Nvidia chief executive Jensen Huang said at GTC Berlin on October 22 that the company expects to charge about $2,000 per GPU per year to license "all the software that Nvidia provides that runs agents on these AI factories." Huang described software licensing as an "increasingly major part" of Nvidia's business, a pointed signal from a company still known chiefly for selling chips.
Nvidia has not said when the fee would take effect, how many GPUs it would cover, or exactly which products it includes. Those gaps matter, because Nvidia already sells an AI Enterprise software platform listed at $4,500 per GPU for a one-year subscription. The reported $2,000 figure is lower, but the company has not confirmed the two prices cover the same products, so it should not be read as a price cut. A scenario outlined by FourWeekMBA, in which five million paid licenses generate $10 billion a year, implies $50 billion over five years, but that rests on assumptions Nvidia has not endorsed.
The licensing push sits alongside new infrastructure aimed at running AI agents safely. On October 21 Nvidia announced its OpenShell reference architecture, which pairs its Vera Rubin systems and BlueField-4 data processing units with software called Sentry that monitors agent activity and can quarantine an agent that tries to exceed set boundaries. Anthropic is integrating OpenShell and BlueField hardware with its Claude Managed Agents, and Salesforce has built a Slack bridge that lets administrators approve or reject agent permissions. The design moves guardrails into the infrastructure rather than leaving them to the model or application.
Nvidia moved on a second front the same week, addressing the supply bottlenecks that cap how much hardware it can deliver. It announced substrate agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron that it says will raise substrate capacity 2.5-fold within three years, and a $1.1 billion prepayment to reserve added capacity at KYEC, which plans facilities in the United States. It also introduced DSX POD, a modular design for prefabricated Vera Rubin factories that it says could cut provisioning at prepared sites from six months to as little as 14 days, with partner shipments targeted for early 2027.
Every one of those figures carries a caveat. The substrate capacity is years out, the 14-day figure applies only to already-prepared sites, and the power and water savings Nvidia cited for its DSX platform are stated as targets rather than results. Demand is not the constraint: Reuters reported that ByteDance, Alibaba and Tencent together ordered more than three million RTX PRO 5500 cards worth over $35 billion, and Nvidia has said it cannot guarantee it will fill those orders in full.
The announcements landed in a jittery market. Nvidia and other AI-linked stocks came under selling pressure on October 23 amid worries about expensive debt financing for data centers and stretched valuations, the same anxiety behind AI Daily's report on distress in neocloud bonds that Nvidia helps backstop. Charging customers a recurring software fee could deepen Nvidia's revenue and widen its moat, or, as one commenter on the pricing news put it, pull more cash from buyers already loading up on debt. Which of those it turns out to be depends on numbers Nvidia has yet to publish.