NVIDIA plans to charge a software subscription fee of about $2,000 per GPU per year

Nvidia is planning to further expand its software business, hoping to obtain more stable recurring revenue through enterprise software licensing in addition to selling AI chips. Nvidia CEO Jensen Huang recently discussed this strategy at the GTC event in Berlin, Germany, saying that software provided for running AI agents could adopt a licensing model charged per GPU, for example at about $2,000 per GPU per year. This figure has drawn industry attention to the potential scale of Nvidia's software business, but it cannot yet be regarded as a formally announced uniform pricing standard.

According to reports, Huang discussed the software licensing business during a related Q&A session on October 22, 2026. He said software licensing will increasingly become an important part of Nvidia's business, especially in providing the software, management tools and supporting capabilities needed to run AI agents for computing equipment in AI factories.

So-called AI agents refer to AI systems that can execute multi-step tasks according to goals and, within authorized scope, invoke tools, access data or run programs. Compared with traditional chatbots, such systems usually require more complex operating environments, security controls and continuous monitoring mechanisms. If enterprises want to deploy AI agents at scale in data centers, they not only need to buy GPUs but also need to address software deployment, permission management, operation monitoring and security maintenance.

Nvidia is trying to integrate these capabilities into a more complete enterprise-grade infrastructure solution. Based on the example price proposed by Huang, if 5 million GPUs adopt a software license of $2,000 per year, this could theoretically generate $10 billion in revenue per year; if maintained at this scale for five consecutive years, cumulative revenue would reach $50 billion. However, this is only a scenario calculated based on assumptions, not a revenue forecast published by Nvidia. At present, the company has not disclosed the actual number of GPUs to which this would apply, nor has it confirmed which products this price would cover, when it would be implemented, or the specific contract terms.

It is worth noting that Nvidia already has a mature enterprise software subscription business. Its NVIDIA AI Enterprise platform is aimed at enterprise data centers and provides software components needed for AI development, deployment and management. Nvidia's published license prices show that a one-year subscription to the platform costs $4,500 per GPU, including corresponding support services; a five-year subscription costs $18,000 per GPU.

Therefore, the $2,000 per year mentioned by Huang cannot simply be understood as a price cut for Nvidia's existing enterprise software. Because it is still unclear whether the software scope, license term and service content corresponding to the two prices are the same, their actual costs cannot currently be directly compared. The new price may target a different software combination or a specific AI agent operating environment, and the specifics still await further explanation from Nvidia.

The reason Nvidia values its software business is related to its desire to reduce dependence on one-time hardware sales. GPU sales are usually closely tied to data center construction cycles, customer procurement plans and chip supply capacity, while software subscriptions can generate recurring revenue while customers continue to use the relevant products. If Nvidia can make enterprises rely on its software platform for the long term to run and manage AI workloads, it will have the opportunity to continuously earn revenue from already deployed GPUs, rather than only recognizing one-time sales revenue when new chips are sold.

This strategy is also related to Nvidia's recently advanced AI agent security architecture. On October 21, 2026, Nvidia announced the OpenShell reference architecture, aimed at providing a more complete security environment for enterprises running AI agents. The solution integrates Nvidia's Vera Rubin computing infrastructure, Vera CPU and BlueField-4 data processors, and introduces a monitoring mechanism called NVIDIA Sentry.

OpenShell provides an operating environment for isolating AI agents and can restrict the tools, data and system resources that agents can access. NVIDIA Sentry is responsible for independently monitoring agent activity and taking isolation measures when it detects that an agent attempts to cross established permission boundaries. Its design approach is to place some security controls in the underlying infrastructure rather than relying entirely on the AI model's own security mechanisms.

This is especially important for enterprises deploying AI agents. Agents may be authorized to read internal files, access business systems, or even execute operations with real consequences. If permissions are improperly configured, or if an agent is induced by malicious input, it could cause data leaks, unauthorized operations or other security risks. Nvidia hopes to provide enterprises with an easier-to-manage deployment solution through operating environment isolation, permission restrictions and independent monitoring.

At present, other enterprises have also begun cooperation around this architecture. Anthropic is integrating OpenShell and BlueField hardware with Claude Managed Agents to provide additional runtime isolation and access control capabilities for enterprise agents. Salesforce has developed Slack Admin Bridge for OpenShell, allowing administrators to directly review agent access requests, approve or deny permissions, and monitor related activities in Slack.

These partnerships show that competition in AI agents is gradually expanding from model performance to runtime environments, security management and enterprise services. As agents begin to take on more complex work, enterprises not only need to judge whether a model can complete a task, but also need to ensure that it respects permission boundaries during execution and can be continuously monitored and managed. Nvidia hopes to integrate GPUs, networking equipment, software platforms and security mechanisms into a complete solution, thereby increasing customers' dependence on its technology ecosystem.

Huang Renxun has also repeatedly emphasized that Nvidia should not be seen merely as a chipmaker. The company's business has expanded into networking equipment, complete systems, enterprise software, and fields such as healthcare, automotive, telecommunications and life sciences. By providing a complete technology system from computing hardware to software management, Nvidia hopes to generate revenue at multiple points in the AI infrastructure chain.

However, whether the software licensing model can become an important new source of revenue for Nvidia still depends on the acceptance of enterprise customers. Some customers may be willing to pay ongoing fees for stable operation, security guarantees and technical support; others may choose open-source software, develop their own management tools, or meet their needs through solutions offered by other suppliers. Charging by the number of GPUs also means that as enterprises expand their deployments, software costs may rise accordingly, so customers will pay closer attention to licensing scope, contract terms and whether there is duplicate payment.

At present, Nvidia has not officially announced the full terms of this licensing plan of about $2,000 per GPU per year, nor has it given a specific implementation date. At this stage, this figure is more suitable as a price example raised by Huang Renxun when explaining the software business strategy, rather than a commercial quote that has already fully taken effect.

In the long term, Nvidia is trying to establish a new revenue structure: hardware sales help customers build AI infrastructure, software licenses continue to generate revenue after the infrastructure is put into use, and security and management tools further increase enterprises' willingness to use its platform over the long term. If this model succeeds, Nvidia's future growth will no longer depend entirely on how many GPUs it can sell each year, but will increasingly depend on how much customers are willing to pay for the software and services that run on those GPUs.