A business model should follow the value delivered to the customer. Not where the engineering happened to run.

That's my whole argument.

I've been sitting with the local-vs-cloud question for a while, mostly through the lens of home security. And the technical story is genuinely exciting. Small models are good now. Person, package, vehicle that runs on the edge today. Face recognition on the edge is no longer a research demo; the accuracy is real. The next wave is large models, LLMs, and VLMs, and for those, an edge AI box starts to look like a clean answer. Compute cost drops. Privacy concerns mostly go away because the footage never leaves the house.

I'm not here to argue against any of that. I'm open to cloud-plus-edge. I'm open to edge-only. If the customer sees value, and on AI cost and privacy, they clearly can, then the architecture should follow the value. That part isn't controversial to me.

Here's my argument.

There's a mindset I keep running into: once the compute and storage live on the device, the subscription should go away. The customer bought the box or camera or any devices with local AI or storage capicity, so now the service should be "free." Buy the hardware, own it forever, no monthly fee, no SaaS. And it's usually framed as the pro-customer position, as if charging a subscription on top of hardware the customer already owns is somehow a trick.

I think that's mixing two different problems into one.

  1. Where does the work run? Cloud, edge, or some split. That's an engineering and economics decision. Token cost, latency, bandwidth, privacy, unit cost, and who eats the capex.

  2. How do you charge for the value? One-time, subscription, usage, tiered. That's a business model decision. It's about what the customer keeps getting, over time, for the money.

These are not the same question. They feel linked because for a decade cloud was the recurring cost — the data center was the thing you were obviously renting. So "subscription" and "cloud" got welded together in people's heads. Move the compute home, the logic goes, and the reason to pay monthly disappears with it.

But the data center was never the product. It was plumbing. The product was the thing the customer actually wanted: accurate detection, event clips, the model that got smarter last month, the patch that closed a vulnerability, the new capability that wasn't there when they bought the box. None of that stops being delivered just because it now executes on a chip in the customer's hallway instead of a server in Washington.

So pushing everything back to "buy the hardware, we'll throw in the software" isn't pro-customer. It's moving backwards. It's the pre-SaaS model wearing an edge-computing costume. The correct move is to push compute and storage to the edge and keep charging for the value — because the value is still flowing.

Tesla gave us the cleanest possible proof of this.

FSD runs on the edge. All of it. The compute is in the car the customer already paid for. There is no data center doing the driving. By the "you own the hardware, so it should be free" logic, FSD should be the most subscription-proof product on earth.

It's the opposite. Tesla went subscription-only and said plainly that the price will rise as the capability improves. Not as the compute moves. As the value grows. The customer isn't renting a cloud server. They're paying for a value that keeps getting better while it sits in their garage. Where the inference happens is irrelevant to why they pay.

I'll be honest about the obvious tension, because it's real. When the box is in your hallway, the question "what am I paying for?" gets louder. With cloud, the customer can at least picture the meter running. With edge, they stare at hardware they own and want a straight answer. Fair. But the answer to "what am I paying for?" was never "the building the GPUs live in." It was always the value. Edge just makes you say it out loud. That's a higher bar, not a different model.

I might be wrong about some of this. Maybe there are categories where the value really is one-and-done, where the box does exactly one thing forever and never needs to change, and a subscription there would be a tax with no service behind it. I'd believe that for some products. I don't believe it for anything with a model inside it that we'll keep improving.

The line I keep coming back to is simple. Don't let an engineering decision masquerade as a pricing decision. Move the compute wherever it serves the customer best. Then charge for what you actually deliver, and wherever it happens to run.

Curious whether people see this differently. Especially anyone who's shipped an edge product and felt the pull to give the software away. I'd like to hear where the argument breaks.