NVIDIA intends to profit from artificial intelligence not just by selling graphics accelerators. Company CEO Jensen Huang stated that software for AI agents could generate around $2,000 per year from each GPU. According to him, software licenses will occupy an increasingly prominent place in the manufacturer's business, and maintaining autonomous agents on specialized computing platforms requires a whole set of software tools. Thus, the company is laying the groundwork for recurring revenue that persists throughout the entire operational lifespan of the hardware, rather than just at the time of purchase.
From Chip Sales to a Subscription Model
Huang announced the $2,000 figure during a Q&A session at the GTC Berlin conference held on October 22, 2026. According to him, users will have to pay annually for each activated accelerator, with the license granting access to the software tools needed to run and support AI agents. This marks a shift in NVIDIA's strategy: the semiconductor manufacturer is increasingly expanding its business beyond hardware solutions and offering comprehensive products for so-called "AI factories" — specialized infrastructure ensuring the operation of artificial intelligence and autonomous agents.
Pricing Is Not So Simple
However, pricing details are not yet entirely clear. The stated $2,000 is an estimated figure rather than an officially announced tariff. The existing NVIDIA AI Enterprise subscription costs $4,500 per single GPU per year with support, or $18,000 for a five-year term. The company has not specified whether the new estimate applies to a separate AI agent package, another set of solutions, or a future licensing model. Therefore, talking about a price cut is premature: for now, this concerns a potential revenue level rather than a fixed price list.
OpenShell and Agentic AI Infrastructure
At the same event, NVIDIA introduced the OpenShell Reference Architecture for agentic AI infrastructure. It integrates Vera Rubin systems for inference, Vera processors for running agents, and BlueField-4 DPUs paired with NVIDIA Sentry to monitor their actions. OpenShell is an open-source environment that sandboxes agents; Huang mentioned it specifically in the context of AI safety. This emphasizes that the software layer for the company is not just an add-on, but a key element of managed and controlled agentic infrastructure.
Scale of Potential Revenues
The potential scale of the new model is impressive: one million licensed GPUs at $2,000 per year would bring the company about $2 billion in annual revenue. According to calculations by the analytical resource FourWeekMBA, selling five million such licenses at $2,000 per year could generate up to $50 billion in revenue over five years, although experts emphasize the hypothetical nature of such estimates. For NVIDIA, this is an opportunity to generate recurring software revenue on top of hardware sales — and essentially a new profit stream tied to every operating accelerator in client fleets.
