Core Automation details AI training run across clusters about 9,000 miles apart

Core Automation describes a training run that used two GPU clusters about 9,000 miles apart for a roughly 32-billion-parameter mixture-of-experts model. The company says the clusters had different accelerators, GB300 and B300, as well as different parallelism schemes and total batch sizes. A bidirectional 25 Gbps link connected the sites.

Rohan Anil, who says he works on training algorithms at CoreAuto, shared the write-up Nov. 2, 2026.

Core Automation frames the run as a test of how to use working clusters that are scattered rather than colocated. It distinguishes that architectural problem from getting power and equipment into an unfinished data center, a physical problem it says software cannot fix.

At equal compute, that the two-site run reached about 1% lower bits per byte than standard training on both of its evaluations. Its general-coding comparison lists about 0.2525 bits per byte for the multi-cluster run and 0.2549 for standard training. Those are the company's evaluation figures.