Nvidia is signaling that the next major business built on its GPUs may not be hardware at all. CEO Jensen Huang said software for running AI agents could cost roughly $2,000 per GPU each year, a model that could become a multibillion-dollar recurring-revenue engine as companies deploy autonomous systems at scale. The proposed Nvidia software fee was discussed during the company’s GTC Berlin event on October 22.
The figure is not yet a confirmed universal price, and Nvidia has not specified exactly which products would be included. Still, the arithmetic explains why Huang is emphasizing licensing. Five million paid GPU subscriptions at $2,000 annually would produce $10 billion in yearly revenue, or $50 billion across five years if customers continued renewing. The estimate is hypothetical rather than a sales forecast, but it illustrates how software could let Nvidia monetize each deployed accelerator long after the initial hardware purchase.
There is already a higher-priced comparison point. Nvidia’s official enterprise licensing guide lists AI Enterprise at $4,500 per GPU for a one-year subscription, with support included. Nvidia’s documentation also confirms that the platform is licensed on a per-GPU basis. Huang’s $2,000 comment may describe a different bundle, a narrower agent-focused offering, or a future licensing structure; it should not be treated as an announced discount on AI Enterprise.
The timing connects the pricing discussion to Nvidia’s expanding security push around agentic software. On October 21, the company announced its OpenShell Reference Architecture, which combines Vera Rubin systems for model inference, Vera CPUs for agent runtimes and tool use, and BlueField-4 data-processing units running Nvidia Sentry. The architecture is designed to give agents useful access to tools while placing limits on what they can reach, with Sentry independently monitoring behavior and enforcing policies. Nvidia says the system can quarantine an agent that attempts to cross those boundaries.
That infrastructure-first approach is significant because enterprise customers are unlikely to treat autonomous agents like ordinary chatbots. Agents may read internal data, call business tools, alter records or initiate transactions, creating demand for audit trails, permissions, isolation and a clear party responsible when something goes wrong. Nvidia’s own developer explanation describes Sentry as a separate enforcement layer built on its DOCA software framework, rather than relying only on safeguards inside the model itself.
For Nvidia, the strategy could bind together chips, operating software, security and support into one enterprise stack. Independent coverage of the OpenShell design describes the same division of labor: models think on Rubin, agent software acts through Vera, and BlueField hardware governs activity. That may make Nvidia’s software harder to replace, while also raising questions about whether customers will accept per-GPU fees for systems that already represent substantial capital investments.
The immediate takeaway is that the $2,000 number remains a marker of Nvidia’s ambition, not a finalized price sheet. The company still needs to clarify the covered products, eligible GPUs, licensing terms and whether OpenShell-related capabilities will be bundled or sold separately. If enterprise agents become a standard workload, however, Huang’s remarks show Nvidia intends to charge not only for the machines that run them, but for the software that keeps them useful and under control.
