AI doesn't kill software. It kills the seat.
Own the volume. Rent the exceptions. Keep the learning.
For twenty years, business software was priced by the person. You bought a license per employee, counted heads, and knew your bill. Headcount moves slowly and predictably, so a flat per-person price made sense. The seat was the ceiling: the thing that capped your software spend at the size of your payroll.
An AI agent is a different kind of user. It does not map to a name on the org chart. It works in loops, around the clock, and spins up by the thousands when there is work to do. So the meter comes off the person and moves onto the task. Off the seat, onto the token, the unit of AI work.
That one shift changes the shape of the bill. Your software expense stops being a fixed subscription and starts behaving like a cost of goods: it scales with how much work you do, not how many people you employ. And the moment a cost behaves like that, every operator asks the oldest question in business. Do I keep renting this, or do I own it?
Think of it like power. For years you rented every kilowatt from the grid. The meter spun and the bill grew. Then solar changed the math. You pay once for the panels, and the next kilowatt costs almost nothing. You do not rip out the grid; you still lean on it for peaks and cloudy days. But the steady, predictable load, you generate that yourself. In AI, that steady load is the claim summaries, the first-draft replies, the contract checks, the thousand small tasks that repeat every day.
Now notice what electricity does not do. It only flows one way. When you buy power, nothing of yours travels back up the wire. When you buy tokens, everything does: your documents, your questions, your workflows, and your judgment. That difference is the whole argument, and the cost half was the easy half.
AI is having its solar moment, and you do not have to take a startup's word for it. This past spring, in the span of two weeks, Microsoft CEO Satya Nadella did two things.
First, he promised "unmetered intelligence," AI you own rather than rent, where the hardware is bought once and each additional task costs almost nothing.
Second, he argued the real prize is bigger than cost. The durable advantage of a company is not the model it rents. It is the learning loop it owns: the way its people, data, and decisions compound into institutional knowledge that gets sharper every time it is used. You can offload a task, or even a job, he wrote, but "you can never offload your learning."
That last point is the one to sit with. Your most valuable work accumulates on a platform you rent, in infrastructure and memory you do not fully control, whatever the contract says about how your data is used. You keep your learning, but only if you own where it happens.
And notice who is saying it. When the company that runs one of the biggest grids on the planet tells you to buy panels and own the loop, that is not marketing. It is the most credible endorsement owned compute will ever get.
There are two versions of the risk. For banks and hospitals, it is where the data sits: residency, jurisdiction, and who can be compelled to hand it over. For every company, it is whether the compounding institutional edge stays behind its own walls or leaks into systems anyone can rent.
The answer starts with compute you control: the AI PC on the desk, the server down the hall, the private environment you run. Real hardware you manage the way you already manage everything else. It does not end there, because identity, access, logging, and governance all matter, but it starts with owning where the work runs.
None of this is a case against the cloud. The novel, the huge, and the once-in-a-while problems still belong on rented, world-class models. The point is not local or cloud. The point is placement.
Most companies do not need an AI ideology. They need a workload map. Which work is high-volume enough to own? Which is sensitive enough to keep close? Which needs the best rented model on the planet?
A word on where I sit. Before Microsoft, I helped sell SGI supercomputers into buildings the data was never allowed to leave. Then I spent fourteen years marketing Surface inside a company that bet its future on the cloud. I have worked both ends of this pendulum. It is swinging.
That is why we started Edge Forward. We help companies decide what AI work to own, what to rent, and how to protect the data, workflows, and institutional learning created along the way.
Own the volume. Rent the exceptions. Keep the learning.
The token bill is the new software bill. Where those tokens run is the new margin decision, and it is now a board-level question. None of this is a slogan. It is arithmetic, and we put the model online so you can check our work.