Decart builds foundational models and systems for real-time inference, focusing on world models and live video generation at millisecond-level latency. The company operates across the full computational stack - from low-level kernel and hardware optimizations through model architecture - to achieve persistent generative capabilities with constrained compute budgets. This systems-first approach addresses a core bottleneck in interactive AI: the latency and efficiency tradeoff that has historically made continuous, low-latency inference either prohibitively expensive or technically infeasible.
The company's products demonstrate this infrastructure focus. Oasis generates interactive 3D environments from user input, while Lucy performs pixel-level edits to live video streams in real time. Both require solving multiple overlapping constraints: maintaining sub-100ms end-to-end latency, sustaining inference continuously rather than on-demand, and reducing per-inference compute requirements relative to conventional approaches. The efficiency target - roughly 100x improvement over current real-time generation - reflects the magnitude of the systems work required, not merely algorithmic gains.
Decart targets interactive domains where latency directly degrades experience: gaming, robotics simulation, and synthetic data generation. The company operates an "AI grid" infrastructure model, positioning edge inference and GPU acceleration as load-bearing components of the architecture rather than optimization afterthoughts. This reflects a production-oriented perspective: real-time systems require rethinking where computation happens, not just making existing models faster.