90 days, $500,000
The Build with Gemini XPRIZE was one of the largest hackathons in the world: 90 days to build a real AI business, $2 million in prizes across 25 winners, judged on business viability, how deeply AI ran the operation, and category impact. More than 26,000 builders registered across five categories. On September 25, 2026, at Moonshots LIVE in Los Angeles, the $500,000 grand prize went to Polyfork — built by Lucas Martinic, representing Germany.
Polyfork's insight starts with a frustration every 3D artist knows: you download a finished 3D model and it's frozen. Wrong size, wrong colors, wrong proportions — and fixing it means rebuilding it yourself. Polyfork sells 3D models as programs rather than fixed digital assets. Users can resize, recolor, and remix each model before downloading, tuning parameters through sliders and color pickers, then importing a customized result.
The business is AI-native to its core. An AI-agent factory generates and publishes new models continuously — more than 175 per week — building a catalog that grows without proportional human effort. Within its first month, the platform had 583 models in its catalog and had generated $4,428 in revenue from 54 paying customers. That's the number that mattered most to the judges: not the technology demo, but the proof that someone would pay.
Models as programs
The competition's framing is worth understanding: 'Revenue isn't the point — it's the proof your solution actually worked.' The XPRIZE challenged builders to create AI-native businesses addressing real-world problems, evaluated equally on viability, AI integration, and impact. Polyfork won because it scored on all three — a real product, deeply automated operations, and a genuine improvement to how 3D assets work.
The 'models as programs' concept is the key differentiator. Traditional 3D marketplaces sell frozen meshes — what you download is what you get. Polyfork's parametric approach means every model is a small program with knobs: drag a slider and the geometry rebuilds, pick a colorway and every part recolors at once. Imported models stay editable afterward, so changing your mind a month later doesn't mean hunting for the asset again.
The technical execution is serious. Every model is flat-shaded vertex color on a single material with no textures, so it draws in one call — kilobytes per model, not megabytes. The Unity connector, published open-source on GitHub, puts the whole catalog inside the Unity editor: browse, preview in an orbitable 3D view, remix with sliders, and import as .glb at real-world scale. It meets 3D artists where they already work instead of asking them to adopt a new workflow.
What the judges recognized — and what competitors should worry about — is that parametric models change the unit economics of 3D content. A traditional marketplace needs artists to make every asset; Polyfork needs its agent factory to generate variations, and each parametric model effectively becomes dozens of products. One model with ten adjustable parameters isn't one SKU — it's a catalog.
The agent factory
The AI-agent factory is what makes Polyfork a genuinely AI-native business rather than an AI-assisted one. Human 3D artists can't produce 175 models a week; AI agents can, with humans setting direction and quality bars. This is the central idea the XPRIZE was designed to demonstrate: small teams — in this case, effectively a solo founder — using AI agents to automate substantial portions of company operations.
The competition specifically focused on businesses built from a clean sheet rather than existing startups bolting AI onto established products. Participants had 90 days to build, launch, and begin generating revenue. Polyfork's first-month numbers — 583 models, 54 paying customers, $4,428 in revenue — were modest in absolute terms but extraordinary as a 30-day-old business built by agents. The trajectory, not the total, is what won.
Martinic's approach also reflects a smart reading of the 3D asset market. The catalog is split: about half free forever (no account needed, commercial use allowed), with an API key unlocking the paid catalog and higher remix allowances. Free users browse and import; serious users pay. It's a funnel built on the product itself, and the free tier doubles as marketing — every free model in someone's project is Polyfork distribution.
Why it works
Polyfork's win is a data point in a larger shift: the businesses that win AI competitions aren't the ones with the cleverest demos, but the ones where AI is load-bearing in the operations. The agent factory isn't a feature — it's the supply chain. Without it, there's no catalog; without the catalog, there's no business.
The 'models as programs' insight is also exportable. In any market where digital goods are sold as frozen artifacts — templates, presets, components — there's an opening for a parametric version that lets buyers customize before they commit. The format becomes the moat: once customers build workflows around remixable assets, going back to static files feels like a downgrade.
For builders watching the next XPRIZE or the next wave of AI-native businesses, Polyfork's playbook is clear: pick a real workflow pain, make AI do the scaling work humans can't, prove it with revenue in 90 days, and distribute inside the tools people already use. Lucas Martinic did all four — and took home half a million dollars to prove it.
The playbook
- Sell the format, not the file — Polyfork doesn't sell finished meshes — it sells parametric programs. The format is the moat, and competitors selling static files can't follow without rebuilding.
- Automate the supply side completely — An AI-agent factory publishes 175+ models a week. The catalog grows while the founder sleeps — that's the AI-native advantage the judges rewarded.
- Distribute inside the workflow — The Unity connector puts the catalog inside the editor where 3D artists already work. Distribution through existing workflows beats asking users to come to you.
- Revenue is the proof, not the point — 583 models and $4,428 from 54 paying customers in month one. The XPRIZE judges evaluated business viability equally with AI-nativeness — revenue was the proof.
- Make half the product free — About half the catalog is free forever. Free users become the top of the funnel for the paid catalog and API keys.