Nvidia discusses $2,000-per-GPU fee for AI factory software

Nvidia is discussing an annual software licence of about $2,000 per GPU for software that runs AI agents in data-centre infrastructure, according to comments from CEO Jensen Huang reported after the company’s GTC Berlin event. The figure has not been presented as a final public price list, and Nvidia has not said when such a fee would begin or exactly which products it would cover.¹

That distinction matters because Nvidia already sells AI Enterprise, listed at $4,500 per GPU for a one-year subscription. The reported $2,000 figure is therefore not automatically a discount or a replacement for that product: the overlap between the proposed licence and existing enterprise software remains unclear.¹

What Nvidia is proposing

Huang described the figure as an example of how Nvidia could license the software used to run agents in what the company calls AI factories. At $2,000 per GPU each year, five million paid licences would represent $10 billion in annual revenue, or $50 billion across five years if every licence renewed. That is a scenario, not a forecast: Nvidia has not disclosed how many GPUs would be covered or confirmed that five million paid licences are planned.¹

The discussion fits Nvidia’s larger attempt to sell a complete AI infrastructure stack rather than chips alone. Its OpenShell Reference Architecture combines computing infrastructure with controls intended to keep agents inside defined boundaries, monitor their activity and quarantine an agent that behaves dangerously, according to Nvidia’s own announcement.² The company has also described Open Agent Safety Platform as part of its wider approach to agent security.

For customers, the practical question is what the licence would actually include. A per-GPU charge could bundle runtime software, security tools, management features or support, but Nvidia has not yet set out those terms. It is also not clear whether organisations would pay for every GPU in a system or only the hardware assigned to agent workloads. Until those details arrive, the $2,000 figure should be treated as a reported indication of direction rather than a confirmed price.

Why this matters

Nvidia’s software ambitions are significant because agent systems need more than a model: they need access controls, data connections, execution environments and monitoring. Turning those layers into a recurring per-GPU charge would give Nvidia a new stream of revenue while tying customers more closely to its hardware and infrastructure ecosystem. It would also make licensing costs a larger part of the total bill for companies building AI factories.

The proposal is not yet a product announcement with a launch date. It is an early signal that Nvidia sees the software around AI hardware becoming commercially important, while the exact relationship between a future $2,000 licence and AI Enterprise remains unresolved.

Our opinion

Nvidia is right to treat agent infrastructure as more than a chip sale, but a per-GPU price needs unusually clear boundaries. Customers will want to know whether they are paying for software they already licence, security controls they can obtain elsewhere or a genuinely new operating layer. Until Nvidia publishes that breakdown, the headline number is interesting chiefly as a statement of intent: the company wants recurring software revenue, and it is testing how much the market will accept.