Nvidia Floats $2,000 a Year per GPU for Agent Software

Jensen Huang put a number on Nvidia's software ambitions at GTC Berlin on 22 October 2026: about $2,000 per GPU per year to run the software that operates AI agents on what the company calls AI factories. He offered it during a question and answer session, as an illustration rather than a price list. The Register reported the remark the same day.

The figure is strange in one direction. Nvidia already sells AI Enterprise at a list price of $4,500 per GPU for a one-year subscription, more than twice what Huang quoted. The company has not said whether the two cover the same products or the same licensing terms, which is why the $2,000 should not be read as a price cut on anything.

What has not been said

Almost everything that would make the number usable is missing. Nvidia has not disclosed how many GPUs would fall under the proposed fee, whether $2,000 is a standard price or an example, which software it would cover, how it would overlap with AI Enterprise, or when it might start. The arithmetic circulating since Huang spoke, five million paid licences producing $10 billion a year and $50 billion over five years, comes from FourWeekMBA rather than from Nvidia, and rests on a licence count nobody outside the company has.

Scale is the reason a software line item registers at all. The Register, reporting the GTC Berlin remarks on 22 October 2026, put Nvidia's fiscal 2026 revenue at $215.9 billion and net profit at $120 billion, and noted that net profits in the first half of fiscal 2027 were already close to the whole of the prior year. Against that, a recurring per-GPU fee is a margin question rather than a survival one.

Why the billing unit matters

Charging per GPU per year does something different from charging for a chip once. It turns an Nvidia estate into a recurring cost that scales with fleet size rather than with how hard the fleet is worked, and it arrives while large buyers are deciding whether to stay on Nvidia silicon. Huang has been arguing the broader case for a while, telling the BG2Pod podcast that "One of these days, people might not categorize us as a chip company anymore", and pointing to Nvidia's reach into networking, systems, enterprise software and AI agent security across healthcare, automotive, telecommunications and life sciences.

Microsoft unveiled Maia 300 on 12 October 2026 and said it would scale its in-house accelerator family to multi-gigawatt deployments from 2027, as Parallax Nexus reported. An annual per-GPU software charge is one more entry on the side of the ledger that custom silicon is meant to avoid, although Nvidia's software is also a large part of what makes its hardware awkward to leave.

A trial balloon with a chosen shape

The honest reading is that this was a trial balloon rather than a launch. A chief executive who meant to announce pricing would not do it in a Q&A, in a round number, below the list price of a product the company already sells, without naming the product. What cuts against dismissing it is that Huang picked the per-GPU, per-year unit out loud, in front of European customers, at an event built around sovereign AI infrastructure. Companies rarely choose a billing unit by accident.

GTC Berlin ran from 20 to 22 October 2026, with Huang's keynote on 21 October and the pricing remark coming the day after. Nvidia has published the keynote in full. It has not confirmed the $2,000 figure as pricing, and the reports that followed cite the Q&A remark and nothing else from the company.