Perspectives

The Backyard Generator

"Our data can't leave the building." The fear is aimed at the wrong target, and the backyard generator usually costs more than the danger it prevents.

Gary
July 6, 2026
A lone backyard generator with one dim bulb, beside a bright, interconnected power grid.

Every serious enterprise deal hits the same moment. The room goes quiet, someone leans forward, and says a version of one sentence: our data can't leave the building. They mean it as a wall. Cloud AI sits on the wrong side of it, and no feature list climbs over.

I used to treat that wall as a hard no. I don't anymore. The fear behind it is aimed at the wrong target, and honoring it usually costs a company more than the danger it's meant to prevent.

Who actually guards your data

Start with who the fear is actually aimed at, and what that party wants. An AI vendor is a business, and its business only grows when its customers do. Usage is the product. Outcomes are the product. A vendor that leaks a client's data, or quietly feeds it to a competitor, has destroyed the one asset keeping it alive: the trust that using it is safe. That incentive isn't a line in a contract. It's structural. Betray one customer's data and every other customer reads about it the same week, and the company is finished. The people guarding the data have more to lose than the client does.

An office building behind a wall, with a data stream passing safely out through a gate to a connected network.

Think of it like the power grid. A company could run its own generator in the backyard. It sits on the owner's property, nobody else touches it. But the grid isn't run by people who want to steal anyone's appliances. It's run by people whose entire business is keeping the lights on, because the moment they can't, the customer leaves and they die. Their interest and the customer's are bolted together.

There's a second alignment most buyers miss. An AI vendor carries the same fear, one layer up. It doesn't want the underlying model trained on its customers' data either, because that data is its edge as much as the customer's. When it pushes the model provider for a no-training guarantee, it's protecting its customers by protecting itself. Customer, vendor, and model provider all pull the same rope the same direction.

Three parties connected on one line, all pointing the same way, showing aligned incentives.

You already made this trade

And most buyers accepted this trade years ago, without noticing. Their money sits at HSBC, on HSBC's servers, not in a safe under the desk. Their term sheets go over Gmail. Their teams run whole companies over WhatsApp. All of it is cloud. They trusted strangers with money and secrets because those strangers' business was to not lose them. The AI objection is the same objection, resolved a decade ago.

When the wall should stay up

Now the honest part, because the fear isn't entirely wrong. Some vendors are badly aligned. If a company's real revenue comes from ads or from reselling behavior, the data is the product, not the thing being guarded, and the wall should stay up. Some data carries legal walls that no incentive argument dissolves. And the dream of a model that runs fully offline, on hardware under the company's own roof, is a real dream that may someday change this math. I'm not certain it won't. I'm saying it hasn't yet.

A dimmer bulb and a second job

Which is the whole problem with the generator. Run everything offline today and the result is a dimmer bulb and a second job. The models a company can host itself are weaker than the ones on the grid, often by a full generation. Standing them up costs real money and real months. Then the company owns the upkeep forever, patching and tuning and buying GPUs instead of building the actual business. The point of buying AI was an edge. A hobbled private model hands that edge straight back to the competitor who plugged into the grid and spent those months on customers instead of maintenance.

A single dim bulb on a private generator next to a row of bright bulbs on the grid.

So the decision isn't the one people think they're making. It isn't safe versus risky. It's a choice about whose incentives should hold the data, and how much capability a company will burn to feel in control. Feeling in control is not free. For most companies it costs the exact advantage they were chasing.

Vet the alignment, not the location

Doing this well means vetting the alignment instead of the location. Read the no-training terms and make them contractual. Ask who actually pays the vendor, because that answer reveals what it protects. Check whether the vendor's business grows when the customer's does, or in spite of it. Get those three right and the walls of the building stop being the point.

Transmission towers resolving into a connected network of nodes, the electrical grid as shared infrastructure.

Keep the generator for the one room that truly needs it. Put the rest on the grid, and go build.