NVIDIA Backs a Plan to Run Inference From Residential Homes at Gigawatt Scale

NVIDIA Backs a Plan to Run Inference From Residential Homes at Gigawatt Scale

The Grid Has Spare Capacity. Nobody Was Using It.

The numbers behind the U.S. power grid tell a counterintuitive story. American data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of national consumption, and analysts expect that share to exceed 9% by 2030. At the same time, the existing distribution network operates at roughly 40 to 45% utilization on average, meaning more than half its capacity sits idle most of the time. Building new infrastructure to close the gap takes years. Some projects already in development have spent years waiting for interconnection approval, and a 100-megawatt data center typically runs upward of $15 million per megawatt to build, with a three-to-five-year construction timeline.

SPAN, which makes smart electrical panels for homes and utilities, has spent several years developing software that monitors a home's real-time electricity use and identifies how much headroom exists between current consumption and the panel's rated capacity. That same intelligence, the company concluded, could schedule AI compute workloads into the unused portion of the circuit. XFRA is the product that comes from that reasoning: an outdoor compute node paired with a SPAN smart panel and a whole-home battery, installed in residential and small commercial buildings, orchestrated by software that routes inference jobs across the distributed fleet based on available power and latency requirements.

The cost comparison is significant. Reaching 100 megawatts using XFRA would require installation in roughly 8,000 homes over about six months, at approximately $3 million per megawatt, compared to $15 million per megawatt for a conventional data center build. A pilot deployment is expected to begin later this year across 100 newly constructed homes, representing about 1.25 megawatts of compute capacity and 1,600 liquid-cooled inference GPUs. SPAN's target is gigawatt-scale capacity by 2027.

What NVIDIA Brings to the Partnership

NVIDIA's involvement goes beyond brand endorsement. The RTX PRO 6000 Blackwell Server Edition GPU that SPAN selected for XFRA is a recently released chip with capabilities specifically oriented toward distributed inference workloads. It carries 96GB of GDDR7 memory, fifth-generation Tensor Cores with support for FP4 precision, and can be partitioned into up to four isolated instances, allowing multiple workloads to run concurrently on a single card. Benchmarks from NVIDIA and third-party cloud providers indicate it delivers up to five times the LLM inference throughput of the previous-generation L40S chip, with roughly twice the price-performance of an H100 system for inference tasks.

Critically, the GPU is passively cooled and rated for 24/7 operation, which matters when the hardware lives in a residential setting rather than a purpose-built machine room. The liquid-cooled Server Edition variant that XFRA uses addresses the thermal management problem that would otherwise make residential compute impractical.

Marc Spieler, NVIDIA's Senior Managing Director of Global Energy Industry, framed the partnership in terms of latency and scale. Inference workloads increasingly need to run close to end users, and the traditional data center model struggles to deliver that proximity at the speed the market requires. NVIDIA has been building a broader strategy around exactly this thesis. Separately from XFRA, the company is working with T-Mobile to explore edge AI applications using the same RTX PRO 6000 Blackwell GPUs at distributed network locations. Akamai has deployed thousands of the chips across more than 4,400 edge locations worldwide to run inference closer to users. SPAN's approach takes the same hardware and pushes the deployment point further out — all the way into individual homes.

Who Else Is Building Distributed Compute

SPAN is not the first company to pursue distributed compute as an answer to centralized infrastructure constraints, though its residential grid-integration angle is distinct. The broader landscape splits into two categories of competitors, and neither maps exactly onto what XFRA is doing.

The first is decentralized compute networks, often called DePIN (Decentralized Physical Infrastructure Networks), which aggregate idle GPU capacity from data centers, crypto miners, and individual hardware owners into pooled marketplaces. Akash Network operates as an open-source cloud marketplace where providers list compute and buyers bid for workloads, typically at 60 to 75% lower cost than AWS or Google Cloud. Render Network, originally a GPU rendering marketplace for 3D artists, launched a dedicated AI compute subnet called Dispersed in late 2025. io.net aggregates GPU clusters from data centers and consumer hardware. Gensyn focuses specifically on distributed machine learning training using cryptographic proof mechanisms to verify compute contributions.

These networks are real and growing. Aethir reported roughly $127 million in 2025 revenue from enterprise AI and gaming clients. Render processed more than 22 million compute frames in 2025. But most DePIN networks face a persistent challenge: the hardware they aggregate is heterogeneous, uptime is variable, and the networks are better suited for batch or asynchronous workloads than for the low-latency, reliable inference that enterprise AI applications require. Token-based incentive models also add coordination complexity that enterprise buyers tend to avoid.

SPAN's approach differs in two meaningful ways. First, XFRA uses uniform, enterprise-grade hardware — the same liquid-cooled NVIDIA GPUs a hyperscaler would install in a purpose-built facility — rather than aggregating whatever happens to be available. Second, SPAN's orchestration software controls power at the circuit level through its smart panel integration, giving it scheduling capability that a pure software marketplace cannot replicate. The company is not asking homeowners to contribute idle gaming rigs. It is installing standardized nodes and managing them like distributed utility assets.

The second category is edge cloud providers. Akamai is deploying Blackwell GPUs at network edge locations, and telecom operators are testing edge AI at cell towers and switching offices. These deployments achieve geographic distribution, but they rely on existing carrier infrastructure rather than residential circuits. SPAN's footprint, if it scales as projected, would be more distributed and closer to end users than anything a carrier-based edge network can achieve.

The Homeowner Side of the Deal

XFRA only works if homeowners participate, and SPAN has structured the incentives accordingly. A household that agrees to host an XFRA node receives a SPAN smart panel, whole-home battery backup, optional solar, and fixed discounted rates on electricity and internet. The compute node itself is designed to be self-contained and quiet, installed outdoors. From the homeowner's perspective, the arrangement is essentially a hardware subsidy in exchange for consenting to have their panel's spare capacity managed by SPAN's software.

PulteGroup, one of the largest homebuilders in the United States, is integrating XFRA into new home construction from the start. The builder's VP of Strategic Sourcing described the model as reducing build costs while delivering homes with lower operating expenses for buyers. Building XFRA into construction from the foundation up is more practical than retrofitting existing homes and allows SPAN to aggregate capacity from entire new developments rather than recruiting individual households one by one.

The utility side also benefits, though indirectly. By scheduling high-draw compute workloads during off-peak periods and optimizing load across residential circuits, XFRA can reduce demand spikes and defer capital expenditures on grid upgrades. This positions SPAN's technology as an asset to utility operators, not just a cost burden — a relationship that matters for the regulatory environment the company will need to navigate as it scales.

The Practical Constraints

The XFRA model has real advantages on cost and deployment speed, but it also carries constraints that centralized facilities do not. Residential settings introduce variability in uptime, connectivity, and environmental conditions that orchestration software cannot fully eliminate. Latency across a residential network differs from latency inside a purpose-built data center, and workloads requiring tight coordination across many GPUs may not perform well when those GPUs are spread across thousands of homes.

SPAN's stated focus on inference rather than training is a deliberate choice that addresses some of these limitations. Inference workloads — running a trained model to generate outputs — are generally more tolerant of distribution than training, which requires tight coordination across large GPU clusters. Inference also accounts for the fastest-growing segment of AI compute demand. By 2030, it is expected to represent more than half of all AI workloads. XFRA is positioned for that growth curve, not for training infrastructure, which remains concentrated in large facilities where coordination overhead can be tightly controlled.

The pilot deployment this year, at 100 homes and 1.25 megawatts, will be the first real test of whether the orchestration software performs as described before the company attempts the much larger scaling operation it has outlined for 2027.

A Different Way to Think About Grid Infrastructure

The fundamental argument SPAN is making is that the distribution grid is an underutilized asset, and that underutilization represents a financing opportunity for whoever can turn spare capacity into billable compute. The grid was not designed for this use, but it was also not designed against it. Electrical headroom is a real and measurable quantity in every home with a SPAN panel, and SPAN's software can observe it in real time.

What makes XFRA notable is not that it is doing something technically unprecedented — distributed compute is a crowded space — but that it uses power infrastructure as the organizing principle rather than software marketplaces or carrier networks. The smart panel is the entry point, the battery provides buffer against fluctuation, and the NVIDIA hardware provides the compute quality that enterprise buyers require. The combination allows SPAN to offer hyperscalers something the DePIN networks generally cannot: predictable, enterprise-grade inference capacity at residential scale, without building new transmission infrastructure or waiting for grid interconnection approval.

Whether that combination performs as promised at gigawatt scale is a question the next 18 months will begin to answer.