Ninety Thousand Games a Day: Roblox's Bet That AI Should Wash the Dishes

Konstantine Buhler opens with the kind of numbers that make Roblox sound less like a game company than a small economy. At the time of the conversation the company had a $29 billion market cap, about $3.5 billion in run-rate revenue, and $600 million in operating profit. Creators had been paid more than $800 million over the course of a year. Players had logged 70 billion hours on the platform in that same stretch.

The audience is no longer only kids. Daily actives sit at 79 million, and 46 million of those users are over 13. Voice chat adds another 3.8 million people talking on the platform each day. Monthly actives, which Roblox does not publish, are in the several hundreds of millions, Corazza says — a figure he finds more staggering than the daily number the company repeats in public.

The creation side is what he wants people to sit with. Every day, roughly 90,000 experiences and games are published on Roblox. Some become worldwide franchises with tens of millions of concurrent players. Others are one person's experiment. Together they are the reminder, he says, of how much this has become a creation platform, not just a place to play.

Corazza did not start in games. About twenty years ago he came from Italy to Stanford on an exchange program, working on computer vision and machine learning for measuring human motion. The work sat on a boundary: one path led to biomedical engineering, the other to animation. He decided animation was the larger opportunity, built video-based motion capture and animation tools, and in 2008 spun out Mixamo. Adobe acquired the company in 2015. Mixamo is still one of the industry's most used machine-learning services for rigging and animating characters.

He spent seven years at Adobe helping build the 3D offering, including Stager and the AR tool Aero, and the acquisition of Allegorithmic, which filled out a materials stack. After that he was working with Adobe's CTO on entirely new ways to generate content. That is when Roblox reached out. A few breakfasts with CEO Dave Baszucki were enough to convince him that generative AI would blossom where the data already existed: a platform with some 15 million experiences played each year, and a chance at worldwide impact.

He still sounds slightly stunned by the compounding. People buy houses on Roblox income. Studios of more than a hundred people have formed, some of them venture-backed. "It's really like an amazing example of a compounding effect," he says, "and we are every day mesmerized ourselves about the success." The thing that made him say yes was not the market cap. It was the dataset hiding in plain sight.

The uniqueness Corazza keeps returning to is that Roblox is one of the most vertically integrated companies on the planet. It owns data centers and the hardware inside them. It owns the app that distributes the games. It owns voice, live chat, and chat translation. It owns Roblox Studio, the creation tool his team leads, and the creation services that sit around it. The stack runs from bare-metal CPU and GPU clusters all the way to the editor.

The commercial rule on top of that stack is just as unusual. Roblox only takes revenue when creators make money. There is no upfront fee to publish, no fee for the tool, and a long list of services that stay free. If a game has one user or a million, the creator does not provision servers. Roblox scales it, pays for the CPU, GPU, and storage, and keeps that cost opaque to the person building. Friction to start is close to zero.

That is also why the company can subsidize AI. Corazza says Roblox is one of the few places with a full-fledged AI assistant for game development — code, materials, textures, assets — offered free to creators. The vertical stack is not a talking point for investors in this telling. It is the reason the assistant does not have a price tag.

Several million people create on Roblox each month. The average creator is in their mid-20s, older than the player base. Most of them are world-builders and artists. Only about 30 to 40 percent write code. Corazza's own recent obsession, for what it is worth, is the car game Driving Empire. The coding share is low for a simple reason: building a world is intuitive, and writing code is not. You have to master abstract constructs before anything in the scene will move.

That friction is why Code Assist shipped first, in March 2023. A year and a half later it had grown into a full assistant with three jobs. On code, it autocompletes, explains, debugs, and attaches scripts to parts in the scene. On documentation, it answers "how do I do this" so creators stop hunting through DevForum threads. On assets, a material generator and a more complex texture generator can wrap a 3D object from a text prompt at decent fidelity. Put together, people can already type natural language and get a simple game from scratch. The future version he describes is spoken on a phone: generate the world, the forest, the boss, and the mechanics.

The games that come out of a two-hour internal game jam are "pretty incredible," he says, but they are not yet at the sophistication of someone who can code. That is not the point. The first script is written for you and attached to a part; next time you know where it goes. Assistant is a companion that shows you by doing, then gets out of the way as the skill sticks. Sonya Huang asks whether more people are coding now that the barrier is lower. Corazza answers with a cohort comparison instead of a ratio. People who use the assistant create 180% more code. People who use the material generator create 60% more materials. On the thing that actually matters — publishing — the lift is about 30%, and most of the product is still in beta.

The KPI his team watches is retention, not hours or Robux. Week-over-week retention for assistant users is higher than for any other feature they have rolled out, and Studio's overall retention has moved with it. Daily users have grown organically. The team has spent nothing on marketing. Free, he notes, is the best marketing there is, even if it is not cheap.

The cost of free is covered by a bargain Corazza states in one breath. Roblox told the community: give us access to your data to train AI, and we will make the best assistant we can and put it back into Studio at no charge. "We're not making money off your data. We're actually helping you create more." The overwhelming majority of creators said yes. That permission is why he can claim not only one of the largest datasets in the world, but one of the most multimodal: code, images, 3D assets, audio, video, the interactivity that glues those together, and the analytics on what people actually play.

Buhler, listening, reaches for an older memory. In 2017 and 2018 he sat in on Roblox board meetings with his friend Craig Sherman. Behind closed doors the talk was not monetization. It was community, and teaching people to code, and lifting the people on the platform. The data bargain, in that light, is not a new personality. It is the same ethos with a training loop attached. The company treats the dataset with high security standards; Corazza is careful to say so. The point of collecting it is not a side business in model APIs.

What they are actually doing, he says, is teaching AI game development. The early tools look like small jobs — make a material, write a texture, complete a function — but they are already converging into larger ones. The model is not being taught how to make an image or how to write code in isolation. It is being taught how a whole game gets made, including, eventually, which levels get played and which games retain people. Those signals are in the data. The tools are the on-ramp to that single objective.

Asked which parts of game development AI will take over, Corazza refuses the premise. At RDC a couple of weeks earlier, the best parallel the team offered was a dishwasher: nobody wants to wash the dishes. The latest release, Assistant Actions, is aimed at those chores. Imagine an open world with a forest of 100,000 trees, and fall arriving. Recoloring every canopy by hand would be enormous. Three lines of text — select all the trees other than the pine trees, and make the leaves yellow — and 57,000 trees change in a few seconds. That feature got the highest community score of any AI capability Studio shipped in two years, by a landslide. It did not replace talent. It did the work nobody wanted.

The implementation sits on three things at once. Assistant is an LLM, including open-source models such as StarCoder that Roblox trains on its own data, and it is getting better at generating code. Some of that code can run at edit time in Studio's command bar, not only at runtime. And because the assistant is inside Studio, it has full awareness of the scene's data model: this is a tree, this is a car. Code generation, command execution, and data-model awareness together let the model manipulate a world in bulk. A good amount of that object understanding, Corazza says, is inference rather than a pile of hand labels.

Zoom out, and he thinks the last thirty years of digital tools were about surfacing more control. Photoshop will let you own every pixel, even if most people never need to. The next generation of tools only works if they capture intent. Text, images, sketches, voice, and even high-level gestures become inputs; velocity jumps by two orders of magnitude. Casual creators may already have enough to post on TikTok or invite friends into a Roblox experience. Pros still need the old fine control. The tension is how to serve both.

His answers are iteration, progressive disclosure, and a refusal to make Studio do everything. One-shot generation is a slot machine: throw in text and an image, pull the lever, see what comes out. Creation, he believes, is iteration, so the same AI has to accept "only the trees above five feet" or "open this texture, I want to paint on it." Deep tools can hide behind a curtain for people who want them. For the rest, interoperability with Blender, Photoshop, and Substance Painter is the honest design. Studio should be good at what it does, not a place where you can do everything and nothing well.

Studio is the incubator. What the team is most excited about is carrying the same AI into live experiences: you are playing, you want to modify a level, you and your friends want to make something new and jump in. Companies such as Ego.live are already experimenting with generative agents that simulate a community. From the player's side, Corazza thinks games can become more personal with very little extra work from the developer. An LLM that remembers past sessions can change the challenge or spin the story when you come back. The same title does not have to play the same way twice.

Buhler asks what that does to infrastructure. Corazza's answer is the reverse of what a systems person might expect. The first step is NPCs hooked to an LLM, which does not take a massive rewrite. Generating whole worlds on the fly is harder, but the biggest challenge is not latency or compute. It is moderation. Safety is Roblox's number one product; the company wants people to connect with civility. The moment players can create anything, the bar goes up. CDNs, caches, pre-generation, level of detail, and streaming are a solved genre in games. Opening the Pandora's box of in-experience creation without building the guardrails is the actual risk.

Between NPCs and a fully generated world he points at constrained creativity. A popular Roblox game, Build a Boat, lets you build only a boat, from a limited set of materials, at a limited size, then race it. The world is the developer's; the player's freedom lives inside the constraint. Because Roblox is a physics sandbox — aerodynamics, fluid dynamics — the wind tells you whether the boat is any good. AI can power that pattern: build a race car, an airplane, a piece of the loadout, and gain an advantage without rewriting the game. On neural physics engines, he is unconvinced. Efficient real-world approximations still matter more than how closely you model Navier-Stokes, especially in a purely digital world.

A world of only NPCs, watched rather than played, sounds like television to him, and he will not rule it out: people already watch other people play. A more realistic first step is hybrid — inject interesting NPCs into a newly launched place that does not yet have a thousand concurrent humans. On user-generated content, avatar auto-setup is the proof. A large community makes a living on clothing and accessories, but 3D modeling in Blender or Maya is a narrow skill. Image- or text-to-avatar substantially increased how many avatars could be created; the early beta is something they plan to double down on. Huang asks whether generative AI blurs user and creator into the same person. He reaches for music. Karaoke, then GarageBand, then type-the-lyrics-and-the-genre. Game development will take longer because the content is more complex, but entropy only increases.

Roblox has an enormous amount of video, audio, text, and 3D. Corazza says the hard part is no longer gathering it. The work is turning raw data into something you can actually train on. A couple of weeks before the taping, the company announced it is building a 3D foundation model, with the intent to open-source it, for synthesizing scenes and worlds from multimodal input. World creation is the first attack surface. After that comes interactivity — doors that open, cars that run — and then gameplay. Because they can see which games get traction and which levels get played, they want the model to make things that are fun, not just things that match a prompt. That will take unsupervised learning and a way to pick up signals of creator intent.

The physics engine and full 3D scenes are, in his view, a structural advantage over video-only world models. If you have the 3D scene, you can anchor spatial and temporal coherence instead of estimating it from 2D. Video-to-video stylization drifts. ControlNet plus a depth map — a two-and-a-half-D patch — made it more reliable, but a character who looks at you, looks away, and looks back can still come back as a different person. Everyone has seen those examples. The industry ran a long way on 2D. Incorporating 3D information is the work that still has to arrive.

Asked whether 2D is a dead end, he splits the question. For a single image — texturing a character, stylizing a frame — 2D has been fantastic. Add time, a moving camera, and two thousand frames that have to stay consistent, and you are at a disadvantage without 3D. The fatal hole is occlusion. If an arm is behind a back and then pops out, 2D data has never seen the hand, so nothing can enforce how it should look. In a 3D representation, even the surfaces the camera is not seeing are still there. Feed that into a generator and the outcome, he thinks, gets much better.

The longer bet is neural rendering. Historically, assets and art style were glued together: you cannot flip Super Mario into Call of Duty without rebuilding everything. Generative rendering would leave geometry and physics capsules in place and restyle the game in real time from text and reference images — a final compute pass, like bloom, that adds no extra downloads. It is still expensive. He believes that in five years this is how games will be built, and that it will run at least on high-end phones. If a developer allows it, players could even choose their own style while sharing the same physics and gameplay with teammates who see a different picture. The assistant leaving beta, meanwhile, is the thing creators can use tomorrow, still free.