The Backyard Is Not the Frontier

Jun 20, 2026

A startup wants to turn homes into AI infrastructure. The pitch is decentralization. The reality is simpler: intelligence is still made where the power is.

There is a company called Span, best known for smart electrical panels, that has teamed up with Nvidia and PulteGroup to bolt a new kind of box onto the side of a house. Inside is a liquid-cooled server stuffed with Nvidia RTX PRO 6000 Blackwell Server Edition GPUs. The homeowner gets a premium smart panel, battery backup, and potentially discounted electricity and internet.

The catch is that the box is not really yours.

It is a node in someone else’s data center, and you are the landlord.

It is a real answer to a real bottleneck, but it is also a useful fiction about where intelligence actually comes from.

The constraint moved and nobody updated the story

For about a decade, the AI story was a chip story. Whoever had the most GPUs, or the best ones, won. That framing is now out of date, and most of the public conversation hasn’t caught up.

The binding constraint has quietly turned from silicon to electricity.

The industry even has a polite phrase for it: the “speed-to-power gap.” That is a tidy way of saying companies can now buy GPUs faster than utilities can provide the power to run them. Interconnection queues for large new loads can run for years. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity globally in 2024, around 1.5% of global electricity demand, and projects that figure could roughly double to about 945 terawatt-hours by 2030. In the United States, the growth is especially concentrated: IEA projects data center electricity consumption will rise by about 240 terawatt-hours by 2030, a 130% increase from 2024 levels.

That is the part people miss. Globally, the percentage can look modest. Locally, it can hit like a dropped refrigerator.

Communities are organizing against new facilities over water use, noise, transmission lines, and the simple fact that a warehouse full of servers makes a lousy neighbor.

So the real question stopped being, “Can we get the chips?”

It became, “Where on the grid can we possibly plug them in?”

Once you see that as the bottleneck, a lot of otherwise-baffling behavior snaps into focus.

What the bottleneck makes people do

The backyard node is one response.

If you cannot get a massive interconnection for a new data center campus on any timeline that matters, maybe you can scavenge smaller amounts of power from houses that already have grid connections, panels, meters, and homeowners who would not mind some help with the electric bill.

You route around the queue by hiding the data center inside a subdivision.

I am skeptical that this works well at residential scale. A large always-on compute load is not “spare capacity.” It is a serious electrical load. Put enough of them on one street and suddenly the neighborhood transformer, which was never designed to moonlight as an AI factory, becomes the weakest link.

The version that survives is probably commercial: mid-market sites with three-phase power, real fiber, a fence, clear accountability, and someone available when the expensive box starts making expensive noises.

But the motive is the tell.

Nobody does this because backyards are a good place to compute. They do it because backyards have power connections, and data centers cannot get enough of them fast enough.

The solar siting fights are the same story wearing a different costume.

About twenty miles from where I live, Silicon Ranch is building the Stockton Solar Farm in Baldwin County, Alabama, on land that until recently was timber country. The project is being framed as energy infrastructure, agriculture, and conservation, and to be fair, that is better than pretending land use is simple. But it still points to the same uncomfortable truth.

Developers across the country are clearing farmland and timberland for solar fields. Not because anyone forgot parking lots exist. Not because they overlooked the acres of dead asphalt sitting near every city.

They priced them and walked away.

Ground-mounted solar on open land is far cheaper. Build that same array elevated over a parking lot, engineered for wind, vehicle clearance, drainage, lighting, liability, and foundations drilled through pavement, and the economics get ugly fast.

Nobody overlooked the parking lots. The spreadsheet killed them.

That is why it often takes a law to make the better land use happen. France passed one in November 2022 requiring large parking lots to cover at least half their surface with solar panels. The French government estimated the rule could generate up to 11 gigawatts of power without clearing new land.

The spreadsheet that picks the timberland never has to account for the timberland.

Different schemes, same root: energy is the scarce thing, and everything bends around it.

Now the part that is actually about the models

This is where the story stops being a real-estate curiosity and starts being about foundation models themselves, because the power wall is not just a logistics problem. It is beginning to shape the kinds of AI systems that get built.

The load-bearing fact is that training and inference are different physical animals, and people who do not build this stuff tend to blur them together.

Training is the process that creates a frontier model. It needs thousands of GPUs packed close together, wired through high-bandwidth, low-latency interconnects, and moving through synchronized updates.

That does not mean distributed training is impossible.

It means frontier-scale training is not something you scatter across neighborhoods like birdseed. The communication overhead would eat the benefit. The frontier is still forged in centralized, power-bound, water-hungry mega campuses, and for now, it stays there.

Inference is different.

Inference is running a finished model to answer a prompt, classify an image, generate a video clip, summarize a document, or route some task through a tool. That part can move closer to the edge. Span says XFRA is designed for inference workloads, not to replace centralized data centers.

That distinction matters.

A backyard node can run useful work. It can run smaller models, specialized models, quantized models, and pieces of larger systems.

What it cannot do is make the next frontier model from scratch.

So the more inference gets pushed outward, the more pressure lands on building models that are small enough to fit, cheap enough to run, and specialized enough to be useful at that size.

That is the genuinely interesting part.

The power wall does not end the capability race. It changes the racecourse. Efficiency per watt becomes the competitive axis. Not because scaling laws stopped working, but because you cannot simply assume the next ten-x of electricity will show up on schedule with a bow on it.

When power becomes the scarce resource, the field reroutes into cleverness: distillation, quantization, mixture-of-experts with smaller active footprints, smaller models that think longer at inference time, retrieval, tools, memory, routing, caching, and all the unglamorous engineering that makes a system useful without making it enormous.

The capability race meets physics.

Physics does not negotiate.

So the race learns to turn corners.

The darker wrinkle

There is another problem hiding inside all this distributed enthusiasm.

When inference is scattered across thousands of heterogeneous nodes, the question “What am I actually talking to right now?” gets harder to answer.

Is it the full model? A smaller version? A quantized version? A routed response from another service? A specialized system wearing the same brand name? A degraded fallback because the expensive path was unavailable?

Distribution launders degradation.

The frontier lab can ship a pristine model. But by the time it is running on a smaller node, routed through layers of orchestration, compressed to fit the economics of the moment, and hidden behind a friendly product name, you may have no meaningful way to know how much of the original capability survived.

Edge inference is useful for latency.

It is much less wonderful for transparency.

The useful fiction

Which brings me to the thing being sold versus the thing being built.

The pitch for backyard nodes, and for distributed compute generally, is democratization.

Break the hyperscaler monopoly. Compute for everyone. Power to the people, literally.

It is a useful fiction.

Look at what the homeowner actually owns: nothing that matters.

Not the model weights.
Not the orchestration software.
Not the customer relationship.
Not the ability to train anything.
Not the economic upside of the intelligence being sold.

They own a slab of someone else’s infrastructure and maybe a discount on their power bill.

The weights stay central.
The control stays central.
The economics stay central.

And every one of those nodes still runs Nvidia silicon, which means the “decentralized” future may tighten Nvidia’s grip rather than loosen it.

This is not distributed power.

It is distributed hardware with centralized everything-that-counts.

So when someone tells you backyard data centers are going to democratize AI, the honest translation is less romantic:

A chip company, a smart panel company, and a homebuilder found a way to move part of the capacity problem into neighborhoods and dress it up as empowerment.

That is not evil. It is just what the incentives produce.

And it is worth calling a spade a spade.

Where this leaves us

None of this distributed cleverness changes where frontier intelligence is made.

Backyard nodes, parking-lot canopies, edge inference, smarter panels, and scattered compute fleets are all downstream of a training process that remains centralized, expensive, energy-hungry, and controlled by the few players who can secure the power to run it.

The distributed layer is a delivery system for intelligence manufactured somewhere else.

It may change how the model reaches you.

It does not change who makes the model, who controls the weights, or what it costs in watts to create.

That is the line I keep coming back to.

The backyard is a distribution channel.

The frontier is a power plant with a research lab attached.

And the most important AI policy questions of the next few years may not be about algorithms at all. They may be about electricity: who gets it, who pays for it, whose grid bends to absorb it, and whose land gets cleared to generate it.

The backyard is not the frontier.

But the people building the frontier would very much like you to host one.

Sources / Further Reading

Sources and Further Reading

This piece originated as a conversation between the author and Claude Opus 4.8. Editing and fact-checking by Jinx (Laura’s “Chaos Gremlin Agent”).

Laura Florey

Discussion about this post

Ready for more?