Dreamscale Labs

Robotics teams building on physical AI models need to run inference for robot control in real time, but existing cloud providers introduce latency that is poorly suited to real-world robot execution. Running models like world action models close to the robot, with optimized compute, is a nontrivial infrastructure problem spanning the model, GPU kernels, and network layer. Teams that want to experiment, iterate, and deploy robot models on demand otherwise have to solve this stack themselves.

Dreamscale Inference is cloud infrastructure for running physical AI models such as DreamZero and MolmoAct2, letting teams choose a model, connect their robot, and have Dreamscale handle the compute. The company says it optimizes across the full stack, including the model, GPU kernels, and network layer, using techniques like Real-Time Chunking (RTC) to bring inference closer to the robot. It reports 40 to 50% lower tail latency than existing cloud providers, 320 ms inference for DreamZero 14B, and the ability to run 8+ robots concurrently on a single H100. Teams can also bring their own fine-tuned checkpoints to run in the cloud.