> ## Content Index
> Fetch the complete content index at: https://www.regenraygun.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Why Roblox is the most AI-pilled game company
- URL: https://www.regenraygun.com/roblox-most-ai-pilled-game-company/
- Published: 2026-11-06T11:47:49.000Z
- Updated: 2026-11-06T13:17:01.000Z
- Description: Quietly, Roblox is combining 20 years of player data with its custom AI infrastructure.
- Author: Jon Jordan
- Tags: Roblox

Roblox runs nearly 400 AI models, and serves 77% of requests from creators using its AI Build platform from infrastructure it owns.

That disclosure, in the company's [Q3 2026 shareholder letter](https://s27.q4cdn.com/984876518/files/content%5Ffiles/Q3-2026-Earnings-Shareholder-Letter.pdf?ref=regenraygun.com), is the first time anyone in games has published the ratio, and it describes a position no-one else in the industry occupies.

The models cover creation, safety, social communication, discovery, economy and infrastructure. They are a mix of open-weight and frontier models, and Roblox fine-tunes the open-weight ones on its own data for the tasks it cares most about: generating game logic, building worlds, creating assets, and reading the player signals that let creators build businesses.

It's worth being precise about what 77% self-hosted means, because it is easily misread as an on-device claim. It is not. This is inference running in Roblox's own data centers, which is a capital expenditure story rather than a story about models getting small enough to run on a phone. The platform is not avoiding the cost of inference. It is choosing to own it.

What makes that possible is not engineering ambition. Nearly 400 models is an unremarkable achievement for a company of this size. It is the training set.

Roblox has more than 20 years of gameplay data, hundreds of thousands of daily active creators, and an average of more than 350 million hours of human interaction per day. 

Its Data Model Recording platform captures gameplay as 4D spatial data down to the instance and frame level, which the letter says lets models learn precise 3D state changes over time from any perspective, from exact telemetry like line of sight, weapon usage and hit locations. The letter's claim is that no other company can replicate these bodies of data: how games are built, which mechanics keep players engaged, and what maximizes player lifetime value.

Models trained on that data produce human-like playtesters and NPCs at higher quality and lower inference cost than frontier models. Roblox is adding purpose-built diffusion models for 3D asset generation, scene verification models, and NPC models trained on human gameplay that playtest and give feedback to the coding agent.

> The report also contains a nicely self-aware line about building art agents for UI, UX and visual assets specifically to avoid generic AI-generated output. A company training its own models so that its AI-generated art does not look AI-generated is the most 2026 sentence in the filing.

Of course, the remaining 23% is where the honest reading lives. Roblox describes a hybrid cloud architecture that couples on-premise capacity with public cloud bursting, allowing it to route each AI prompt and compute workload to the most efficient layer. 

So self-hosting the majority is not a claim that tuned open weights are as capable as frontier models. It is a claim that most requests do not need a frontier model, and that a company which knows its own workloads can tell which ones do. Cheap, high-frequency, well-defined tasks stay in-house. The hard reasoning goes out. 

That is a routing decision, not a quality concession, and it is the part of this that generalizes to everybody else.

However, the bill is also visible. Certain infrastructure and trust and safety expenses, a measure that strips out personnel and depreciation, grew 24% year-on-year to $258 million, which the letter attributes to AI-related investment including training and inference costs. 

The full infrastructure and trust and safety line on the income statement ran $385 million against $321 million a year earlier.

On the call, CFO Naveen Chopra said that a lot of the investments being made in-house now take more of the form of fixed expense as opposed to variable, which is the accounting consequence of owning inference rather than renting it. 

CEO Dave Baszucki said the company will be adding more compute and more GPUs, naming Build specifically, and noted that part of the secret sauce has been running infrastructure very efficiently. Q4 guidance puts $380 million of capital expenditure and intangibles against free cash flow guided to between negative $190 million and negative $105 million.

Some of that cost is about to be passed along. Build, the prompt-to-game tool going to a broader US release in December, will keep a free base tier with paid tiers for additional prompt capacity. Metered prompts are what inference economics look like when they reach the creator.

Roblox Reality is the same bet placed one level higher: the cloud and engine stack coupled to generative video world models, aimed at multiplayer photorealism, with early access scheduled for the first half of 2027\. 

Video world models are the most expensive thing anyone is currently trying to serve interactively, and the argument for Roblox attempting it is identical to the argument for the 77% — twenty years of gameplay data, and somewhere to put it.

Which is the real finding. The moat here is not the models, which anyone can fine-tune, or the infrastructure, which anyone can buy. It is a training set that cannot be bought at any price, including by the frontier labs, because it is the accumulated record of what several hundred million people actually did inside games.