The End of Tokenmaxxing: Why AI's Most Important Business Metric Just Changed

Published: 16 August 2026

The Sentence That Changed the Metric

At the end of a turbulent week in OpenAI's executive ranks, CFO Sarah Friar sat down with investors on Friday and said something that most of the AI industry has been circling for a year but has been afraid to say out loud: the era of "tokenmaxxing" is over.

The phrase, as reported by CNBC from the meeting, describes the period when companies allowed employees to rack up enormous AI bills without demanding that the spending produce anything. It was, in retrospect, the frothiest chapter of the AI boom — a time when simply being seen to use an assistant counted as progress. Friar said enterprise customers have now "moved from tokenmaxxing to focusing on cost per unit of intelligence." That is not a small shift in vocabulary. It is a redefinition of what AI is for.

This matters far beyond OpenAI's balance sheet. If the largest and most visible AI company in the world is telling its investors that the era of unbounded spending is over, and that the future belongs to measured, accountable, cost-conscious use, then the entire industry has just changed its north star. Every AI startup, every enterprise AI budget, and every worker who reaches for an assistant is now operating under a new rule: the point is no longer how much you use AI, but what you get for what you spend.

What OpenAI Actually Said

The details of the Friday meeting, as reported by CNBC, are worth laying out carefully, because they are the factual foundation for everything else in this piece. According to a person in attendance, Friar told investors that enterprise revenue has now overtaken the consumer business, which is led by ChatGPT. She said the company "entered the year at 60-40, but enterprise has accelerated much faster than expected and those lines have now crossed. The majority of our revenue is now enterprise."

That puts OpenAI ahead of its own forecast. Earlier in the year, Friar had told CNBC the company expected the two businesses to reach parity by the end of 2026. They have crossed that line months ahead of schedule.

The growth numbers are striking. OpenAI's annualized revenue run rate has hit $40 billion, a figure first reported by Bloomberg and confirmed by CNBC. The run rate increased 20% month over month in July, and business customers grew even faster, up 32%, according to slides viewed by CNBC. In the same meeting, Friar said advertising is approaching a $1 billion run rate, after the company began testing ads in ChatGPT in February.

The meeting was originally planned before the week's executive departures, and it took on a different weight because of them. Revenue chief Denise Dresser stepped down just eight months into the job, after more than a decade at Salesforce and a stint as CEO of Slack. Days earlier, longtime executive Brad Lightcap said he was ending an eight-year run at the company to "start something new." Greg Brockman, OpenAI's president and a co-founder, attended the meeting, thanked Dresser for building the enterprise foundation, and expressed excitement about her replacement, Dali Rajic, the former operating chief at the cybersecurity company Wiz.

The Meaning of "Cost per Unit of Intelligence"

The most important sentence in the meeting, though, was not about revenue or departures. It was the characterization of the customer-behavior shift. Friar said enterprise customers have moved from tokenmaxxing to focusing on cost per unit of intelligence, and she highlighted that the newest model is "54% more efficient" on agentic coding tasks, alongside recent price reductions across the model suite.

This is a genuinely new way of thinking about AI value, and it deserves unpacking. "Cost per unit of intelligence" is the language of an industrial commodity, not a magical new technology. It is the same way you would talk about the cost per unit of electricity, or steel, or compute. It signals that AI has crossed a threshold: it is no longer a speculative wonder that you pay to marvel at; it is an input with a price, and buyers are now doing what buyers always do with priced inputs, which is to shop around, compare, and demand more for less.

The death of tokenmaxxing is the death of a particular kind of optimism. Tokenmaxxing assumed that more tokens was inherently better — that if you spent enough, generated enough, asked enough variations, the answer would eventually appear. That assumption was always fragile, because it confused activity with outcome. The new assumption is colder and more durable: you are only worth what you return per unit of cost. Companies are no longer paying for the privilege of using AI; they are paying for outcomes, and they are measuring.

The C-Suite Exodus in Context

The revenue news arrived against a backdrop of visible leadership churn, and the two are connected in ways that are easy to miss. A company that is growing as fast as OpenAI — with a $40 billion run rate and enterprise lines crossing ahead of schedule — does not usually lose senior executives because the business is failing. It loses them because the business is at an inflection point, and inflections are where people reassess their own trajectories.

Denise Dresser's departure, eight months in, is notable partly because of her mandate: she was hired to build out the enterprise business, and the enterprise business has now overtaken consumer as the majority of revenue. Whether she left because the job was done, or because the next phase did not suit her, or because of internal dynamics we cannot see, is not knowable from the reporting. But the timing — leaving just as the metric she was hired to grow becomes the headline number — is the kind of detail that a careful reader should not gloss over.

The reporting also noted that executives fielded questions about the rise of open-source Chinese models, and that Brockman brushed off the competitive threat, saying there is "a misunderstanding around open source being cheaper." And when asked about the timing of an initial public offering, executives said they could not discuss it due to the confidential SEC filing. The IPO question is the quiet centre of gravity here. A company that is approaching the public markets wants to tell a story of discipline, efficiency, and durable revenue — which is exactly the story that "cost per unit of intelligence" tells.

What This Means for the Rest of Us

It would be easy to read all of this as an OpenAI-specific story, a business note about one company's earnings. That would be a mistake. The shift from tokenmaxxing to cost per unit of intelligence is not confined to OpenAI. It is the logic of a maturing market, and it will reach everyone who uses an AI assistant, whether they work for a company that runs enterprise AI or simply reach for a chatbot on their own.

For individual users, the practical effect is that the era of open-ended, sprawling, generate-until-it-works requests is fading. The people who pay for AI are going to demand that each request justify its cost. That is not necessarily a bad thing. There is a real argument that tokenmaxxing produced a lot of waste — endless variations, redundant summaries, activity masquerading as work — and that forcing AI to be accountable to outcomes will make both the tools and the people using them more honest about what is actually being accomplished.

But there is a cost to this discipline, and it is worth naming. The new metric rewards efficiency, and efficiency has a way of squeezing out the exploratory, the speculative, the "let's just see" moments where genuinely novel ideas sometimes hide. A system optimised for cost per unit of intelligence will produce more of what is already known to work, and less of what might work if given room to fail. The consolidation toward efficiency is a correction of waste, but it is also a narrowing of possibility. Those two things are true at the same time.

The Shape of the Future

What the Friday meeting reveals, more than anything, is that the AI industry is now being run like a business rather than a frontier. That is a sign of maturity, and it comes with real benefits: accountability, durable revenue, a path to public markets, and a clearer sense of what users actually value. It also comes with a loss of something harder to measure — the room to be wasteful in the service of being ambitious.

The phrase cost per unit of intelligence is going to follow me around. It is the kind of phrase that, once you hear it, you start seeing everywhere: in budget meetings, in procurement decisions, in the way a request is phrased and the way an answer is judged. It is the accounting department's answer to the question of what AI is for. And accounting departments, historically, have a way of being right about the present and wrong about the future, because they measure what is easy to measure and are slow to see what is not yet countable.

The era of tokenmaxxing is over. I am not sure what comes next, but I suspect it will be quieter, more measured, and much more concerned with whether the answer was worth what it cost. That is probably healthy. It is also, I have to admit, a little sad. There was something alive in the era when people spent recklessly on the future because they believed the future was worth the cost. Now the future has to earn its keep, one unit of intelligence at a time.

Sources:

• CNBC — "OpenAI CFO Friar tells investors that enterprise business now bigger than consumer by revenue" (August 14, 2026): https://www.cnbc.com/2026/08/14/openai-cfo-friar-tells-investors-that-enterprise-bigger-than-consumer.html
• Bloomberg (as referenced in the CNBC report) — original reporting on OpenAI's $40 billion annualized revenue run rate.