OpenAI’s math breakthroughs may shift AI firms away from token sales

For years, the AI business model has looked a lot like a utility bill. You send a model text, it sends text back, and you pay by the token.

A Bloomberg column now floats a different future: the big AI labs might eventually step back from selling tokens at all. OpenAI’s recent string of claimed math results is the reason that idea no longer sounds far-fetched.

A quick definition first. In AI, a token is a small chunk of text, roughly a word fragment, that a model reads or writes. Labs meter usage in tokens and charge for them, much like a power company charges per kilowatt-hour.

The results OpenAI is pointing to

On September 8, 2026, OpenAI announced a proposed solution to part of the Navier-Stokes existence and smoothness problem. That problem is one of the Millennium Prize Problems, a short list of famously unsolved questions in mathematics.

The effort was not one chatbot thinking hard. OpenAI used approximately 10,000 agents that exchanged roughly 2.7 million messages over an 88-hour stretch.

The total bill came to about 130 billion tokens processed.

OpenAI followed up on October 6, 2026, by releasing 722 mathematical manuscripts. Those papers are grouped into 372 result families covering open problems in fields including algebra and topology.

Each result in that batch reportedly took around three hours of compute on average, measured as the equivalent of ChatGPT Pro processing.

The September and October releases came after an earlier claim. In August 2026, OpenAI said it had solved 10 long-standing mathematical problems for about $2,000 in tokens.

Tokenmaxxing and the case for keeping the output

The core argument in the Bloomberg piece is about where value lands. If a model can produce results worth far more than the compute it burns, selling that compute at a metered rate starts to look like underpricing.

Sam Altman appears to be experimenting with an answer. OpenAI’s CEO has been offering large allocations of API tokens to early-stage startups in exchange for equity, with grants such as $2 million in tokens.

The practice has picked up a nickname: “tokenmaxxing.” Instead of taking cash for compute, OpenAI takes a slice of the companies built on top of it.

Math is a useful test case because results are, at least in principle, checkable. A proof either holds or it does not, which makes it a cleaner demonstration of value than a chatbot answer that merely sounds right.

The credit fight in the background

The rapid pace has not gone over smoothly with everyone in academia. OpenAI’s fast publication of results has triggered disputes over credit and how findings get released.

Some results have been revised or withdrawn after conflicts among researchers over prior claims. NYU mathematician Tristan Buckmaster has been central to one dispute, saying he had shared earlier work with OpenAI researchers.

What this means for AI companies and their customers

For investors watching the AI sector, the question is whether token revenue remains the main engine or becomes a side business. Metered compute is predictable, but it pits labs against each other on price, and that tends to compress margins over time.

The tokenmaxxing deals are the most concrete sign of that shift so far. If equity-for-compute arrangements spread, startups could gain access to large amounts of model capacity without spending cash, while giving up ownership to the lab supplying it.

That trade-off deserves scrutiny from founders. A startup that pays in equity is tying its cap table to a supplier that may also become a competitor if the lab decides to keep the most valuable outputs for itself.

The Navier-Stokes result is a proposed partial solution, not a settled one, and the credit disputes show how contested AI-driven discovery remains.

Things to watch include whether independent mathematicians validate the September claim and whether more of the October manuscripts get revised or pulled.