Polyfork, a start-up that sells 3D models built as small programs rather than fixed shapes, won the $500,000 first prize in the Build with Gemini XPRIZE at a live ceremony on September 25, 2026. According to XPRIZE's winners announcement, founder Lucas Martinic's catalogue held 583 models and earned $4,428 from 54 paying customers in its first month.
Peter Diamandis, founder of XPRIZE, presented the awards during a live edition of his podcast Moonshots, published on September 28. XPRIZE runs prize competitions for set goals. This one, backed by Google, gave teams 90 days to turn an idea into a business that uses AI and has real customers. Gemini is Google's family of AI models. When Diamandis handed over the check, he called the money "non-dilutive seed capital," meaning early funding for which founders do not give up any ownership of their company. "It's the best kind," he said.
The winners and the money
Diamandis announced the five finalists in reverse order. He brought the top two teams on stage together, saying the result would be obvious once either place was announced. The runner-up joined by Zoom.
- First: Polyfork, Lucas Martinic, $500,000
- Second: TICIAN Cierre, José Edgar Peña Ramirez, $200,000. Diamandis said he didn't know what that came to in pesos.
- Third: MyFixam, Jane Chukwuyen Odum and Ayodele Daniel Dada, $100,000
- Fourth: LaunchBridge, Chase Rosen and Leonardo Fall, $100,000
- Fifth: DodoPrep, Naveed Ahmed, $100,000
The episode is titled as if $2 million went to five builders, but that is not how the money was split. According to XPRIZE, the five main prizes add up to $1 million. The other $1 million of the $2 million pool went to 20 teams as $50,000 awards: five for categories and fifteen for runners-up. On stage, Diamandis mentioned that there were "20 other winning teams." XPRIZE says more than 26,000 builders registered.
Four judges decided the final placings: Palmer Luckey; Cathie Wood, CEO and chief investment officer of the investment firm ARK Invest; Mark Pincus, the entrepreneur best known for founding the game company Zynga; and Logan Kilpatrick, product lead for Google AI Studio and the Gemini API at Google DeepMind, who served as chief judge.
Why the challenge demanded paying customers
Diamandis opened with his reason for holding the competition. Many people are worried about jobs, he said, and the old deal of doing well in school, earning a degree and getting a job is "cooked." In its place he offered a "new social contract": find a purpose, which he defines as "something you love doing that helps other people." Then find a problem that "should not exist" and can be solved, and "put on your AI hat, put on your entrepreneurship hat and build that."
He said the prize asked teams to pick a meaningful problem affecting 100,000 people and to build a "revenue generating, profitable company" in 90 days, starting "from a clean sheet of paper." The winning companies themselves were not the point, he said. "The mission here is showing that it can be done," and inspiring people to do it themselves.
The official rules spell out what that meant in practice. Teams had to build and submit a new business between May 19 and August 17, 2026, and use a Google Cloud product. If their app used a language model, it had to make at least one live call to Gemini through its API, the interface programs use to reach the model. Other model providers were allowed as well. Entries had to show real users, revenue from customers with no connection to the team, monthly revenue, expenses and spending on finding customers.
Entries that passed an eligibility check were scored on three equally weighted criteria. The first was business viability: actual revenue and whether the business could last. The second was "AI-native operations," meaning AI carries out important decisions in the live product. The third was impact in the chosen category.
Wood added a practical caveat to the call for entrepreneurship. Diamandis had asked the judges what they would do if they had to start again from nothing. Wood said she expects an "entrepreneurial explosion" rather than mass joblessness. She also said entry-level jobs are not filling up the way they used to: people coming out of school are unemployed for six months on average, with a median of three. She called this "a blessing in disguise." Her advice to young people was to start a company "with AI and no employees," but also to "keep interviewing," because she expects the initiative and hands-on AI experience to help them get hired quickly. "Or even better," Diamandis replied, "your company takes off and you tell them to go away."
What the five finalists built
The finalists' own presentations are not part of the episode. The descriptions below come from each team's competition submission.
Polyfork makes 3D models in which every model is a program with adjustable settings. In Martinic's submission, making a traffic cone taller doesn't simply stretch it. The cone's geometry is rebuilt, an extra reflective band appears, and the number of triangles in the model rises from 280 to 472. Models come in kits that share dimensions, colour palettes and construction rules, so the parts fit together. Customers can recolour, resize and remix them, and software can search the catalogue and use the models directly.
Production is largely done by AI agents, programs that carry out multi-step tasks on their own. A person writes a short request. Gemini turns it into a brief and reference art, another agent writes the code that produces the shape, and a rendering tool measures the result. A separate agent then reviews it in a fresh session. Martinic reports that this automated reviewer sent back roughly two-thirds of more than 1,000 attempts. Even so, he says human review is still needed, especially to get proportions right, so "AI-native" does not mean unattended. XPRIZE puts the agent-assisted production rate at more than 175 models a week.
TICIAN Cierre helps Mexican businesses recover VAT credits. VAT is a consumption tax, and businesses can often recover what they paid on their purchases as credits. Diamandis described the claim as a complex process of preparing, auditing and submitting. According to the builder's submission, the system matches Mexico's digital tax receipts against bank statements and prepares an eight-part file for the SAT, the country's tax authority. Ordinary Python code removes duplicate receipts and calculates the totals. A separate Gemini service then checks the summary and records its reasoning. If that checker is unavailable, the reconciliation still finishes.
The submission also lists defects found along the way. In one, credit notes were passed to the auditing service as positive amounts while the main engine subtracted them, which produced a false discrepancy. It was fixed in August. A later check on an accountant's 227,000 real receipts found no duplicates and exactly reproduced totals that had been delivered before, the builder reports.
MyFixam is a marketplace, which Diamandis described as "a task rabbit" for Nigeria's informal artisan economy. According to the team, it covers 29 trades. Customers pay before a job starts, and the money is held until they confirm the work is finished. The platform keeps 10% of completed bookings. Interviews with artisans changed the team's focus from helping them get noticed to helping them get paid reliably.
Gemini drafts artisans' profiles from photos and short descriptions, interprets customer requests and moderates posts. Prices come from the company's own stored data rather than being made up by the model, and people check artisans in person and settle disputes. A September 9 update reported 620 users and at least 20 booking requests a week in the tenth week after launch. To cut down on haggling, the team was starting to show fixed prices for standard jobs, taken from artisans' own rate sheets, beginning with painting.
LaunchBridge turns an idea into a registered business in under 72 hours, but for now only in Virginia. Its $350 package includes the $100 state filing fee and covers forming an LLC (a common legal structure for small businesses), obtaining an employer identification number, building a website and setting up Stripe payments. A registered-agent service costs $110 a year extra. Founders describe their business in plain language, and Gemini helps classify it. The team says the hard part was automating government websites that offer no API. Its software has to operate those sites through a browser, coping with fragile forms that change without warning. Expanding to about ten more states remains a plan, and each state needs its own pricing and workflow.
DodoPrep is an AI tutor. According to Ahmed's submission, it turns PDFs, slides, notes, lecture recordings and YouTube videos into lessons, quizzes, flashcards, mind maps and study plans. Uploaded material becomes a searchable library. Before answering a question, the tutor searches that library by meaning (semantic search) and bases its reply on the passages it finds, a method called retrieval-augmented generation. It also adjusts to each student's progress and gaps. Ahmed lists the cost of running the AI as one of his engineering problems, since he had to balance it against speed. In an August 30 update, he reported that cumulative revenue had grown from an earlier milestone of more than $1,500 to more than $4,500.