Nvidia CEO Jensen Huang has put a rough price on the software side of its AI factories: "about $2,000 per GPU per year." He said it at GTC Berlin on October 22, 2026, according to Tom's Hardware's report, which quotes a video of the comments. Social posts quickly turned it into a $50 billion revenue story. This post separates what Huang said from what others calculated and from what Nvidia has not told anyone.
TL;DR: said, calculated, unconfirmed
| Item | Status |
|---|---|
| Huang: about $2,000 per GPU per year to run Nvidia's agent software | Said on video at GTC Berlin, October 22 (quoted by Tom's Hardware) |
| Software licensing "will become increasingly a major part" of Nvidia's business | Said by Huang, same remarks |
| $10B per year, $50B over five years | Third-party scenario assuming five million paid licenses; not an Nvidia forecast |
| Which software is covered | Not disclosed |
| Whether $2,000 is a standard price | Not disclosed |
| Start date | Not disclosed |
| Relationship to AI Enterprise ($4,500 per GPU per year list) | Not confirmed; not a confirmed price cut |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What exactly did Huang say?
The quote, as reproduced by Tom's Hardware: "Software licensing will become increasingly a major part of our company's business. The way that we license will be... for example, it's about $2,000 per GPU per year to run all the software that Nvidia provides that runs agents on these AI factories."
Two details matter. "For example" signals an illustration, not a price sheet. And "all the software that Nvidia provides that runs agents" describes a bundle without naming its contents. Neither a press release nor a pricing page has appeared, so what exists is a spoken remark, reported by one outlet from a video.
Where the $50 billion comes from
Tom's Hardware attributes the big number to a proposed scenario from the analysis site FourWeekMBA: at $2,000 per GPU, five million paid licenses would bring in $10 billion a year, or $50 billion over five years if customers keep renewing. A separate post on X from an investor, embedded in the article, assumes 10 million eligible enterprise GPUs, 50% paid attach and the same price to reach the same $10 billion.
Those are assumptions stacked on a single spoken number. Nvidia has not said how many GPUs would be covered. The arithmetic is correct; the premise is a guess. Treat "$50 billion" as an illustration of scale, not guidance.
How it compares with AI Enterprise pricing
Nvidia already sells enterprise software. Tom's Hardware reports that its AI Enterprise platform is listed at $4,500 per GPU for a one-year subscription, which is more than twice Huang's figure. The outlet is careful: Nvidia has not confirmed the two prices cover the same products or terms, so the $2,000 should not be read as a cut. It could be a different bundle, a volume rate, a new agent-focused tier, or a loose figure.
For anyone budgeting a cluster, the practical reading is: expect software to be priced per GPU per year, and expect it to be a line item next to hardware. Do not plug $2,000 into a model until Nvidia publishes terms.
Why the agent angle: OpenShell and Sentry
A gripper working inside a transparent dome, illustrating how Nvidia software fee tied to agent sandboxes
The timing fits a broader push. On October 21, Nvidia announced its OpenShell Reference Architecture, which it describes as a full-stack design for building, containing and monitoring agents. Per Nvidia's announcement, Vera Rubin systems handle inference, Vera CPUs run agent harnesses and tools, and BlueField-4 DPUs run the Nvidia Sentry framework that monitors and enforces limits. Tom's Hardware adds that Sentry watches agent activity independently and can quarantine an agent that crosses its boundaries.
We have covered the pieces in detail: the Open Agent Safety Platform from September, the OpenShell GitHub project and what developers said on Hacker News, a technical look at how Sentry works on BlueField-4, and the full GTC Berlin keynote recap.
Important caveat: Nvidia has not said that OpenShell or Sentry are what the $2,000 pays for. Linking them is an inference from the "runs agents" wording and the timing. Tom's Hardware also notes that partners are building on the stack: Anthropic is integrating OpenShell and BlueField with Claude Managed Agents, and Salesforce built a Slack Admin Bridge for OpenShell.
Why Nvidia wants software revenue
Huang has been telling investors for months that Nvidia is not just a chip company. Tom's Hardware quotes him on the BG2Pod podcast: "One of these days, people might not categorize us as a chip company anymore." Software licensing is a recurring revenue line that does not depend on shipping a new GPU generation, and it deepens the dependence on Nvidia's stack, which readers will recognize from the CUDA debate.
Reaction in the Tom's Hardware comments, which are opinion rather than evidence, ranged from reading it as a sign of a bubble to calls to break Nvidia up. The Hacker News thread had 12 points and fewer than ten comments when we looked, too thin to draw conclusions.
What this means for what you build or pay
- If you rent GPUs from a cloud or neocloud: software costs are usually bundled into the hourly rate. A per-GPU license could surface as a separate line or as a price increase. Ask your provider whether Nvidia software fees will be passed through.
- If you buy hardware: budget for a software subscription on top of the capital cost, and ask your Nvidia rep for the bundle contents in writing. Compare against the Vera Rubin NVL72 efficiency claims to see whether the total cost of ownership story still holds.
- If you run agents at scale: open-source options exist for sandboxing and scheduling, such as Nvidia's own open-source KAI scheduler work. The question is which pieces will stay free and which move behind the license.
- If you secure agents: infrastructure-level enforcement is one layer. For runtime action controls on the agents themselves, AgentBeam is the agent security platform from the explainx.ai team, built to stop AI agents before they take dangerous actions.
A worked example: what $2,000 per GPU means for a cluster
Illustrative arithmetic only, using Huang's quoted figure as a placeholder and not a confirmed price. A 72-GPU rack, the size of the NVL72 systems Nvidia sells, would carry $144,000 per year in software if every GPU were licensed at $2,000. A 1,000-GPU cluster would carry $2 million per year, and a 10,000-GPU deployment $20 million. Over a typical three-to-five-year depreciation window, the license can add a meaningful fraction on top of the hardware bill, which is why buyers will want the bundle contents spelled out.
Compare that with the AI Enterprise list price reported by Tom's Hardware, $4,500 per GPU per year: the same 1,000-GPU cluster would be $4.5 million per year. If the $2,000 tier replaced AI Enterprise for agent workloads, customers would save; if it stacks on top, the effective bill would be $6,500 per GPU. Nobody outside Nvidia knows which, and that gap of $4,500 per GPU per year across a large fleet is exactly why scope matters more than the headline.
Per-GPU licensing in context
Per-seat and per-device licensing is how enterprise software has long worked: you pay for each unit that runs it. Nvidia's version attaches the meter to the accelerator rather than the user, which scales with deployment size and not with headcount. For agent workloads this is a notable choice, because agents multiply work without multiplying people. A per-GPU model charges for capacity, not for how many agents run on it.
The alternative models readers will see elsewhere are usage-based (per token or per sandbox-second) and outcome-based pricing. Our timeline of how token prices have fallen shows the model providers moving the opposite direction at the top of the stack: cheaper per token while the infrastructure vendor looks for recurring fees underneath. If you are weighing a hybrid of open tooling and vendor support, our Nvidia Mission Control 2.5 coverage shows another example of Nvidia packaging operations software around the cluster.
What to watch next
- A pricing or licensing page from Nvidia naming the bundle and the price.
- Whether the figure is a list price, a volume rate, or an illustration.
- How it maps to AI Enterprise, and whether existing customers migrate.
- Nvidia's next earnings call, where software revenue commentary would be the first firm data.
- Cloud providers' statements on whether they will pass through the fee.
How to check claims like this yourself
When a CEO quotes a number on stage, look for four things: a written source from the company, a defined scope, a date, and a unit. The $2,000 has a unit (per GPU per year) and a source (Huang on video) but no scope or date. The downstream $50 billion has none of the four. That is why we label it a scenario. Our earlier claimed-vs-verified posts on Nvidia events follow the same method.
Bottom line
Huang gave a ballpark: about $2,000 per GPU per year for the software that runs agents on AI factories. Everything built on it, the $10 billion and $50 billion annual and five-year figures, is someone else's arithmetic. Scope, timing, and the relationship to the $4,500 AI Enterprise list price are open. We will add an update section when Nvidia publishes terms.
Related reading
- GTC Berlin 2026 keynote: everything announced
- Nvidia Open Agent Safety Platform: OpenShell and Sentry
- Nvidia Sentry on BlueField-4: how it works
- OpenShell on GitHub and Hacker News
- Vera Rubin NVL72 claims, checked
- Nvidia run:ai and the KAI scheduler
Figures are as reported on October 26, 2026. Nvidia has not published pricing terms for the quoted fee.