A new kind of customer showed up at the grid

For most of the past two decades, the electricity grid was a slow-growing, predictable business. Data centers were an important customer, but a steady one, drawing power for websites, corporate software, and streaming video. Artificial intelligence changed that math almost overnight. Training a single large language model can consume as much electricity as a small town uses in a year, and the facilities built to run those models now rank among the largest single consumers of power on Earth.

That mismatch between demand and supply is the entire story behind the current wave of AI data center investing. The computers are ready today. The grid, the water, and the land are not.

What an AI data center actually is

At its most basic, an AI data center is a warehouse full of servers packed with graphics processing units, the chips that do the heavy lifting of training and running models. The difference from a traditional data center comes down to density. A conventional server rack might draw 5 to 15 kilowatts. A rack full of AI servers drawing 100 kilowatts or more is now routine, and that density compounds into campuses measured in hundreds of megawatts.

That power has to arrive somewhere, and the chips have to stay cool. Both requirements strain the same infrastructure. Many facilities rely on evaporative cooling that consumes millions of gallons of water a day, and they need transmission lines and substations that utilities cannot build quickly.

The bottleneck nobody can code around

Power is the binding constraint. A single large data center campus can draw as much power as a mid-sized city, and it needs that power every hour of every day, not just when the sun shines or the wind blows. Utilities take years to permit and construct new generation and transmission, which means demand is racing ahead of supply in the markets where AI is concentrated.

This is exactly the pressure point the Angel Publishing Homestacks promotion leans on. Its central claim is that giant centralized data centers are hitting power, water, and heat limits, so some of that compute will move to a distributed appliance installed on new homes. The pitch calls these units Homestacks, and it frames them as a coming replacement for the big campus model.

The real product behind the name

The actual hardware behind the Homestacks idea is not called a Homestack at all. It is SPAN’s XFRA, an outdoor unit roughly the size of an air-conditioning compressor. Each unit pairs NVIDIA Blackwell GPUs, well over $150,000 of equipment per box, with a smart electrical panel and a home battery. SPAN covers the host homeowner’s power and internet bills and installs the battery and panel at no charge. The first pilot is about 100 units in build-to-rent communities in Arizona and Nevada, built with homebuilder PulteGroup.

It is a genuinely interesting idea, and SPAN’s own press release is careful about what it claims. The company says XFRA is meant to augment centralized data centers, not replace them, which is a far more modest statement than the promo’s headline. NVIDIA agreed to sell SPAN the Blackwell chips. It did not invest in the company and did not discount the hardware.

The three stocks behind the pitch

The promotion points to three real, liquid, exchange-traded companies. nVent Electric (NVT) makes the enclosures and housings that protect electrical and networking equipment, and data centers now account for roughly 40% of its revenue. Generac Holdings (GNRC) is the home energy name, selling batteries, software, and grid controls, with a data-center backup-power backlog of about $700 million. Vistra Corp. (VST) is the Texas power giant with purchase agreements with Amazon and Meta and a long-term deal for half of its Comanche Peak nuclear plant.

All three are real data-center-adjacent businesses at sane valuations. What they are not is a bet on a 100-unit housing pilot. The honest way to read the promo is as a data-center buildout thesis, where the Homestack is the headline and the fundamentals are the substance. That is the same theme we walked through in our explainer on the Frontier AI data center buildout and our look at the data center stocks behind the AI Black Paper pitch.

Why the campus model is not going away

For all the appeal of a distributed appliance, the economics of AI compute still favor the big campus. The most demanding workloads, the frontier model training runs, need thousands of chips wired together at close range, because the speed of communication between chips is what determines how fast a model trains. That tight coupling is far easier to achieve inside one building than across a neighborhood of homes.

Cooling tells the same story. A dense rack of AI servers throws off an enormous amount of heat, and managing it at scale is a specialized industrial problem. Data center operators have built liquid cooling systems and water treatment plants designed specifically for that job. A home appliance the size of an air-conditioning compressor cannot easily match that, which is why SPAN’s own press release describes XFRA as a way to augment centralized data centers rather than replace them.

None of this makes the Homestacks idea worthless. It points at a real constraint, the power and water limits on big facilities, and it offers one partial answer. But it is best understood as a niche, a way to soak up smaller, less latency-sensitive workloads, rather than a replacement for the data center empire the promo’s headline describes.

Ready to see the research? Click here to access the Angel Publishing report.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.