Dreamscale Labs

Dreamscale Labs runs robot AI models in the cloud so robotics teams can deploy large robot foundation models without buying expensive local GPUs. A team chooses a model, connects its robot, and Dreamscale handles the compute, optimizing the model, the GPU kernels, the network path and real time chunking so that inference results reach the robot fast enough to control it. The service is aimed at robotics companies and research groups that want to experiment with, iterate on and deploy frontier robot models. The vendor demonstrates real time execution of NVIDIA's DreamZero world action model over the cloud and lists MolmoAct2 among the models it can run. It states that it can onboard models from providers including NVIDIA, Ai2, Physical Intelligence, Robbyant, Perceptron and Hugging Face, and that teams can bring their own checkpoint fine tuned for a specific robot and task. The company reports 320 ms inference for the 14 billion parameter DreamZero model and more than eight robots running concurrently on a single H100, figures that come from its own benchmarks. Pricing is not published. There is no pricing page, and the site asks prospective customers to schedule a demo or talk through their model with the team, so access is quote based. The offering is positioned as on demand compute for robotics rather than a general purpose GPU cloud, with the performance work spread across the model, kernel, network and chunking layers instead of the model alone.