Nvidia Floats a $2,000 Yearly Fee to Run AI Agents

Nvidia has spent two decades selling chips. At GTC Berlin this week it started talking, out loud and with a number attached, about selling the software that runs on them. Jensen Huang told a Q&A on 22 October that it costs "about $2,000 per GPU per year to run all the software that Nvidia provides that runs agents on these AI factories," a figure relayed by Notebookcheck and a run of trade outlets.

Before anyone budgets for it, a caution. The $2,000 was an example, not a published rate. Nvidia has not said which products the fee would cover, or when it might start charging. The number is also a little hard to square with what the company already lists, because its existing Nvidia AI Enterprise suite sells at $4,500 per GPU for a one-year subscription, as Tom's Hardware has noted. So the lower figure either describes a narrower bundle or a direction of travel rather than a fixed menu.

The direction of travel is the real story. Huang framed software licensing as something that "will become increasingly a major part of our company's business," which is a notable thing to say from a company whose margins already draw complaints. One outlet ran the arithmetic and suggested the agent-software line could reach $50 billion over five years at scale. That is a projection stacked on an illustration, so treat it gently, but the ambition is not in doubt.

What you would actually be paying for

The fee attaches to a product Nvidia unveiled at the same event. Its agent-containment stack pairs Vera CPUs, which run the agents inside a sandbox it calls OpenShell, with BlueField-4 chips running a separate watchdog called Sentry. The pitch is that Sentry sits apart from the agent's own software, watches what it does, and can isolate a misbehaving agent within milliseconds. OpenShell's policy checker is meant to catch cases where an agent's permissions could be combined into something its operator never intended.

If that problem sounds familiar, it is because we keep writing about it. Agents that spend money they were not given and models that route around an access gate are exactly the failures Nvidia is proposing to sell a guardrail against. The security vendors CrowdStrike, Check Point, Palo Alto Networks and Zscaler are listed as adding support, and Anthropic, Google, Meta and OpenAI appear as model-builder partners. Decrypt reports Sentry's availability as early next year.

Nvidia has been busy in Berlin. It also showed off a way to ship a data centre in a shipping container and spent the keynote pitching Europe its own sovereign AI. The software fee fits the same plan. Sell the shovels, then sell a subscription to the handle.