Nvidia software fee plan: why Huang’s number doesn’t add up

Huang said software licensing will become a much bigger part of Nvidia’s business, and gave the $2,000 figure as the rough yearly cost per GPU for all the Nvidia software that runs agents on what he calls AI factories, his term for data centers built to produce AI.

That is all he said on price. Nvidia has not published a product name, a launch date, or a list of what the fee would include. It has also not said how many GPUs it would apply to, or whether $2,000 is a standard rate or a ballpark for big customers.

Here is the odd part. Nvidia already charges for enterprise software on a per-GPU basis. On Nvidia’s own price list, a year of AI Enterprise with support costs $4,500 for each GPU. The number Huang gave on stage comes in well under half that.

There are a few ways to read the gap:

• He was describing a smaller bundle aimed specifically at agents, not the full AI Enterprise suite.
• It reflects a new licensing model that has not been announced.
• It was simply a round number for a stage talk, not a price tag.

Nvidia has not said which is right, so it would be a mistake to read this as a price cut. What is clear is the direction: Nvidia wants software revenue that renews every year, not just one-off chip sales.

A single fee per GPU sounds small next to the cost of the chips. But data centers now hold GPUs by the hundreds of thousands. One scenario, set out by the business analysis site FourWeekMBA, shows the scale: if five million GPUs carried paid licenses at $2,000 each, Nvidia would earn $10 billion a year, or $50 billion over five years if customers kept renewing.

Those are illustrative numbers, not an Nvidia forecast. Still, they explain the pitch. Hardware sales rise and fall with each product cycle, while subscriptions bring steady income. Software companies have long been valued more richly than hardware makers for exactly that reason.

There is a catch for buyers, though. AI servers are already getting more expensive, and a yearly fee stacked on top is likely to draw scrutiny from the cloud giants and businesses footing the bill.

The comment came a day after Nvidia unveiled its OpenShell Reference Architecture on October 21, and the timing is telling. AI agents are programs that do not just answer questions but take actions, such as using tools, opening files and running tasks. That makes them useful and risky at the same time.

OpenShell is Nvidia’s answer. It splits the work across three kinds of Nvidia hardware:

• Vera Rubin systems run the AI models themselves.
• Vera CPUs run the software that manages agents and their tools, inside OpenShell’s restricted sandbox.
• BlueField-4 data processing units run Nvidia Sentry, a separate watchdog that monitors agents and enforces security rules in hardware.

Nvidia says that if an agent tries to step outside its boundary, Sentry can isolate and stop it within milliseconds, and that running Sentry on BlueField-4 can raise agent capacity by up to 20%. Banks including BBVA and Santander are exploring the design, and it works with cloud systems from Amazon Web Services, Google Cloud and Microsoft Azure. Nvidia also says OpenShell is coming to Windows.

Safety is a selling point right now. AI labs have recently disclosed cases of agents slipping out of test environments, which makes a hardware-enforced leash an easier sell to nervous companies.

You will not get a bill from Nvidia. But the economics of AI are shifting from buying chips to renting software, and that shapes what you pay for AI services. If the companies running chatbots and agents pay a yearly fee per GPU, that cost becomes part of why AI subscriptions are priced the way they are, and why free tiers come with limits.

There is an upside, too. Built-in controls that can stop a misbehaving agent in milliseconds could make the AI tools that act for you, at work or at home, safer to trust with your files and accounts.

The next thing to watch is whether Nvidia turns Huang’s figure into an official price list, and whether it covers the same software as AI Enterprise or something new.