NVIDIA CEO Jensen Huang has cited a figure of roughly $2,000 per GPU per year as an example of what the company might charge for software that runs AI agents in AI factories. According to The Register's on-the-ground report dated October 22, the figure came up during a Q&A session at GTC Berlin in Germany, and it suggests NVIDIA wants to expand software licensing as a core pillar of its business. However, this was not a formal announcement of new pricing, and the products it would cover, when it would become available and how it would relate to existing contracts have not been made clear. NVIDIA already offers an enterprise software subscription at $4,500 per GPU per year, and some GPUs come with software entitlements included. What the $2,000 figure would cover, and what additional burden it would place on customers, needs to be examined carefully.
According to remarks published by Tom's Hardware, Huang said, "Software licensing will become an increasingly large part of our business." He offered the $2,000 figure with the qualifier "for example," so it was not presented as a confirmed price. The cost of buying GPU hardware is separate from the fee for continuing to use the software that runs on it, and without knowing which features and support a contract includes, it is impossible to judge the actual burden on customers.
Existing AI Enterprise costs $4,500 per year, with entitlements bundled with some GPUs
NVIDIA's current enterprise software platform, "NVIDIA AI Enterprise," already has a per-GPU licensing structure. According to the official pricing guide, list prices for systems that companies manage themselves are $4,500 per GPU for a one-year subscription and $18,000 per GPU for a five-year subscription. Both are partner-channel prices that include standard support, and the five-year subscription provides five years of use for the price of four. Looking only at the one-year price overlooks the fact that the effective annual cost changes with the contract term.
In addition, depending on the GPU product and generation, an AI Enterprise entitlement comes bundled with the hardware. Based on NVIDIA's official licensing guide, the main purchase types and conditions are summarized below.
| Purchase / usage type | Price / bundled entitlement | Main conditions |
|---|---|---|
| One-year subscription for self-managed systems | $4,500 per GPU | Includes standard support |
| Five-year subscription for self-managed systems | $18,000 per GPU | Five years of use for the price of four; includes standard support |
| H100 PCIe / H100 NVL, H200 NVL | Five-year entitlement bundled | Activation required to start using it |
| Hopper-generation DGX | Included in the software bundle | Bundled license terms apply |
| Blackwell-generation DGX | AI Enterprise purchased separately | Purchase terms differ from Hopper generation |
In other words, even today's AI Enterprise is not a system in which every customer pays a flat $4,500 per year. Besides term-based discounts, some GPUs come with an entitlement at the time of purchase. The $2,000 per year that Huang cited therefore cannot be read as a price cut to the current AI Enterprise or as a new add-on fee imposed on all NVIDIA GPUs. It is not even known whether the $2,000 covers the same functions as the existing product.
The terms of bundled entitlements also vary by GPU type. For example, the A800 40GB Active comes with a three-year term, which does not match the five years for the H100 PCIe/NVL and H200 NVL. Also, Hopper-generation DGX includes AI Enterprise as part of its software bundle, while Blackwell-generation DGX requires a separate purchase. Whether a company that deploys GPUs must buy additional software depends first on what entitlements come with the hardware it owns.
How licenses are counted also matters. On servers and workstations running NVIDIA AI Enterprise, a license is generally required for every installed GPU. In environments with no GPUs, one license is counted per server or instance. During the subscription term, maintenance, bug fixes, security updates and access to new software versions are provided, but continuing the subscription after the term ends requires renewal. Activating an entitlement bundled with hardware and renewing after a contract expires are separate matters.
The cloud has a different pricing structure as well. For production use, a rate of $1 per GPU per hour applies, with the cloud instance fees charged separately. Support terms also differ from annual contracts, so simply converting this into an annual figure for comparison is not appropriate. Even existing AI Enterprise offers a range of purchase types and contract conditions, so judging what the $2,000 remark means requires an explanation of the products it covers and the specific contract terms.
OpenShell's role in monitoring and controlling AI agent execution
One piece of software NVIDIA is developing for AI agents is "OpenShell." It is an open-source runtime that runs AI agents in an isolated environment and controls their access to external systems. According to NVIDIA's official description, rather than giving agents unrestricted access to networks and files, it supervises external communication and restricts operations according to defined rules. The design separates the AI model's reasoning about its next action from the decision on whether that action is actually permitted.
For example, when an agent tries to connect to an external service, the supervising mechanism checks the destination, communication method and access path against the rules and decides whether to allow or deny the connection. The result is logged so that what the agent attempted can be traced. This differs from simply instructing the AI model not to perform prohibited operations: it restricts actual behavior from outside the model. Including file access permissions, it aims to manage the range of operations an agent can perform, separately from techniques that improve the accuracy of AI responses.
This division of roles is also reflected in the OpenShell Reference Architecture that NVIDIA announced on October 21. In this reference architecture, Vera Rubin handles AI model inference, and the Vera CPU handles agent execution management and tool-based processing. In addition, "Sentry," running on the BlueField-4 DPU, monitors operations from outside the agent and applies configured security policies. By splitting inference, agent execution management and security monitoring across different processing platforms, the design aims to improve safety and manageability when operating AI agents at scale.
NVIDIA explains that this setup allows it to take measures such as isolation or suspension if an agent tries to exceed its permitted scope. However, this is the company's description of its design and features, and independent verification has not guaranteed that it can prevent every attack or unauthorized operation. Nor is the Vera CPU required to use OpenShell itself. The hardware configuration NVIDIA recommends should be distinguished from the OpenShell runtime itself.
This distinction also matters when considering the $2,000 annual fee. The reference architecture shows a structure in which software and hardware for agent management and monitoring are combined around GPU-based inference. But it is not clear whether the fee Huang cited would apply to OpenShell itself, include monitoring features such as Sentry, or refer to a software contract for the entire AI agent platform. The fact that OpenShell's code is published as open source is also no basis for judging which features and support a paid license would include.
NVIDIA's recurring revenue hinges on contract counts and renewal rates more than GPU shipments
If NVIDIA expands its software business, what matters for revenue is not the number of GPUs shipped but the number of GPUs covered by paid licenses and the length of the contracts. Even if the same number of GPUs is deployed, the additional payments differ between products that come with a software entitlement and those that require a separate license purchase. After a contract term ends, whether customers choose to renew changes the recurring revenue. For this reason, growth in GPU shipments cannot be directly tied to growth in annual software sales.
If paid contracts at $2,000 per GPU per year were to run for one year across 1 million GPUs, contract value would come to $2 billion a year. However, this is merely a simple calculation of the relationship between price and contract count, not an indication of NVIDIA's actual contract numbers or revenue forecasts. It does not take into account discounts, entitlements bundled with hardware or mid-term cancellations, nor does it include the timing of revenue recognition for accounting purposes. Without knowing how many of the shipped GPUs would be covered by paid contracts, the actual scale of revenue cannot be estimated.
For customers, too, existing entitlements and renewal terms are important in calculating deployment costs. For example, license costs at deployment differ between a GPU that comes with a five-year AI Enterprise entitlement and a Blackwell-generation DGX that requires the software to be purchased separately. Furthermore, renewal costs may arise if a company wants to keep using the software after a bundled entitlement ends. Rather than comparing the $2,000 annual figure alone with existing contracts, companies need to evaluate it together with entitlements they already hold, new features they would purchase, contract terms and renewal conditions.
For NVIDIA, software that supports the execution and monitoring of AI agents could become a new business opportunity for earning recurring revenue on top of GPU sales. In particular, as companies run AI agents in production, demand may arise not only for inference performance but also for access control, monitoring, incident response and software maintenance. However, the spread of technologies like OpenShell is not the same as growth in the number of paid license contracts. Without clarity on which operational features companies will pay for and keep renewing, it is difficult to judge how large the software business could become.
Huang's remarks suggest that NVIDIA aims to earn recurring revenue not only from selling AI infrastructure but also from the agent operations software that runs on it. But the $2,000 annual figure is only an example, and no formal pricing structure has been announced. Going forward, the focus will be on whether the covered products and GPUs, the relationship to existing AI Enterprise contracts, and the paid operational features, support and renewal terms become clear. Only when that information is available can companies compare the costs of adopting and continuously operating an AI agent platform, and only then can NVIDIA's software business be evaluated based on actual contracts.