Owning the compute
The AI work across Labs depends on inference capacity, and renting it means cost that scales with success and a dependency on providers whose terms and availability can change.
Titan Infrastructure is a long-term programme to bring that capacity in house for AI, backend services, rendering, simulation and research.
What is worth owning
Not everything should be self-hosted. Steady, predictable workloads favour owned hardware; bursty and unpredictable ones favour rented capacity.
Inference for live game features is the clearest candidate: continuous, latency-sensitive, and expensive to rent at scale.
Beyond inference
Owned compute also enables work that is impractical to rent: training and fine-tuning, large-scale simulation, and rendering that is not economical by the hour.
Edge networking matters because latency to a model is latency in the conversation, and a fast model reached over a slow path still feels slow.