A startup backed by NVIDIA is placing enterprise-grade GPU (graphics processing unit) hardware on the outside of residential homes, fundamentally changing how data centers are built and powered. Span's XFRA nodes pack 16 NVIDIA Blackwell GPUs and consume as much electricity in three days as a typical US household uses in a month, yet the company claims this distributed approach solves a critical bottleneck in AI infrastructure expansion.

Why Are Data Centers Moving Into Neighborhoods?

The traditional path to building AI data centers has become a years-long bureaucratic nightmare. Substation upgrades for a standard 100-megawatt data center take four to seven years in most parts of the United States, according to Lawrence Berkeley National Laboratory. More than 2,060 gigawatts of energy capacity sat stuck in grid interconnection queues as of late 2025. This infrastructure bottleneck is strangling the AI industry's ability to scale.

Span's solution sidesteps the problem entirely. Instead of waiting years for new grid infrastructure, the company taps electrical capacity that already exists in residential neighborhoods. The average American home only uses about 40% of its electrical capacity at any given time, leaving roughly 80 amps of idle power in homes with standard 200-amp utility connections. Span captures that unused headroom by installing a liquid-cooled server unit outside your home, next to your air conditioning unit or electrical box.

The math is compelling. Span can deploy 8,000 XFRA units approximately six times faster and at roughly one-fifth the cost of a comparable centralized data center. The company is targeting one to two megawatts of compute by year-end, scaling to over 1 gigawatt annually from 2027.

What Hardware Are These Home Data Centers Packing?

Each XFRA node is a serious piece of computing equipment. Inside the compact, air-conditioner-sized unit sits 16 NVIDIA RTX Pro 6000 Blackwell Server Edition GPUs, four AMD EPYC CPUs, and 3 terabytes of memory in a liquid-cooled Dell server setup. According to Carnegie Mellon computer scientist Mahadev Satyanarayanan, "even a modest-sized large language model could run on a 16-GPU cluster," making this hardware capable of handling substantial AI workloads.

The economics are straightforward. Homeowners pay Span a flat monthly fee of around $150, and in return, Span covers their entire electricity and internet bills. Span then sells the computing power your node generates to AI companies, cloud gaming providers, and other high-performance computing customers.

How Is This Different From Traditional GPU-Based Data Centers?

While Span focuses on distributed residential deployment, the broader AI chip market is fragmenting in interesting ways. D-Matrix, a startup located just three miles from NVIDIA's Silicon Valley headquarters, has begun production of its Corsair inference chip, which claims to run inference workloads 10 times faster and using five times less energy than a standalone NVIDIA GPU, though with important caveats.

D-Matrix's approach relies on a different memory architecture than traditional GPUs. Like competitors Cerebras and Groq, D-Matrix uses SRAM (static random-access memory), a type of memory that can be integrated directly onto the same chip as the computing logic. This tight integration eliminates the latency penalty of moving data between separate memory and compute components. NVIDIA's GPUs, by contrast, rely on large amounts of DRAM (dynamic random-access memory) packaged into stacks of high-bandwidth memory added around the logic chip.

The trade-off is significant. SRAM-based designs excel at inference tasks where speed matters more than model size, but they cannot handle massive reasoning models with trillions of parameters. D-Matrix's Corsair is optimized for interactive AI applications like chatbots, voice agents, and agentic tools, not for training or running the largest language models.

"We're not running into a chokepoint around DRAM with our product because our product doesn't really rely on DRAM to be successful," said Sid Sheth, D-Matrix co-founder and CEO.

Sid Sheth, Co-founder and CEO at D-Matrix

D-Matrix began shipping Corsair chips to customers this month, with about 90% of them located in the United States and overseas customers in the Middle East and Southeast Asia. The company packages four Corsair chips together inside a card that slides into data center server racks, with each card costing tens of thousands of dollars.

What Are the Real-World Challenges to Home-Based Data Centers?

Span's fanless, liquid-cooled design addresses one major concern: noise. But the broader push to place AI infrastructure into residential areas has already triggered serious backlash in communities across the country. In Dowagiac, Michigan, residents near the Hyperscale data center filed a class-action lawsuit in federal court, reporting noise readings as high as 84 decibels affecting more than 1,300 homes within a one-mile radius. One resident described the sound as "like having a vacuum cleaner running all the time." In New Jersey, two neighbors sued DataOne USA over an industrial hum that never stops. Virginia residents near Google's "Mango Farm" complex describe the sound as "an internal organ vibration" that triggers anxiety attacks.

The Environmental and Energy Study Institute points out that most local noise ordinances cover noisy parties, not 24/7 industrial facilities running indefinitely. While XFRA's fanless design addresses this directly, skeptics note the 100-home trial is small. Scaling to thousands of units across diverse neighborhoods will surface new challenges.

Beyond noise, the XFRA setup raises data privacy questions. Span manages both your electrical connection and internet service, meaning your internet traffic routes through Span's network. That arrangement could potentially allow Span to observe your browsing patterns, drawing comparisons to public Wi-Fi risks. Span has not yet published detailed data-handling policies for this arrangement.

How to Evaluate Whether a Home Data Center Makes Sense

• Financial Trade-off: You receive roughly $150 per month in utility coverage (electricity and internet) in exchange for hosting enterprise-grade GPU hardware on your property. Calculate whether this covers your typical monthly bills and whether the arrangement makes financial sense for your household.
• Noise and Disruption: Although Span claims XFRA units are fanless and run quietly, verify the actual noise levels in prototype deployments before committing. Research whether your local noise ordinances would protect you if the unit becomes disruptive.
• Data Privacy Concerns: Review Span's data-handling policies carefully before signing up. Understand what internet traffic passes through Span's network and what data the company collects about your household's usage patterns.
• Physical Security: Data centers carry significant physical security for good reason, given the value of the hardware inside. Consider whether theft or tampering risks are acceptable for your neighborhood and property.
• Long-term Commitment: Understand the contract terms, cancellation policies, and how long Span plans to operate the unit on your property before agreeing to host one.

The broader infrastructure play is clear. Span founder and CEO Arch Rao told CNBC that "fundamentally, it's an infrastructure play. We're uniquely positioned to build infrastructure that can simultaneously help us meet what is clearly an insatiable demand for more compute, much more cost-effectively".

Arch Rao

Meanwhile, AMD is positioning itself as a broad-portfolio alternative to NVIDIA, offering CPU, GPU, and adaptive computing solutions with emphasis on power efficiency. AMD claims its data center CPU and GPU offerings deliver leadership performance-per-watt, meaning it can take less space and power utilization to achieve the same results as competing systems.

The 100-home trial for XFRA units rolls out later in 2026, with prototype deployments already live in Northern California. Whether neighborhoods will accept distributed compute nodes at scale remains an open question, but the economic incentive is undeniable. As grid interconnection queues continue to grow and AI demand accelerates, the pressure to decentralize data center infrastructure will only intensify.