Nvidia has outlined a plan to offer software for running AI agents at about $2,000 per GPU annually, introducing a new direction for Nvidia subscriptions. Nvidia CEO Jensen Huang discussed the potential of a software licensing business that could generate recurring revenue beyond hardware sales at a GTC event held in Berlin, Germany, on October 22, 2026.
The core of the proposal is the total cost of operating all the Nvidia software required to run agents in an AI factory. It is a strategy to bundle security, access management, and infrastructure operations into a single platform, extending Nvidia’s hardware-focused revenue model into a subscription-based structure.
The $2,000 Figure Is Not a Final Product Price
The announced $2,000 figure appears to be an example offered to explain the licensing model rather than a confirmed selling price. Nvidia has not disclosed when the pricing would take effect, the specific software configuration, or the licensing terms. It also remains unclear how the offering would be distinguished from existing products.
Nvidia’s enterprise AI software platform, AI Enterprise, is currently sold for $4,500 per GPU annually. It is unclear whether the newly mentioned amount covers a narrower agent-specific package or represents a separate licensing system, making it difficult to interpret the figure as a price reduction for existing offerings.
The Scale of Recurring Revenue and Its Unconfirmed Assumptions
If subscriptions are introduced on a large scale, they could generate recurring revenue independent of GPU sales cycles, which is the primary commercial value of Nvidia subscriptions. If five million GPUs are covered by paid licenses, the annual revenue would amount to $10 billion, or $50 billion cumulatively if the subscriptions are maintained for five years.
However, these figures do not represent an actual revenue forecast. Nvidia has not disclosed the number of GPUs eligible for paid licenses or provided a revenue outlook, so the five-million-GPU and five-year figures should be understood as a hypothetical scenario. It is therefore difficult to determine the size of the market based solely on the $2,000 announcement.
OpenShell and Sentry-Managed Agent Security
AI agents can access external tools and data and autonomously perform multiple tasks, so large-scale deployment in enterprise environments requires a management system capable of controlling their execution scope and permissions. On October 21, 2026, Nvidia unveiled an open-shell reference architecture that supports isolating AI agent execution environments and restricting access.
OpenShell is based on Vera Rubin computing infrastructure, the Vera CPU, and the BlueField-4 data processing unit, and it monitors agent activity through Nvidia Sentry. Sentry can isolate agents that attempt to cross configured security boundaries. The key is that control is implemented at the infrastructure layer rather than relying solely on the model’s built-in safeguards.
Integration with external companies is also beginning. Anthropic is working to integrate OpenShell and BlueField hardware with managed Claude agents to manage access permissions for enterprise agents, while Salesforce is developing administrative capabilities for reviewing agents’ access requests in Slack and approving or rejecting them. These examples show that the security management layer can be applied to real services rather than remaining merely a reference design. A related report from The AI Times also highlighted Nvidia’s subscription expansion and its broader ecosystem impact.
From Hardware to Full Stack, With Lock-In Risks Remaining
Nvidia CEO Jensen Huang said people may eventually stop viewing Nvidia as merely a chip company. His remarks emphasized Nvidia’s direction of providing a technology stack that includes software and systems, alongside semiconductors, to industries ranging from healthcare and automotive to telecommunications and life sciences.
As enterprises build security and management systems on Nvidia’s infrastructure, the cost and effort required to migrate to another platform may increase. Recurring revenue and high switching costs are both advantages of Nvidia subscriptions and factors that could accelerate ecosystem lock-in. As the AI agent market expands, the value of the security layer will also grow. However, a meaningful evaluation of Nvidia’s business plan will require clarity on the rollout timeline, eligible GPUs, included software, licensing units and terms, and whether the offering overlaps with AI Enterprise.
Source Article
This article was written with reference to the following original source: The AI Times — Nvidia Considers $2,000 Annual Software Subscription per GPU…Accelerating Ecosystem ‘Lock-In’

Leave a Reply