The Last Mile Has a Power Bill

The hardest problem in Physical AI at the edge isn’t the intelligence. It’s the watts - and almost no one designs for it.

Most conversations about Physical AI begin with the intelligence: the models, the silicon, the autonomy. Spend a few days where this technology actually gets deployed - a solar farm, a construction site, a remote yard - and you learn the conversation usually should have started somewhere far more ordinary. Where does the power come from?

At an off-grid site there is no wall to plug into. So you bring your own power: solar panels and batteries, usually on a portable trailer. By the time that trailer carries the radios, the compute, the cameras, and the storage a real deployment needs, it can cost more than a hundred thousand dollars. And its reliability is only as good as the sky. String together a few overcast days and even a well-sized battery bank runs out of headroom.

Then comes the part nobody designs for.

Enterprise gear was built for a wall socket. Standard cameras, switches, and appliances were engineered in a world of effectively unlimited grid power, and they carry that assumption everywhere they go. Drop them onto a battery and the assumption quietly breaks. The sharpest example surprises every team the first time they see it: surveillance cameras draw the most power at night. To see in the dark they fire infrared illuminators, and those illuminators are hungry. So your peak demand arrives in the middle of the night - exactly when your solar array is producing nothing at all. Peak load meets zero supply, and the battery has to carry the entire night, every night.

Now put edge AI in that same trailer. The whole point of Physical AI is to run inference on-site - vision, control, decisions made in real time - because the cloud is too far away and too slow. But GPU inference is power-hungry too, and it lives off the same battery as everything else. If the compute isn’t co-designed for the energy envelope, it competes for watts with the very cameras and sensors it exists to watch. The site browns out. The “autonomous” operation goes dark at the worst possible moment.

The lesson the field teaches, over and over, is that you cannot bolt best-in-class components together and call it an edge AI system. The camera, the radio, the compute, and the power source are not four products that happen to share a trailer. They are one coupled system, and the coupling is the design. Get the power model wrong and the best model in the world never gets to run.

This is the part that doesn’t show up in a demo. In a lab, power is infinite and free. In the field it is finite, variable, and the first thing to run out. So the real engineering question at the edge isn’t “how smart is the model.” It’s “how many watts do I actually have at three in the afternoon in August, and what happens to them at two in the morning in December.”

This is the layer we build at Ramen. We model the full energy budget of a site - what the panels generate, what the batteries hold, and the real, time-of-day demand of every device on it - and manage capture and inference to live inside that envelope: prioritizing workloads, scheduling compute for when the power is there, and choosing inference platforms tuned for low-power operation, so the site keeps running through the night and through a run of cloudy days. The intelligence on top only pays off if the infrastructure underneath survives the real world.

Physical AI at the edge is, in the end, an energy-systems problem as much as an AI problem. The teams that win in the field will be the ones who treated power as the first design decision, not the last.

The last mile has a power bill. Most people find that out the hard way.

‍

^{Field Notes from the Last Mile is a running series on what Physical AI actually takes to deploy in the real world - from the people building it. Subscribe to get the next one.}

‍