The one-screen version: on October 22, Nvidia CEO Jensen Huang floated a figure of roughly $2,000 per GPU per year for the software that "runs agents on these AI factories." It is a floated number, not a price sheet — Nvidia hasn't said what it covers or when it starts — and The Register flagged its own line as a paraphrase. But the direction is the news: the GPU you rent is growing a recurring software toll on top of the silicon. The same week, Daytona raised $60M at a $1B valuation to sell the agent sandbox as its own infra layer. The founder's read: the rental rate you track on a price map is no longer the whole cost of running an agent — budget the fully-loaded cost per agent-run, not dollars per GPU-hour.

What actually happened#

Two numbers, four days apart, pointing the same way.

One is a chip vendor reaching up into software. One is a startup building a layer that didn't exist in 2024. They describe the same thing from opposite ends: running an agent now accretes recurring infrastructure, billed in separate lines, and almost none of it shows up on the hourly GPU rate you've been tracking.

The one real idea: the GPU rate is the floor, not the bill#

For two years the honest way to compare compute was dollars per GPU-hour. You could put an H100, an H200 and a B200 on a price map, pick the cheapest that fit, and know your floor. That number is still real — but it's become the floor, not the bill.

Stack what an agent workload now actually carries:

  1. Raw compute — the GPU-hours. Still trackable, still substitutable (multi-cloud, spot, open hardware).
  2. The model — per-token, and as we wrote yesterday, increasingly swappable inside a single cloud catalog.
  3. The vendor software stack — the layer Huang just put a recurring, per-GPU price on.
  4. The sandbox — the isolated runtime an agent needs to execute code safely, now its own venture-funded market.

Only layer 1 appears on a rental price map. Layers 3 and 4 are recurring, they're priced per-unit-of-the-thing-you-already-pay-for, and they compound. This is the same mechanism as the reseller-margin math that bites wrappers: your real cost of goods is several times the naive "one call," because the taxes hide in layers the headline rate never named. A vendor that sells you the chip, and the software that runs on the chip, and takes a cut of the agent runtime has three bites of the same workload — and markets only the first one as a price.

What to do with it#

Nvidia floating a per-GPU software fee the same week a sandbox startup hit $1B is not a coincidence — it's the agent-compute bill revealing its true shape. The hardware was always the easy part to price. The recurring layers stacking on top of it are the part that will quietly set your margin, and the rental rate is just the first line on a bill that's getting longer.

Also this week: Huang, in the same remarks, waved off talk of an "AI bubble," pointing to demand he called far greater than current capacity — the standard vendor framing for why the toll is worth paying. Price it like a vendor would: as the opening number in a negotiation, not a fact.