The compute that never sleeps
Most of what people think of as artificial intelligence is the finished product, a chatbot answering a question or an image generator drawing a picture. But before any of that works, the model has to be trained, and training is the hungry part of the business. It is the phase where millions of dollars of graphics processors run flat out for weeks or months, chewing through electricity and data to teach a network how to do its job.
That phase is the reason AI uses so much power, and it is the quiet engine behind the entire data-center buildout. When a newsletter promotion talks about AI data centers, it is really talking about the buildings where this training happens.
What training actually involves
Training a modern large language model means running huge batches of text through a network of interconnected chips, adjusting billions of weights a tiny amount at a time until the model starts producing useful answers. The work does not lend itself to small setups. It needs thousands of chips talking to each other at high speed, which is why it happens in purpose-built facilities rather than in a closet.
The hardware that does this work is dominated by NVIDIA’s graphics processing units, including the Blackwell chips that the Homestacks story is built around. Our breakdown of the AI infrastructure behind the Accelerated AI pitch walks through how those chips move from factory to data center.
Why training eats the grid
A training run is relentless. Unlike serving a model to users, which rises and falls with traffic, a training run runs at near-peak load for as long as it takes. That constant draw is what stresses power grids, and it is why data center operators have become the grid’s most demanding customer.
The Angel Publishing Homestacks promotion hangs its argument on this exact pressure. Its claim is that centralized data centers are hitting power, water, and heat limits, and that some AI compute will move to a distributed appliance installed on new homes. The real hardware behind that idea is SPAN’s XFRA, an outdoor unit about the size of an air-conditioning compressor that pairs NVIDIA Blackwell GPUs with a smart electrical panel and a home battery.
What the pilot actually is
SPAN’s first pilot is small by design. About 100 units are going into build-to-rent communities in Arizona and Nevada with homebuilder PulteGroup. SPAN covers the host homeowner’s power and internet bills and installs the battery and panel at no charge. The company has raised roughly $500 million and took a $75 million strategic investment from Eaton.
The important detail for investors is what SPAN says about itself. Its press release describes XFRA as a way to augment centralized data centers, not replace them. NVIDIA agreed to sell SPAN the Blackwell chips. It did not invest and did not offer a discount.
The three names riding the theme
The promo directs readers toward three real, exchange-traded companies. nVent Electric (NVT) builds the enclosures that house electrical and networking gear, with data centers near 40% of revenue. Generac Holdings (GNRC) brings the home-energy angle through batteries, software, and grid controls, plus a roughly $700 million data-center backup-power backlog. Vistra Corp. (VST) supplies the electricity itself as a Texas power producer with purchase agreements with Amazon and Meta.
Each of these businesses is real and reasonably valued, but none of them depends on a 100-unit housing pilot. The honest read is that training demand is the durable trend, and these names are a way to own the buildout around it. We covered that demand story in our explainer on AI data centers.
The demand that does not slow down
The most important number in the AI story is not any single stock price. It is the forward order book for the chips that do the training. Hyperscalers have been signaling for quarters that they intend to keep spending heavily on AI infrastructure, because the competitive pressure to build bigger, smarter models has not let up. That spending is the demand that fills data centers, and it is the reason the buildout keeps growing.
This is where the training angle matters for investors. Training is the compute-hungry phase, and it is the phase the biggest companies cannot afford to pause. If a company stops training, it falls behind its rivals, so the spending is closer to a fixed cost of staying in the race than a discretionary item. That is why analysts expect the data center buildout to keep running for years, and it is the durable trend behind every stock in the Homestacks pitch.
For a fuller view of how that chip demand becomes infrastructure, our breakdown of the AI infrastructure behind the Accelerated AI pitch traces the same spending through the supply chain.
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.