Put a data center in your home! SPAN aims to build a decentralized AI computing network with 80,000 nodes

San Francisco startup SPAN has announced the launch of its "distributed data center" solution, XFRA, which plans to install liquid-cooled nodes equipped with Nvidia RTX Pro 6000 Blackwell GPUs next to residential homes in the U.S., offering subsidized electricity and internet as incentives. (Context: GPU computing shortage repeats: OpenAI, Anthropic, and other major players consume supply, leaving AI startups waiting until the end of the year) (Background: Quitting a multi-million dollar option package to resign, Riot executive's departure reveals the pain points of mining companies pivoting to AI computing power) Data centers don't necessarily have to be built on barren land in the suburbs? San Francisco startup SPAN has provided another answer: break them down into tens of thousands of small boxes and install them in the yards or driveways of U.S. homes. This solution, called XFRA, is essentially a liquid-cooled edge compute node with a built-in Nvidia RTX Pro 6000 Blackwell Server Edition GPU. Homeowners who install it can receive subsidized electricity, high-speed internet, and a backup battery pack. SPAN has completed initial pilots and expects to launch a 100-home trial this year. The construction logic for traditional large-scale data centers is clear: the larger the scale, the lower the unit cost of computing power. However, this logic is hitting several hard walls. First is land. A 100 MW data center requires dozens of hectares of land and must be located near a stable power source. In many U.S. states, there has been community opposition to data centers due to concerns over noise, rising electricity costs, and excessive water consumption. Second is water. Data centers generally use evaporative cooling systems, and a medium-sized facility can consume millions of liters of water per day, which is particularly controversial in drought-prone areas. Finally, there is time. From site selection and permitting to infrastructure construction, a new data center often takes three to five years to come online, while the demand for AI computing power cannot wait that long. SPAN's solution attempts to bypass these three walls. XFRA nodes are installed next to residences, requiring no independent land; liquid cooling is more efficient than air cooling and does not rely on large amounts of water; nodes can be deployed in sync with residential construction, and the expansion speed is theoretically much faster than traditional data centers. Numbers are the most powerful argument for SPAN's solution. According to the company's interview with CNBC, the cost of deploying 8,000 XFRA nodes is only one-fifth the cost of building a 100 MW traditional data center that provides equivalent computing power. The node itself is equipped with the Nvidia RTX Pro 6000 Blackwell, a professional GPU designed by Nvidia for server workloads, supporting large-scale parallel computing suitable for AI inference tasks. The liquid-cooled design keeps noise within a range acceptable for residential areas. SPAN's expansion plan is quite aggressive: starting in 2027, it aims to scale XFRA nodes to 80,000, establishing a distributed computing network of over 1 GW within the U.S. What is the concept of 1 GW? It is roughly equivalent to the output power of a medium-sized nuclear power plant. However, SPAN also clearly defines the boundaries of this network. XFRA targets AI inference, cloud gaming, and content streaming, rather than AI model training. Training large language models requires thousands of high-end GPUs like H100 or B200 working in concert for long periods, which is the domain of hyperscale cloud providers like Google and Microsoft. XFRA fills the computing demand "after training": taking trained models and running them to respond to real-world requests. Of course, this model is not without challenges. The first is network stability. Nodes scattered in residential areas face consumer-grade network environments, where bandwidth and latency consistency are far inferior to the backbone networks of data centers. For some AI inference applications that require low latency, this is a clear technical bottleneck. The second is cybersecurity. Processing enterprise-level data on nodes next to homes involves physical security, data sovereignty, and compliance requirements that are more complex than those of traditional data centers. SPAN has not yet publicly explained how it will address these risks. The third is the sustainability of homeowner participation. Subsidized electricity and internet are attractive incentives, but if nodes malfunction, noise exceeds expectations, or subsidy conditions change, whether the exit mechanism is clear enough will directly affect the reliability of the entire network. The logical appeal of such distributed computing solutions is undeniable. But there is a long way to go from a pilot to 80,000 nodes. The war for computing power has never been just a technical issue; it is a three-way gamble involving deployment speed, capital structure