NVIDIA Just Put a Price on Running Agents. It Is Per GPU, Per Year, and It Does Not Care Whether They Worked.

At GTC Berlin on 22 October, Jensen Huang put a number on something most companies have not thought to budget for. Running NVIDIA's agent software on AI factories, he said, costs about $2,000 per GPU per year. He went further and described software licensing as a major part of NVIDIA's business in the future.

The number is interesting. The shape of it is the real story.

What is actually confirmed, and what is not

Let us be careful here, because a lot of coverage has not been.

Confirmed: NVIDIA already sells software per GPU. NVIDIA AI Enterprise lists at $4,500 per GPU for a one-year subscription, or $18,000 per GPU for five years. The per-GPU software meter is not a thought experiment. It exists, it has a price list, and customers pay it today.

Not confirmed: almost everything about the $2,000. NVIDIA has not said what it covers, when it would apply, which GPUs qualify, or how it relates to the existing suite. It could be a narrower bundle, a future discount tier, or Huang rounding in conversation. Several outlets have extrapolated multi-billion dollar revenue projections from the remark. Those are assumptions stacked on a figure that has no product attached to it yet.

So this is not a pricing announcement. It is a senior executive saying out loud where he expects the money to come from. That is worth paying attention to for a different reason than a price list would be.

A per-GPU meter does not behave like a per-token one

Everybody costing out AI agents right now is thinking in tokens. Tokens are usage. If your agent sits idle for a quarter, usage-based billing is merciful: you pay close to nothing.

A per-GPU annual licence does not work like that. It attaches to hardware, not to work. You pay it whether your agents ran continuously or sat untouched through a slow summer. It is a floor rather than a slope.

That difference matters more than the absolute figure, because it changes what a bad quarter costs you. Usage billing punishes success, since the more your agents do, the more you pay. Capacity billing punishes hesitation, since the hardware is provisioned and the meter runs regardless. Most businesses have planned for the first and not the second.

And notice that these are not alternatives. If you run agents on your own GPUs with licensed software, you can end up paying both.

How many meters are on a single agent?

Here is the question worth taking into a budget meeting: if you added up every layer you pay to keep one agent working, how many separate meters would you count?

Tokens. The obvious one, and the only one most teams track.

Per-seat credits. Notion rebuilt its agents on 19 October with 1,500 Standard and 500 Select credits per seat per month. Credits do not roll over. The expensive pool is the one you touch when you pick your own model or run a custom agent, which is to say the pool you touch when you do anything interesting.

Per-GPU software. The layer Huang was talking about, if you run your own infrastructure.

The model provider's own floor. Minimums, commitments, enterprise tiers.

The person. The one nobody writes down. Somebody's job quietly became keeping the agent working, and that salary is the largest line in the stack by a distance.

Very few companies can produce that list. Almost none can produce the numbers next to it.

The week this landed in

The remark did not arrive in isolation. In the same handful of days, Dell expanded its AI Factory with NVIDIA for running agents on-premises rather than in public cloud. Akamai launched a managed inference service and an isolated sandbox for executing agent-generated code on its edge network. Oracle said Google's TPUs will reach Oracle Cloud Infrastructure in 2027, naming agentic and long-context workloads as the target.

Read together, that is an infrastructure layer being built out and monetised at the same time. The agent boom has reached the stage every platform shift reaches, where the people selling the shovels start itemising.

None of this is sinister. Infrastructure costs money and somebody has to charge for it. But the companies that get hurt in this phase are not the ones who pay. They are the ones who did not know what they were paying for until the invoice arrived.

What to do before somebody prices it for you

Count your meters. Write down every layer billing you for one agent. If the list takes more than ten minutes to assemble, that is the finding.

Separate your floor from your slope. Which costs continue if your agents do nothing next month? That number is your real commitment. The rest is variable and you can manage it.

Check what happens when you want out. Can you move your workload to a different model, a different provider, your own hardware? If the answer requires a conversation with an account team, you are not holding the leverage you think you are.

Price the person. If someone on your team spends a day a week keeping agents running, put a number on that day. It will almost certainly dwarf the per-GPU line you were worried about.

Why we built it the way we did

We have argued for bring your own key and self-hosting since the beginning, and the honest objection has always been that it sounds like ideology. Of course the vendor with a BYOA model says the meters matter.

Weeks like this one are why it is not ideology. When the layers multiply and the pricing for each one is announced in conversation at a conference rather than on a page you can read, the only genuinely defensible position is the one where you can see every meter and switch off the ones you do not want. Your own API keys mean the model bill is yours, transparent, and portable. Self-hosting means the infrastructure line is a number you chose rather than one you inherited.

That does not make your agents free. It makes them countable, which is the part that actually survives a budget review.

The useful question coming out of GTC Berlin is not whether $2,000 per GPU is a lot of money. For a company running a serious fleet, it may well be reasonable. The question is whether you would have noticed it on the invoice, and whether you could name every other line sitting above it.

Want to test the most advanced AI employees? Try it here: https://Geta.Team