Physical Intelligence opens Pi API beta, cloud inference hits 208ms edge

What happened

Physical Intelligence opened a Pi API beta on October 13 that lets companies upload robot data, fine-tune on PI's compute, and run inference through a hosted gateway; the author adds PI's measured 108ms remote-H100 inference to a 100ms Azure East Japan–US West round trip to get roughly 208ms, or d=10 at 50Hz — exactly the edge of PI's trained 0–10 step delay range.

Why it matters

That roughly 208ms round trip sits right at the boundary of the delay PI's model was trained to handle, so companies running robots from overseas cloud capacity have little margin before performance may degrade, according to the author's calculation.