Data center developers are struggling to connect to the power grid and, not unrelatedly, connect with people.
Perhaps half the data center projects due to start operating this year won’t arrive on time, according to Currence, an artificial intelligence analytics firm. Trouble accessing electricity is a big reason, but so is resistance. Concerns about rising power bills, along with land and water usage and AI’s threat to jobs, have more than 70% of Americans opposed to data center construction in their area. Some 75 projects worth an estimated $130 billion were blocked or delayed in the first quarter alone, according to Data Center Watch.
One start-up, Span.io Inc., is proposing a solution: If people are worried about giant data centers being near their homes, why not just co-opt those homes? Span’s model exchanges backyard access for discounted power and internet services, a way for AI to reward, rather than burden, households. An ingenious idea, but also a potent sign of how a technology sector used to the frictionless scaling of software is confronting the physical complications of energy.
Hyperscalers clearly have the will, and the financing, to build giant data centers, but they can’t easily get around yearslong grid-interconnection queues and equipment backlogs. Out of more than 80 gigawatts of new data center capacity planned for 2027 and 2028, analysts at Jefferies Financial Group Inc. think 30-40 gigawatts would be a positive outcome.
One problem is that while giant single-site data center complexes are the most efficient means for training leading-edge AI models, they also present a huge strain for power grids. A 1 gigawatt data center running 90% of the time draws more electricity than the entire state of Rhode Island. By 2030, training runs for the most advanced AI models could draw between 4 and 16 gigawatts, according to Epoch AI, a research nonprofit.An alternative is to distribute the training load across multiple sites, linked by high-bandwidth fiber. While this adds complexity, it takes advantage of different levels of power availability in different areas, without overwhelming local utilities. Microsoft Corp. has already launched such a network at scale in Wisconsin and Georgia.
Now Span wants to take that down to the micro level with its XFRA nodes. These units, housing Nvidia Corp. chips and a cooling system, attach to a house and work in conjunction with Span’s core product, a smart electrical panel that manages a home’s electrical loads to expand its usable capacity, plus a home battery. A network of 8,000 nodes would equate to a 100 megawatt data center. Span says the homeowner gets a smart electrical panel plus power and internet service at a flat, discounted fee, while the company sells computational capacity to AI developers.
Like Span’s panels, these nodes leverage untapped resources hiding in plain sight: homes already connected to the grid but which rarely use anything like their entire electrical capacity. Installing nodes in subdivisions should take months rather than years. There are drawbacks, not least relying on slower internet connections. Still, Epoch AI concludes this is not an insurmountable barrier for distributed compute. Such networks’ proximity to end-users should lend itself well to inference applications, which generate output from trained models.
Whether suburbia starts sprouting mini-data centers remains to be seen. Regardless, the concept itself says much about AI’s energy predicament.
Hyperscalers’ want power faster than utilities can respond. The result has been workarounds.
Span’s approach fits a wider pivot to distributed systems. Hybrid data centers, which use onsite generation to reduce their load on the grid, are growing fast and now represent 15% of planned capacity, according to Jefferies. At the extreme end, Space Exploration Technologies Inc.’s “Colossus” data center sites in Tennessee and Mississippi have used or are installing dozens of onsite natural gas turbines, sometimes without permits.
