You don’t buy the model. You rent the meter.

The industry still sells models as if they were objects you could own.

Most teams do not own them. They rent a meter: tokens, GPUs by the hour, a rate limit that moves when a lab needs the capacity back. The demo is infinite. The invoice is not.

The human scene

A five-person shop wires an agent into support. In week one it is magic. In week six the bill is a salary. They shrink context, then features, then they turn the agent off for the night and hope nobody notices.

A well-funded lab does not have that week six. That is the story. Not the leaderboard.

What “open” does not fix

Open weights can be a door. They can also be a hallway to a locked machine room. If you cannot afford to serve the thing you downloaded, you have a file, not a factory.

Licenses matter. So do power, chips, and the patience of whoever owns the cluster. We already covered the license that called itself open and forbade the commercial use startups actually need. The sequel is quieter: even a generous license does not pay the electricity.

What to watch

• Contracted capacity vs spot. Who is bumped when the lab’s own product spikes.
• Features that only exist at the expensive tier, then get described as “the model.”
• Local silicon (NPUs, small models) that actually change a daily default — camera, offline speech — versus a slide.

Chip stories matter when they change who can build, ship, or afford intelligence. They do not matter when they only move a stock chart for an afternoon.

If your product plan assumes the meter stays cheap, you do not have a plan. You have a demo schedule.