TL;DR · 30-second read
The Short Version
Nvidia, whose chips power much of today’s artificial intelligence, says software subscriptions will become a bigger part of its business.
Its chief executive, Jensen Huang, said running Nvidia’s software for artificial intelligence “agents” (programs that carry out tasks on their own) costs about $2,000 per chip, every year. Nvidia already lists another software package at $4,500 per chip per year.
Nvidia has not said whether buyers would pay one price or both. For a company running 10,000 chips, that means a yearly software bill anywhere from $20 million to $65 million.
Nvidia CEO Jensen Huang said at the company’s GTC Berlin event on October 22 that it costs “about $2,000 per GPU per year to run all the software that Nvidia provides that runs agents on these AI factories,” Tom’s Hardware reported. He added that “software licensing will become increasingly a major part of our company’s business.”
Nvidia has not said when the price would take effect, which software it covers, or how it relates to Nvidia AI Enterprise, which is listed at $4,500 per GPU for a one-year subscription. The remarks came a day after Nvidia announced its OpenShell Reference Architecture, a hardware-and-software design for running AI agents securely in enterprise data centers.
Executive Summary
Nvidia’s chief executive has attached a price to the software layer that runs AI agents, meaning AI systems that use tools, access data and carry out tasks rather than only answering questions. The figure, about $2,000 per GPU per year, is a recurring charge on each graphics processor (GPU), the chip at the center of modern AI computing. It is also the clearest signal yet that Nvidia intends to earn subscription revenue from every chip it ships, not only a one-time sale.
The figure matters less for its size than for what it leaves open. Nvidia already sells AI Enterprise at a $4,500 per-GPU annual list price, and it has not said whether the agent software replaces, sits inside, or stacks on top of that subscription. Until it does, the software line in an AI data center budget can only be modeled as a range. At list prices, that range runs from $2,000 to $6,500 per GPU per year.
The pricing is tied to Nvidia’s push into agent security. OpenShell and its Sentry monitoring software put guardrails for AI agents into the data center infrastructure itself. Anthropic and Salesforce are already building integrations on that platform.
One Quote, Three Possible Budget Lines
Huang’s figure is priced per GPU, per year. That structure makes software behave like power or cooling: a cost that grows with the size of the fleet and recurs for as long as the hardware runs, rather than a one-time entry in a capital budget. Nvidia already sells software this way. AI Enterprise, its subscription suite for running AI on Nvidia hardware, is listed at $4,500 per GPU for one year. Nvidia has not said how the two prices relate, and that is where buyers’ planning gets difficult.
The quote supports three readings. If the agent software replaces or repackages AI Enterprise, the list cost per GPU falls to $2,000. If Huang was describing software already included in AI Enterprise, the list cost stays at $4,500. If the agent runtime is licensed separately and added on top, the cost reaches $6,500, more than three times the low case. For an illustrative 10,000-GPU deployment, that works out to $20 million, $45 million or $65 million a year at list, renewing annually. Nvidia has not confirmed the $2,000 figure as a reduction from $4,500, so it should not be read as a price cut.
Three groups carry this uncertainty. Enterprise buyers need total cost of ownership, the all-in cost of hardware, software, power and support over a system’s life, to justify their purchases. GPU cloud providers have to decide whether and how to pass a per-GPU software charge into rental prices. Finance teams comparing an Nvidia stack against alternatives need a firm software figure to do so. Negotiated discounts may land below list, but a threefold spread in a recurring line item is too wide to use as a planning assumption.
Why Nvidia Is Pricing Software by the Chip
Huang has described Nvidia as more than a chip supplier. “One of these days, people might not categorize us as a chip company anymore,” he said on the BG2Pod podcast, pointing to the company’s full technology stacks in healthcare, automotive, telecommunications and life sciences. Per-GPU licensing links that ambition directly to hardware volume. Each GPU sold becomes a potential annual subscription, and each year of shipments adds to the base of licensed chips.
The arithmetic explains the appeal. A FourWeekMBA scenario assumes 10 million eligible enterprise GPUs shipped in a year, half of them with paid software at $2,000 each. That produces $10 billion in annual recurring revenue, or $50 billion over five years if every licence renews. Each of those inputs is an assumption, not Nvidia guidance. The result moves in direct proportion to the share of GPUs that carry a paid licence (the attach rate) and to how many customers renew. A lower attach rate or customer churn shrinks it quickly.
The same mechanism also determines what customers pay. Recurring revenue for Nvidia is a recurring cost for whoever owns the GPUs. As fleets grow from thousands to hundreds of thousands of chips, a per-GPU software fee becomes one of the few infrastructure costs that scales exactly with the number of chips, with no economies of scale built in unless Nvidia offers volume terms.
Agent Security Is Moving Into the Hardware
The pricing comes as Nvidia builds out infrastructure for managing AI agents. Because agents can call tools, read data and take actions, a mistake or a manipulated instruction has consequences beyond a wrong answer. Nvidia’s OpenShell Reference Architecture, announced October 21, combines Vera Rubin computing infrastructure, Vera CPUs and BlueField-4 data processing units (DPUs). DPUs are specialized processors that handle networking and security tasks separately from the main chips. They run Nvidia Sentry, which according to Nvidia monitors agent activity independently and can quarantine an agent that tries to cross its permitted boundaries.
The design choice is to enforce controls in the infrastructure rather than rely only on safeguards inside the AI model or application. It builds on the Open Agent Safety Platform that Nvidia introduced in September. It is also gaining partners. Anthropic is integrating OpenShell and BlueField hardware with Claude Managed Agents to control what agents can access inside execution sandboxes. Salesforce has built a Slack Admin Bridge that lets administrators review, approve, reject and monitor agent permission requests in Slack.
For buyers, the trade-off is integration versus portability. Enforcement anchored in BlueField hardware and Nvidia software works across the stack, but it also ties a business’s agent governance to Nvidia infrastructure. Nvidia has not confirmed that the $2,000 licence covers OpenShell or Sentry. Whether these security features come inside an existing subscription or as a new line item will decide how much of this architecture enterprises actually adopt.
Background
Nvidia designs GPUs, processors originally built for computer graphics that became the workhorse chips for training and running AI models. In recent years the company has expanded into networking, complete server systems, enterprise software and data center infrastructure. It now sells full technology stacks into industries such as healthcare, automotive, telecommunications and life sciences. Vera Rubin is its next-generation computing platform, and BlueField is its line of data processing units, chips that take over networking and security work from the main processors.
Nvidia already charges for software on a subscription basis. Nvidia AI Enterprise, its suite for businesses running AI on Nvidia hardware, is listed at $4,500 per GPU per year. In September the company introduced its Open Agent Safety Platform, and in October it followed with the OpenShell Reference Architecture, which extends agent security controls into enterprise infrastructure and edge systems. Source: Nvidia discusses $2,000-per-GPU annual fee for software that ‘runs agents on AI factories’ (Tom’s Hardware): Jensen Huang’s GTC Berlin remarks on per-GPU software licensing and how the figure compares with Nvidia AI Enterprise pricing.Sources