Decart.

Decart builds infrastructure for live AI. The company works at the layer where models stop being offline artifacts and start responding in real time: inference, optimization, video generation, world models, and the compute stack underneath them.

That distinction matters. A model that is impressive in a batch pipeline can still feel too slow or too expensive to become a product. Decart is focused on the part of AI that has to run at the speed of interaction, where latency, cost, and hardware utilization decide what can exist.

What they're building

The company describes its Decart Optimization Stack as software that runs across NVIDIA GPUs, AWS Trainium, and Google TPUs. The goal is to make large models cheaper and faster to train and serve, without tying the entire system to one chip vendor or one cloud. That is infrastructure work in the practical sense: memory, kernels, scheduling, and the unglamorous machinery that turns compute into usable product.

On top of that infrastructure, Decart is building interactive models. Oasis is a real-time world model aimed at physical AI use cases including robotics, autonomous vehicles, manufacturing, and drones. Lucy is a live video editing model that transforms video as it runs. Both products point at the same premise: AI should be able to generate, simulate, and respond while the user or machine is still in the loop.

Why we backed the founders and team

Dean, Orian, Moshe, and the Decart team are building in the part of AI where demos collide with physics. A real-time model is not just a better checkpoint. It is a systems problem across chips, serving infrastructure, model design, and product constraints. That is the kind of assembly we like.

The frontier is moving from text boxes into worlds: robots, vehicles, games, factories, and live video. Those systems need models that are fast enough to steer reality instead of merely describing it afterward. Decart is building the stack for that moment.