Polyfork Wins $500,000 Gemini XPRIZE on $4,428 Month-One Revenue

XPRIZE published the winner’s own operating figures — a rare disclosure that lets us read exactly what “revenue-generating” meant in ninety days.

Every figure below comes from XPRIZE’s own announcement page, including the winner’s revenue. These are competition-stated figures, not audited accounts and not a filing. The 113x ratio, the figure of about $7.60 per model and the figure of about $82 per paying customer are this publication’s arithmetic on numbers XPRIZE published. A prize is paid for what a business might become rather than for what it billed last month, and nothing below characterises the award as too large or the winner as undeserving. Nothing here is investment advice.

What Happened

On September 25, 2026, XPRIZE announced that Polyfork — built by Lucas Martinic of Germany — had won the grand prize in the Build with Gemini XPRIZE, a two-million-dollar global competition that gave entrants ninety days to build a revenue-generating business from a clean sheet. More than twenty-six thousand builders registered. Twenty-five projects received prize awards. Polyfork took the top position and the five-hundred-thousand-dollar grand prize. All figures in this piece are competition-stated by XPRIZE, not audited accounts.

What made the announcement structurally unusual — and worth analyzing — is that XPRIZE published Polyfork’s own operating figures alongside the winner’s name. That is not standard practice for competition write-ups. The disclosed numbers: a catalogue of five hundred and eighty-three model programs, output of more than one hundred and seventy-five model programs published per week, and first-month revenue of four thousand four hundred and twenty-eight dollars from fifty-four paying customers.

The rest of the prize structure, as published by XPRIZE: two hundred thousand dollars for second place (TICIAN CIERRE, José Edgar Peña Ramírez, Mexico); one hundred thousand dollars each for third (MYFIXAM, Jane Chukwuyem Odum, Nigeria and the USA), fourth (LAUNCHBRIDGE, Chase Rosen, USA), and fifth (DODOPREP, Naveed Ahmed, India); five category prizes of fifty thousand dollars each; and fifteen runner-up awards of fifty thousand dollars each.

The prize list is reported exactly as published and is not summed here. Judges included Cathie Wood of ARK Invest, Palmer Luckey of Anduril, Mark Pincus of Zynga, and Logan Kilpatrick of Google DeepMind, who leads Google AI Studio. Awards were presented by XPRIZE chief executive Anousheh Ansari.

Build With Gemini XPRIZE — Sequence of Record

Competition Open: More than 26,000 builders registered across a 90-day build window — the only constraint: produce revenue before the clock ran out.

Month One (Competition-Stated): Polyfork’s catalogue reaches 583 model programs; 54 paying customers generate $4,428 in revenue. Output rate: more than 175 model programs per week.

September 25, 2026: XPRIZE announces Polyfork as grand-prize winner. Prize: $500,000. XPRIZE publishes the operating figures above — an unusual disclosure for a competition announcement.

What Is Not Published: No month-two figure, run-rate, cost, margin, or valuation. No revenue figure for any other winner. No count of how many of the 26,000 registrants finished.

The Structural Read

The disclosure is the story. Most competition write-ups publish a name and a cheque. XPRIZE published what the business actually took in. That earns the figures serious analytical attention — while also requiring that every number be handled carefully, because these are competition-stated, not audited accounts, and no second month exists to corroborate the first.

Two derived figures are worth sitting with. On this publication’s arithmetic: five hundred and eighty-three models generating four thousand four hundred and twenty-eight dollars implies roughly seven dollars and sixty cents per model across the catalogue. Fifty-four paying customers generating four thousand four hundred and twenty-eight dollars implies roughly eighty-two dollars per paying customer in that first month. Both figures are our own arithmetic on XPRIZE’s published numbers — not stated by XPRIZE.

The more structurally interesting observation comes from reading the throughput figure against the demand figure. Polyfork publishes more than one hundred and seventy-five model programs a week. It had fifty-four paying customers in month one. The two sides of this business are not scaling on the same clock. Generation is cheap and close to unbounded in the AI era — the supply side of a model-catalogue business can grow almost frictionlessly.

The demand side, acquiring and converting paying customers, operates on a very different rhythm. That asymmetry is visible from exactly two published numbers, and the data stops there.

XPRIZE publishing operating figures alongside the winner’s name is the meaningful structural precedent here. If prize competitions begin disclosing what winning businesses actually earned during the build window — not just a name and a cheque — evaluators, founders, and builders get a calibration point that currently does not exist. “Revenue-generating” as a binary test tells you almost nothing. Four thousand four hundred and twenty-eight dollars from fifty-four customers in month one tells you what clearing that bar actually looks like in practice. That specificity has value independent of whether Polyfork scales.

WHAT A PRIZE IS ACTUALLY PRICING

The approximately one-hundred-and-thirteen-times gap between the grand prize and month-one revenue — our arithmetic on XPRIZE’s stated numbers — is not evidence of mispricing. It is a description of how prize formats work: the award is for potential, architecture, and proof-of-concept, not for trailing twelve months. The ninety-day constraint makes this structural rather than incidental. Anyone reading a prize announcement as a revenue validation is misreading the format. The correct read is: this business cleared the binary test and the judges assessed its ceiling.

THE CATALOGUE ECONOMICS OF AI-NATIVE PRODUCTS

Roughly seven dollars and sixty cents per model across five hundred and eighty-three programs, and roughly eighty-two dollars per paying customer in month one — both our arithmetic — describe a business where the unit of supply (a model program) is nearly free to produce and the unit of demand (a paying customer) is the binding constraint. This is the central economic pattern of AI-harnessing businesses at early stage: generation scales, conversion does not.

The strategic question — which the published data does not answer — is how the demand side is acquired once the supply side is already effectively unbounded.

The Bottom Line

Polyfork winning five hundred thousand dollars on four thousand four hundred and twenty-eight dollars of competition-stated, first-month revenue is not a paradox — it is a precise description of what prize formats are for. The creditable part of this announcement is not the ratio; it is that XPRIZE published the ratio at all. That transparency gives every builder who registered, every judge who evaluates, and every analyst who reads the announcement a factual baseline that competition write-ups almost never supply.

A business that publishes more than one hundred and seventy-five model programs a week has solved the supply problem that AI makes solvable. The demand side — fifty-four paying customers — is the problem that has always been hard and remains so. That is where the work is, and no prize changes it.