Add preferred source
Nvidia is considering an annual software fee of about $2,000 per GPU for the tools that run AI agents on its chips, CEO Jensen Huang said at GTC Berlin on Oct. 22. At that rate, five million paid licenses would generate roughly $10 billion a year, or $50 billion over five years with renewals, though Nvidia has not set eligibility, final pricing or product scope. The company already charges $4,500 per GPU for a one-year AI Enterprise subscription, and has not said whether the two offerings overlap. The proposal follows the Oct. 21 launch of Nvidia’s OpenShell Reference Architecture, a secure runtime for agents built on Vera Rubin systems, Vera CPUs and BlueField-4 processors running Nvidia Sentry, with Anthropic and Salesforce among partners.
Key Elements
Nvidia Corp. is floating an annual software charge of roughly $2,000 per GPU for the tools that run AI agents on its hardware, a move that would deepen the chipmaker’s push into software and could eventually generate tens of billions of dollars in recurring revenue.
Chief Executive Jensen Huang outlined the idea at GTC Berlin on Oct. 22, framing licensing as a growing piece of the company’s business. “Software licensing will become increasingly a major part of our company’s business,” he said. “The way that we license will be… for example, it’s about $2,000 per GPU per year to run all the software that Nvidia provides that runs agents on these AI factories.”
The arithmetic is striking. At $2,000 per GPU a year, five million paid licenses would produce about $10 billion in annual revenue, or $50 billion over five years if customers renewed. Those figures are illustrative scenarios, not company guidance: Nvidia has not said how many GPUs would qualify for the license, whether $2,000 is a standard or final price, or which products the fee would cover.
The proposal sits uneasily alongside Nvidia’s existing software business. The company already sells AI Enterprise, listed at $4,500 per GPU for a one-year subscription — more than double the figure Huang cited. Nvidia has not confirmed that the two prices cover the same software or licensing terms, so the $2,000 number should not be read as a price cut.
| Item | Figure |
|---|---|
| Proposed annual fee per GPU | About $2,000 |
| AI Enterprise one-year subscription | $4,500 per GPU |
| Annual revenue at 5 million paid licenses | $10 billion |
| Five-year revenue if subscriptions renew | $50 billion |
Note: The revenue figures are illustrative calculations based on the $2,000 rate, not Nvidia forecasts. Nvidia has not disclosed eligibility or final pricing.
The timing is tied to the rise of AI agents — systems that can call tools, pull in data and complete tasks with little human oversight. Companies running them at scale need software to manage workloads and to stop agents from reaching resources they should not touch.
On Oct. 21, Nvidia introduced its OpenShell Reference Architecture, a secure runtime for agents that sets limits on their actions and access. It combines the company’s Vera Rubin computing infrastructure, Vera CPUs and BlueField-4 data processing units running Nvidia Sentry, which watches agent behavior independently and can quarantine an agent that tries to cross a security boundary. The design pushes controls into the underlying infrastructure rather than relying only on safeguards inside a model or application, building on the Open Agent Safety Platform Nvidia introduced in September.
Partners are already building on it. Anthropic is integrating OpenShell and BlueField hardware with Claude Managed Agents to give businesses more control over what agents can reach inside execution sandboxes. Salesforce has developed a Slack Admin Bridge for OpenShell that lets administrators review access requests, approve or deny permissions, and monitor activity from within Slack.
The licensing talk fits Huang’s broader argument that Nvidia is outgrowing its chip label. The company has expanded into networking, full systems, enterprise software and infrastructure for AI workloads. On the BG2Pod podcast, he said, “One of these days, people might not categorize us as a chip company anymore,” pointing to its presence in healthcare, automotive, telecommunications and life sciences, where it sells broader technology stacks.
For investors, the significance lies in the shape of the revenue. Hardware sales are lumpy and tied to <a href="https://bitcomme.com/top-venture-<a href="https://bitcomme.com/top-venture-capital-firms-vcj-50-2/” title=”Top Venture Capital Firms | VCJ 50″>capital-firms-vcj-50-3/” title=”Top Venture Capital Firms | VCJ 50″>capital spending cycles, while per-GPU software subscriptions would recur annually and carry higher margins. If agent deployments spread through corporate data centers, a licensing layer on top of installed GPUs could turn Nvidia’s installed base into a subscription annuity. The open questions are how much customers will pay on top of hardware, whether rivals offer comparable agent security for less, and whether enterprises accept per-GPU pricing as agents move from pilots into production.
