Roblox VP of AI on 25M concurrent users, 3D generation models, and safety at scale

Summary

Roblox hit 25 million concurrent users this week, a new platform record, and a single game simultaneously reached 16 million players, claimed as a Guinness World Record for a single title. Anu Pom, VP of AI and Infrastructure at Roblox, attributes the platform's resilience to an engineering culture where every executive, including Pom himself, rotates through on-call shifts alongside the site reliability team. The company runs weekly stress tests internally called 'Taco Tuesdays,' deliberately starving specific services of capacity to simulate overload, alongside a chaos testing program designed to simulate data center failures.

AI Architecture Across Three Layers

Roblox structures its AI deployment across three distinct tiers. The first is classic machine learning, used for years to power homepage recommendations and content discovery. The second is large model inference at scale, which Pom acknowledges remains economically challenging. At 250,000 invocations per second, even a fraction-of-a-cent cost per call becomes commercially unworkable, so the team uses model distillation and quantization to compress large models into smaller, cheaper inference targets.

The third and most strategically differentiated layer is safety. Every chat message on the platform passes through a content filter, and Roblox has moved from traditional ML-based filtering to deploying large models with persistent memory of user chat history, enabling detection of behavioral patterns over time rather than single-message snapshots.

Proprietary 3D Generation Model

Approximately 18 months ago, Roblox began building its own large model for 3D content generation, a domain it argues existing generative AI stacks, which stop at text, image, and video, do not address. The model is open-sourced and available on Hugging Face, allowing developers to generate full 3D scenes from text prompts. The intended workflow targets game creators who would otherwise spend hours in tools like Blender or Cinema 4D.

NPC and Creator Infrastructure

Rather than prescribing how NPCs should behave, Roblox exposes large models as a managed service, abstracting away inference costs and model selection from creators entirely. Developers then use that infrastructure to let players build and configure their own NPCs. The platform's philosophy remains API-first, pointing to emergent hits like Dress to Impress, Grow a Garden, Fish, and Dead Rails as evidence that creator-led development consistently outperforms top-down product design. A new all-time high game has emerged roughly every two to three weeks in recent months.

Scale and Capacity Planning

Roblox operates 24 data centers globally and plans infrastructure at the platform level rather than forecasting individual game breakouts, which Pom describes as a deliberate choice to avoid artificially capping growth expectations. Founder Dave Baszucki has noted that the platform has set a new concurrent user record roughly every other weekend for the past 16 years.