“Fireworks and Cursor's distributed training system runs across 5 to 6 global data center regions, utilizing scattered GPUs to run massive reinforcement learning jobs.”