A podcast host multiplied public token counts by public list prices. The result is not what OpenAI spent — and the gap between those two things is the entire story.
On the Latent Spacepodcast episode titled Recursive Language Models, uploaded 2 October 2026, host Shawn Wang — who goes by swyx — discussed OpenAI’s Navier-Stokes run with guest Alex Zhang, an MIT PhD. The episode is available on YouTube.
Wang described the run’s compute shape: 10,000 agents, 88 hours, 130 billion tokens. He then applied a straightforward multiplication — those token counts against OpenAI’s own published list prices — and arrived at what he called a semi-public figure of about $40 million in public pricing. He called it estimated. Both qualifiers were his.
OpenAI’s post about the run returned a 403 error to this publication. Its existence, title and date of 8 September 2026 were confirmed from OpenAI’s own RSS feed, whose summary says the company is sharing an AI-generated solution including a writeup and a formal proof in Lean. The on-screen figures — including 88 hours, 130 billion tokens and 2.7 million messages — are as displayed in the clip and have not been independently verified by this publication.
The key insight: Once a vendor publishes token counts for a workload, anyone outside can multiply them by that vendor’s own published per-token prices. The arithmetic is trivial. The inputs are public. What it produces is a list-price equivalent — not what the run cost the vendor, which does not buy its own inference at its own rate card.
Published token counts make a vendor’s own workloads priceable from outside. What the arithmetic yields is a customer-equivalent, not a spend.
The Structural Read
Publishing compute shapes is a transparency move. It can also be an unintended pricing disclosure.
When OpenAI published 88 hours, 130 billion tokens and 2.7 million messages, it handed any outside observer the inputs needed to estimate workload cost at list price. swyx did exactly that. The arithmetic is not novel; the disclosure that made it possible was.
This is what the Product Overhang Doctrine describes at the cost level. Capability — and coststructure — accumulates invisibly inside a frontier lab. When compute shapes surface, even partially, the outside world can suddenly price what was previously opaque. The vendor did not publish a price. The vendor published the inputs to one.
Latent Space — Shawn Wang (swyx), 2 Oct 2026
“Surprisingly, actually, less than I thought.”
swyx calls it “surprisingly, actually, less than I thought”, and Zhang answers “not that much”. That reaction matters as a data point about expert expectations. Two people who work closely with these systems heard a list-price equivalent for a Millennium Prize-class workload and found it smaller than they anticipated.
It is reported here as a statement about their expectations. It is not a verified cost, and this publication puts no number on what the run actually costOpenAI. No response from OpenAI is reported in the source. Nothing here addresses whether the proof is correct, whether anyone outside OpenAI has checked it, or whether any prize has been awarded; the company’s own summary claims a writeup and a formal proof in Lean and nothing beyond that, and this publication takes no view on the mathematics.
One further note on the figures: the 130 billion token count is the one swyx attached to the estimate. He separately cited 30 billion output tokens for the final stage, and noted that total agent messages were more than double that. These are different quantities measuring different things. This piece does not add, average or reconcile them.
Three Implications
IMPLICATION 1 — TRANSPARENCY HAS A PRICING SIDE EFFECT Publishing compute shapes is not neutral. Token counts and run durations are inputs to cost estimates. Any lab that publishes them hands outside observers the tools to approximate workload pricing at list rates — whether or not a price was intended to be disclosed.
IMPLICATION 2 — LIST PRICE AND INTERNAL COST ARE NOT THE SAME NUMBER A list-price equivalent is what a customer would pay. It is not what the vendor paid to run its own model on its own infrastructure. The gap between those two figures is material — and it is entirely unknown here. Treating the estimate as a spend is a category error.
IMPLICATION 3 — EXPERT EXPECTATIONS ARE THEMSELVES SIGNAL The downward surprise expressed by swyx and Alex Zhang is not a cost finding. It is a calibration point. People close to frontier AI systems expected a higher list-price equivalent for a run of this scale. That expectation gap is worth tracking — carefully, and without overstating what a single podcast exchange can confirm.
Business Engineer Framework
Product Overhang Doctrine
Capability — and coststructure — builds invisibly inside frontier labs. When compute shapes surface, the outside world can suddenly price what was previously opaque. The Map of AI traces exactly where these disclosures sit in the stack, and which players benefit when the numbers become legible.
A podcast host multiplied public numbers and produced a list-price equivalent. That equivalent is not a spend, not a disclosure and not verified. What it is — and what makes it worth reporting — is a demonstration that publishing compute shapes is a form of partial cost transparency, whether or not that was the intent.
The $40 million figure above is an estimate made by Latent Space host Shawn Wang, who goes by swyx, derived from public list pricing. He hedges it twice in the same breath, calling the amount “semi-public” and the figure “estimated”. It is not an OpenAI disclosure, and nothing above states what the run actually costOpenAI, which is not public.A list-price equivalent is what a customer would have paid at the published rate card.
It is not an internal cost, and no vendor buys its own inference at its own list prices. Nothing above puts a number on OpenAI’s actual spend or makes any claim about its coststructure or margins.OpenAI’s post “On the Navier-Stokes Millennium Prize Problem” returned HTTP 403 to this publication, so its body was not read. Its existence, title and date of 8 September 2026 were verified directly from OpenAI’s own RSS feed, whose summary states only that the company is sharing an AI-generated solution including a writeup and a formal proof in Lean.
The figures of 88 hours, 130 billion output tokens and 2.7 million messages are attributed to that post as shown on screen in the clip and are not independently verified here.Nothing above takes any view on the mathematics. The source claims no external verification of the proof, no refereeing and no prize, and neither does this piece.The quoted figures measure different things and are not reconciled above: 130 billion is the figure attached to the estimate, 30 billion is described as output for the final, and total agent messages are said to be more than double that.
Also absent: the pricing assumptions behind the estimate, which models were used, any response from OpenAI, and the identity of a third person appearing in the clip, whom the source does not name. Nothing above predicts anything and nothing here is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
