Nvidia is moving toward a new financial strategy aimed at reducing reliance on one-time hardware sales and focusing instead on sustainable cash flows through software licensing. During the GTC Berlin conference, CEO Jensen Huang indicated that software licensing would become a central part of the company's business, citing $2,000 as a model for an annual fee per GPU to run AI agents.
These fees are intended to cover the costs of the software required to run agents within what the company calls "AI factories," which are data centers specifically designed for AI workloads. Interestingly, this figure ($2,000) is lower than the current "Nvidia AI Enterprise" package pricing, which costs approximately $4,500 per year per GPU, or $3,600 per year when committing to a five-year contract. The company has not yet clarified the relationship between these prices or the actual implementation date for this new pricing.
Development of Infrastructure and Secure Software
In parallel with this financial direction, Nvidia is accelerating the development of agent software, having recently launched the "OpenShell" reference architecture. This platform provides a sandbox environment that allows agents to operate using "Vera" processors with granular control over their access to files, networks, and available tools. Additionally, "BlueField-4" network chips, supported by "Sentry" software, contribute to monitoring these agents from the outside, with the ability to isolate any agent that exceeds permitted boundaries within fractions of a second, thereby enhancing security in these complex environments.
What Changes Practically?
This shift represents an attempt by Nvidia to secure recurring revenue that does not depend entirely on new hardware purchase cycles. If the company succeeds in licensing one million GPUs at this price, it would generate $2 billion annually in recurring revenue, a massive figure to add to its multi-billion dollar earnings. However, a fundamental technical and commercial question remains regarding how a "GPU" is defined for licensing purposes; while there were previous trends to count each silicon chip as an independent GPU (which could double costs for companies), it appears the company has returned to adopting the concept of complete units in its new systems such as "Vera Rubin," which will directly impact the total cost that data centers will bear.