Part 1 of 4: The Idea Is Enough Now
An overview of the AI game creation platform landscape — and what it means that anyone can ship a game today
There’s a line on the Crayon AI website that stops you if you’ve spent any time thinking about game development:
“Most people who want to make a game never make one. The idea is there. The world is there. But between that idea and a real game sits an entire stack of tools, tutorials, engines, code, assets, and production work.”
It’s a diagnosis that’s held true for fifty years. The distance between having an idea and shipping a game was measured in years of learning, thousands of dollars in tools, and the willingness to become a different kind of person. A writer had to become a programmer first. An artist had to learn to think like an engineer. A designer had to figure out what “designer” even meant in software terms.
That distance has collapsed in the last 18 months. It has effectively disappeared for a wide category of game ideas, and the platforms responsible are growing faster than the traditional industry wants to admit.
The Numbers
In 2025 and early 2026, a cluster of AI-native game creation platforms crossed thresholds that would have seemed implausible three years ago.
Rosebud AI hosts over 500,000 games made by more than 70,000 creators on a platform that is essentially bootstrapped. No Sequoia check. CEO Lisha Li, a former Amplify Partners principal with a Berkeley PhD, has mapped out a roadmap from “text-to-game” to “prompt-to-world,” with local world generation on devices by 2027 and, by 2030, worlds indistinguishable from designed experiences. Whether the timeline holds is debatable. The 500,000 games exist right now.
Astrocade, backed by Sequoia, Google’s AI Futures Fund, and NVIDIA, with Fei-Fei Li of Stanford’s Human-Centered AI Institute as a co-founder, hit 5 million monthly active users and 140 million game plays per month eight months after launch. The company raised $56 million, including a $10 million creator fund. Verse8, backed by Nexon, Netmarble, and Google, reached 3.5 million monthly active users since launching in July 2025, and is the only platform with native real-time multiplayer generated from a single prompt. Playabl.ai, a Y Combinator entry, crossed 100,000 beta users and solved the problem nobody else thought to solve: not making game creation easier, but making game discovery easier, with a TikTok-style feed that surfaces a new playable game every scroll.
Playo.ai, a Tel Aviv startup backed by iAngels and Overwolf, is running a different kind of numbers game. It isn’t a creator platform. Complete 3D games from a single prompt, under a minute, at roughly three dollars each. The plan is to go from 50 games in year one to 5,000 in year three with a team under thirty. The bet isn’t on creators getting better tools. It’s on AI generating the supply side of games entirely.
And then there’s HypeHype. Formerly Frogmind, the Helsinki studio behind BADLAND (Apple Design Award winner, hundreds of millions of players), now backed by Supercell and Benchmark with its entire identity rebuilt around one sentence: “We’ve always believed that anyone with an idea should be able to build a game. AI is finally making that real.”
Supercell had €2.65 billion in revenue in 2025. They know what it costs to make games and what it takes to make games people actually play. They also back Bitmagic, Nilo, and 721 Games. Four coordinated bets on AI-native game creation, from the people who built Clash of Clans. That’s worth paying attention to.
What These Platforms Do
The category moved past “type a sentence and watch a game appear” about a year ago. The better platforms work more like a design conversation than a generation event.
Crayon AI runs a clarification step before building anything. It asks what kind of game you want (“Chill pearl-collecting ride? Fast arcade dodger? Level-based?”), negotiates scope, then generates. Games publish directly to Poki and Crazy Games, actual web distribution channels with real audiences. Jabali.ai, backed by Sony’s Innovation Fund and BITKRAFT Ventures, offers two modes: Vibe Code for mechanics-minded creators, Design Mode for visual storytellers. Full source code ships with every project. Sony backing a platform that hands creators their own code outright is a quiet statement about where the industry thinks ownership is heading.
Arcade (joinarcade.ai) is pre-launch, lean, and has early adopters from Epic Games, Meta, Microsoft, Nintendo, and Riot Games. When people from Nintendo use your tool personally before it’s publicly available, something is being communicated about the gap between what these companies build for professionally and what they believe is actually possible. SEELE AI exports to Unity, Three.js, and Unreal Engine 5, the only platform with that triple coverage, meaning games built there aren’t locked to any single distribution channel.
GDevelop sits in a different category from all of them. The open-source, no-code engine has been used in thousands of schools long before AI entered this conversation. What its AI agent adds is not text-to-game but text-to-feature: describe a mechanic and the AI writes the event logic against your specific objects and configuration. It already knows your game. That’s a narrower claim than generating a whole game from a blank prompt, but a more reliable one for creators who want to stay in control of what they’re building.
The Sandbox Studio, announced June 2026 and in closed alpha, inverts the standard model entirely. It doesn’t come with its own AI. The pitch is bring-your-own-LLM: connect Claude Code, Cursor, or Codex, and the Studio provides the runtime, six TypeScript templates, 300,000-plus voxel assets, and distribution rails to browsers and Telegram Mini Apps, with Steam to follow. CEO Robby Yung’s framing is direct: “AI tools generate code. The Sandbox Studio generates games. Players will see the difference.” The claim is that the runtime, the asset library, and the publishing pipeline are the hard parts — and the model you use to write the code is interchangeable. Public access is Q4 2026. Real thesis, roadmap not yet shipped.
And then there’s Roblox, which is doing all of this at a scale that makes every other platform in this piece look like a prototype. Roblox has 380 million monthly active users, 144 million daily active users as of Q4 2025, and $4.9 billion in revenue last year. In April 2026, the company reported that 44% of its top 1,000 creators already use Roblox Assistant or third-party AI tools via MCP to plan, build, and test their games. Their 4D generation feature, launched in February 2026, lets creators generate fully interactive in-game objects from a text prompt — not static props but objects that behave correctly in the game world. Players in games using it showed a 64% increase in play time. Roblox has also built MCP server integration into Studio, connecting directly to Claude, Cursor, and Codex. Roblox is not a platform watching AI game creation happen. It is one, at massive scale, and has been for longer than most of its competitors have existed.
Pieter Levels, and the Wall He Found
Pieter Levels built a multiplayer MMO flight simulator in hours in February 2025. It scaled to hundreds of thousands of users and was earning over $50,000 a month in sponsorships. He organized the Vibe Coding Game Jam, which drew 1,170 submissions, all at least 80% AI-generated. The 2026 edition has $40,000 in prizes.
Every winner of the Jam was a free browser game. The Great Taxi Assignment, Vibeware, Vector Tango: fun, technically impressive, genuinely creative. Not commercial products. There’s a documented complexity ceiling with AI-generated game code, the point where a 1,400-line JavaScript file becomes unnavigable and the AI, as one developer put it, “reaches a dead end and can’t recover.” That ceiling is real.
But it’s not a ceiling on the tools. It’s a ceiling on the workflow. Prompting an AI and accepting whatever it produces is one approach. It has a complexity wall. A different approach, where the person’s taste, judgment, and understanding of what makes something good are doing real work alongside the AI, doesn’t hit that wall the same way. Most creators on Rosebud and Astrocade haven’t found that yet. Some have. That’s the more interesting story, and it’s what Part 2 is about.
Why Now
Owen Mahoney ran Nexon for a decade. His essay “Deus Ex Machina” offers the clearest framework I’ve read for what’s actually happening. He traces the history of communication as a series of unbundlings: writing unbundled speech from presence, the printing press unbundled creation from duplication, the internet unbundled duplication from distribution. Each one changed who the winners were.
AI, he argues, is unbundling the substantiation of an idea from the idea itself. Every previous tool still required you to do the work of turning a thought into a functional artifact. Engines, asset stores, no-code editors: all of them still needed you to build the thing. AI removes that last step. When you describe a game to Rosebud or Verse8 or Crayon, you form the thought and the AI gives it form.
The 70,000 creators on Rosebud have shipped half a million games. Most of them are rough. That doesn’t mean they don’t count. The printing press’s first commercial outputs were indulgence certificates and calendars, not bibles. The medium proves itself before it gets good.
What’s more interesting than the platform numbers is what happens to the creators inside them who start caring about quality. Who iterate past the first playable build. Who tune a mechanic, question a design decision, and somewhere in that process stop asking “how do I make this work?” and start asking “how do I make this good?” That shift, from generation to judgment, isn’t a departure from the platform. It’s the platform working exactly as intended.
The tools don’t cap you. They open a door. What’s on the other side isn’t harder. It’s just different, and considerably more powerful than anything a single prompt can reach on its own.
In Part 2: What that next layer looks like in practice, and how a professional game designer with 15 years of experience built a complete Castlevania-style vertical slice in one month for $50, using AI agents and the instincts he already had.
Sources and further reading
Rosebud AI — “From Text-to-Game to Prompt-to-World”
https://lab.rosebud.ai/blog/from-text-to-game-to-prompt-to-world
Fortune — Sequoia-backed Astrocade raises $56 million
https://fortune.com/2026/05/05/astrocade-raises-56-million-series-b-sequoia-video-games-platform-ali-amir-sadeghian/
Verse8 seed funding (PocketGamer.biz)
https://www.pocketgamer.biz/verse8-secures-5m-to-scale-its-ai-driven-game-creation-platform/
Supercell Investments portfolio
https://investments.supercell.com/investments/
Owen Mahoney — “Deus Ex Machina”
https://www.owenmahoney.ai/owen-mahoney-blog/deus-ex-machina
Vibe Jam 2025 results (Indie Hackers)
https://www.indiehackers.com/post/tech/pieter-levels-just-announced-the-winners-of-the-2025-vibe-code-game-jam-Uz0wHG4pI3KBOiFhP5YR
Jabali.ai seed funding (BITKRAFT Ventures)
https://www.jabali.ai/blog/jabali-raises-funding
Crayon AI
https://usecrayon.ai
Arcade
https://joinarcade.ai
GDevelop AI agent blog
https://gdevelop.io/blog/gdevelop-ai-build-games-faster-ai-powered-tools
The Sandbox Studio announcement (Crypto Briefing)
https://cryptobriefing.com/sandbox-ai-game-engine-studio-launch/
Playo.ai — “5,000 games in two years” (Calcalist Tech)
https://www.calcalistech.com/ctechnews/article/z25izg7su
