Strike Nvidia's $2,000-per-GPU agent fee out of your contract before it becomes standard

At GTC Berlin on 22 October, Jensen Huang said that running the Nvidia software behind agents on what he calls AI factories costs about $2,000 per GPU per year. He offered it as an example. There was no product list, no start date and no explanation of how it relates to AI Enterprise, which Nvidia's own licensing guide lists at $4,500 per GPU for a one-year subscription with support. I think buyers should treat that offhand number as an opening bid and kill it in their hardware contracts now, while they still hold the cheque.

I expected the analysts to start multiplying, and they did. FourWeekMBA modelled five million paid licences producing $10bn a year and $50bn over five years. What I did not expect was how plainly Huang stated the intent: "Software licensing will become increasingly a major part of our company's business." Read that as a buyer. The meter runs on the number of GPUs you bought, whether or not an agent runs on them that day. Hardware you paid for up front turns into a recurring bill indexed to your own capacity, and every expansion you approve raises it automatically.

Nvidia chose to meter per GPU because it controls that number. The fee grows when you grow, it needs no usage audit, and any feature Nvidia ships later can be folded into the same licence to justify a higher rate at renewal. Procurement teams spent a decade learning what seat-based software licences do to a budget once they are embedded. A per-chip licence on infrastructure you cannot easily swap out is the same trap with a bigger unit price.

Nvidia's own price list tells you how soft these numbers are. AI Enterprise is listed at $4,500 per GPU for one year or $18,000 for five, and the same guide carries a $1,125 rate for education and Inception customers. A list price that drops by three quarters for the right kind of customer is a negotiating position. The $2,000 agent figure is softer still, because nobody outside Nvidia knows what it covers or whether it stacks on top of AI Enterprise.

Look at what is actually being sold. OpenShell, the runtime that sandboxes agents, is open source. The pieces that look billable are Nvidia Sentry running on BlueField-4 DPUs, which watches agent activity and can quarantine an agent that oversteps, plus support. Meanwhile Anthropic is wiring OpenShell and BlueField into Claude Managed Agents, and Salesforce has built a Slack Admin Bridge so administrators can approve agent permissions from Slack. Nvidia gets partners to make the free runtime a standard, then charges per chip for the monitoring and accountability enterprises will feel unable to skip. It is a clever strategy, and buyers should refuse to fund it on Nvidia's terms.

The case for paying is reasonable. Agents that use tools, touch data and execute tasks need security and governance, and a large company wants one vendor to call when something breaks. TechSpot pointed out that every million licensed GPUs brings Nvidia $2bn a year, small change next to the $96.2bn it booked last quarter. So why fight over it? Because that asymmetry favours the buyer. A fee that barely moves Nvidia's quarter is a fee an account team chasing a large Vera Rubin order will concede to close. For you, the fee grows with every rack you add and ties your agent guardrails to BlueField hardware, which makes evaluating a rival chip more expensive.

Nvidia has been heading here for years. A secondary source reports Nvidia's CFO claiming a $1bn annualised run rate in software and services back in the fourth quarter of fiscal 2023, with AI Enterprise at the centre. Huang told the BG2Pod podcast that people may one day stop thinking of Nvidia as a chip company. A company that says this wants recurring revenue from hardware it has already sold, which means from your data centre.

If you have a Vera Rubin or BlueField-4 order in negotiation this quarter, add a clause stating that all Nvidia software required to run and secure agents on those GPUs is included for the hardware's service life, or capped at your existing AI Enterprise terms, with no per-GPU uplift when you expand. If the account team will not sign it, you have learned that the $2,000 is coming to your bill, and you still have time to price an alternative before the hardware lands.

Prompted by Nvidia discusses $2,000-per-GPU annual fee for software that 'runs agents on AI factories': details unclear about potential overlaps, firm already charges more than twice that amount for its AI Enterprise software platform, tomshardware.com.