Nvidia Wants $2,000 a Year in Software Fees for Every GPU

October 26 — Nvidia chief executive Jensen Huang says the annual licence fee for the accompanying AI software will be about $2,000 per GPU (around 13,355 yuan at current exchange rates), to run “the full suite of software Nvidia provides that keeps AI agents running on these AI factories.” Video of the remarks surfaced recently.

Speaking at the Berlin GTC conference on 22 October, Huang discussed the pricing, and Nvidia also unveiled broader plans there: an AI infrastructure business spanning chips, software and the security of AI agents. “Software licensing will gradually become a very important part of our business. Our licensing model is roughly like this… for example, running the full suite of Nvidia software for deploying agents on AI factories costs about $2,000 per GPU per year.”

The size of this revenue is easy to estimate. If 5 million GPUs are covered by paid licences, at $2,000 each per year Nvidia would take in $10 billion a year; as long as customers keep renewing, that comes to $50 billion over five years. But there is uncertainty here: Nvidia has not said whether the $2,000 is charged on top of existing fees or replaces them.

Making that distinction is crucial, because Nvidia already sells paid enterprise software. Its AI Enterprise platform is listed at $4,500 per GPU per year. The $2,000 Huang mentioned is lower, but Nvidia has not confirmed that the two prices cover identical products and licence terms, so it cannot be read directly as an official price cut.

The background to this discussion is that Nvidia is building more supporting infrastructure to help companies deploy and manage AI agents. Such systems can call tools, read data and execute tasks automatically, and when companies roll them out at scale they create new security and management problems.

On 21 October Nvidia released the OpenShell reference architecture, which combines Vera Rubin compute infrastructure, the Vera CPU and the BlueField-4 data processing unit (DPU) carrying Nvidia’s Sentry security component. OpenShell gives AI agents a secure runtime, defines the boundaries of what they may do, and limits their permissions to call tools and access resources. Nvidia says Sentry can independently monitor agent behaviour, and that when an agent attempts to cross a boundary, it steps in.

Nvidia launched an open agent security platform back in September, and this new reference architecture extends it, carrying the control approach into enterprise infrastructure and edge devices.

Other vendors are adding similar controls. Anthropic has combined OpenShell with BlueField hardware and Claude managed agents so that companies can control what agents may access inside execution sandboxes. Salesforce has built a Slack management bridge for OpenShell, letting administrators review agent access requests in Slack, approve or reject permissions and monitor operations.

The new software pricing also echoes Huang’s long-standing argument that Nvidia is no longer just a GPU supplier. Its business has extended to networking equipment, whole systems, enterprise software and infrastructure for AI workloads. In a recent appearance on the BG2Pod podcast Huang said: “One day, people may no longer classify us as a chip company.” He noted that Nvidia has entered healthcare, automotive, telecoms and life sciences, providing a full technology stack rather than just chips.

Even so, Nvidia has not said when the $2,000-per-GPU-per-year charge will take effect, which software it includes, or how this pricing will connect with its existing enterprise software business.