids may already be congested due to oversubscription.
• Structure Revenue Sharing: Establish contracts where the compute operator pays for GPU equipment, installation, and energy costs while sharing revenue with building owners and landlords.
• Install Submetering: Deploy submeter systems to track actual energy consumption and ensure accurate billing and revenue allocation between parties.

The financial model is straightforward: Perimeter will pay for the GPU equipment, the installation, and the extra energy the system uses, which a submeter will track. The company will also pay to lease space from landlords and share a portion of the revenue from selling compute with them.

What Are the Remaining Challenges?

Despite the promise, significant hurdles remain. Perimeter does not yet have signed contracts with either tech companies or building owners. The structure of these contracts, and whether they will ultimately be attractive enough to scale more broadly, remains unclear.

Grid stability is another concern. Some local distribution grids in urban hubs like New York and Boston may already be congested, in part because utilities "oversubscribe" connections. For example, a local substation might handle 80 megawatts of capacity at any given time, but a utility approved 90 megawatts across various homes and buildings because none run at 100% capacity at all hours. GPUs, by contrast, need power all the time, potentially upending a utility's carefully managed load.

The broader context underscores why this innovation matters. Hyperscalers are allocating unprecedented capital to AI infrastructure, with forecasts for Big Tech's 2026 AI infrastructure spending ranging from $500 billion to $690 billion, with consensus estimates centering around $630 billion. This massive outlay is creating a "capex-to-revenue gap," where the cost of building and operating AI systems is growing faster than the revenue they produce. Energy is a primary operational cost, making any solution that unlocks existing power infrastructure potentially transformative for the industry.