OpenAI and Nvidia want AI buyers to measure ‘intelligence per dollar’

Two essays, one phrase

The AI industry may have found its next favourite measure: intelligence per dollar.

OpenAI and Nvidia both used the phrase in articles published on Friday. The timing is notable. The definitions are not the same.

OpenAI's chief financial officer, Sarah Friar, proposed 'Useful Intelligence per Dollar' as a scorecard for companies buying AI. Nvidia used 'intelligence per dollar' to describe the economics of continuously refining agent models on its hardware and software.

Both arguments are trying to move attention away from a simple price per token. That is sensible. A cheap answer is not cheap if it fails, needs three retries and takes a person twenty minutes to repair.

But the shared phrase can make the agreement look stronger than it is. OpenAI starts at the business outcome. Nvidia starts much closer to the compute stack. Neither post defines an industry standard that lets a buyer compare one provider with another.

OpenAI starts with the finished work

OpenAI's framework asks four questions. Did the system complete useful work? What did each successful task cost? How often was the result dependable? And did each dollar produce more value as usage grew?

The important unit is a successful task, not a token. A support team might count resolved cases. An engineering team might count code changes that pass tests. A finance team might count forecasts prepared to an agreed quality bar.

The full cost should include the model, employee time, human review, retries and rework. Divide that amount by the tasks that actually met the bar, and the cheapest model on a price list may no longer be the cheapest system to use.

This is a useful discipline. It is not a finished metric. 'Useful work' depends on the organisation, and quality bars can be adjusted to make a result look better or worse. The value of a reviewed contract is also difficult to compare with the value of a resolved support ticket.

OpenAI illustrates the argument with its own products and GPT-5.6 model family. Those examples are part of a sales case. The underlying advice can still be applied to another provider, or to a workflow with no OpenAI model in it.

Nvidia starts in the machine room

Nvidia uses the same words for a different layer of the system.

Its article focuses on post-training: the work that turns a pretrained model into one that can follow instructions, use tools and improve through reinforcement learning. For agents, Nvidia argues, this is becoming a continuous loop rather than a final stage completed once before release.

The company places cost per token inside a broader idea of intelligence per dollar. Cheaper inference makes it less expensive to generate the many attempts used during reinforcement learning. More efficient training hardware then lowers the cost of updating the model from those attempts.

Nvidia points to its Nemotron 3 Ultra model and Vera Rubin platform as examples. It says Vera Rubin can train the largest models with one quarter of the GPUs needed by the previous Blackwell generation. That is a Nvidia claim about a platform not yet independently tested in the conditions described in the post.

Here, intelligence per dollar is largely a measure of how economically a model can be improved and kept current. OpenAI's version asks whether the resulting system completed work a customer values. The concepts touch, but they are not interchangeable.

It is not a standard yet

A real cross-industry metric would need an agreed numerator and denominator.

What counts as intelligence: a benchmark score, a successful task, revenue gained, time saved or errors avoided? What counts as cost: API fees, chips, energy, staff time, integration, security review and the failures that never reach a customer?

It would also need rules for quality. A system that finishes twice as many tasks is not more valuable if its mistakes are expensive or hard to detect. Dependability cannot sit in a footnote.

Neither company provides a common test suite, an audit method or a way to compare unlike workflows. There is no neutral conversion from Nvidia's post-training efficiency to OpenAI's completed business task.

That does not make the phrase empty. It means it is a direction of travel, and currently a marketing claim, rather than a number buyers can request in a procurement table.

A buyer can still use the idea

The useful version is local and a little boring.

Choose one recurring workflow. Write down what a good result looks like before testing models. Count every attempt, correction and escalation. Include the time people spend checking the work, not only the invoice from the model provider. Then run the same measurement again after the system changes.

This will not produce a universal intelligence score. It can reveal whether an AI deployment is improving inside one organisation, which is usually the decision that matters.

Infrastructure teams can do the same one layer down: measure how much it costs to reach a fixed model-quality target, rather than celebrating the largest training run or the lowest isolated token price.

OpenAI and Nvidia are right about the weakness of the old comparison. Tokens are an input. Companies buy outcomes.

The next step is harder: turning a neat phrase into a measurement that survives outside the company selling it.