Physical Intelligence Runs Real-Time Remote Inference for Robotic Control on Modal

Physical Intelligence (Pi) runs real-time remote inference for robotic control using Modal's cloud GPU infrastructure. Their Visual-Language-Action model requires continuous inference cycles across a fleet of robots. Rather than using standard TCP-based Modal Tunnels, Pi and Modal co-developed a QUIC-based transport over UDP with automatic NAT traversal, adding only 10-15ms of network overhead. This setup allows Pi to use larger data-center-class GPUs, load model checkpoints in under 30 seconds via Modal Volumes, and expand globally by pinning inference to regions close to their robots—without shipping hardware or building local clusters.