In context: Most of Nvidia's money comes from selling the hardware that fills AI data centers. Those sales are largely one-off deals, though, so Jensen Huang seems keen on something that keeps paying after the chips ship. At GTC Berlin last week, the CEO said it costs about $2,000 per GPU every year to run all of the company's software for AI agents.
The Register was first to report Huang's quote from a question-and-answer session at the event on October 22. The CEO told attendees that software licensing "will become increasingly a major part of our company's business." The $2,000 figure was his example of how that licensing would be priced.
That money covers the Nvidia software needed to run agents inside what the company calls AI factories – its name for data centers built around AI work. However, it's not clear how this fits with what Nvidia already charges. The company sells an AI Enterprise suite that businesses license to build and run AI on its hardware. It's listed at $4,500 per GPU for a year, support included. Committing to five years brings the total to $18,000 per GPU, equivalent to $3,600 a year.
Huang's figure is lower than both existing prices, but the company hasn't explained how the two prices relate. There's also no word on when the $2,000 pricing might arrive.
The company has been building out that agent software at a fast pace. A day before the Q&A, it launched the OpenShell Reference Architecture, an expansion of a safety platform introduced in September. In it, agents run on Vera CPUs inside OpenShell, an open-source sandbox that controls which files, networks, and tools they can reach. BlueField-4 networking chips running Nvidia's Sentry software watch from the outside and can quarantine an agent within milliseconds if it oversteps those limits.
Huang leaned on OpenShell when asked about AI safety during the same session. "If you're making an omelette, you don't need a hammer," he said.
At $2,000 each, every million licensed GPUs would bring Nvidia $2 billion in recurring annual revenue, as long as customers continued running the software. That's not much next to the $96.2 billion the company brought in last quarter. Still, it's money that doesn't depend on anyone buying new chips.
Since the fee is charged per GPU, how Nvidia counts a GPU matters, too. When Huang said last year that the company would start counting each silicon die as a separate GPU, The Register noted that the change could double the number of AI Enterprise licenses some servers need. Nvidia has since gone back to counting whole modules for Vera Rubin, at least in how it describes its systems.
