Hey everyone Benedict here, one of the founders of Jungle Grid. This started from a pretty frustrating pattern while running AI workloads: Most failures weren’t model issues they were infrastructure problems. Picking the wrong GPU → OOM Choosing the wrong region → latency or availability issues Jobs failing due to CUDA / environment mismatches Or just sitting in queues waiting for capacity We realized we were spending more time managing infra than actually running models. So we asked a simple question: What if you didn’t have to think about GPUs at all? That led to Jungle Grid. Instead of selecting hardware, you describe the workload: * inference, training, batch * model size * whether you care more about cost or speed From there, the system: * classifies the workload * routes it across available GPUs (4090 → H100) * handles retries and failover automatically * shows a cost estimate before execution The biggest shift for us has been mental: you stop thinking in GPUs and start thinking in outcomes. We’ve already run 500+ workloads through the system, and it’s been shaped heavily by real failures and edge cases we kept hitting along the way. Still early, and we’re actively iterating. If you’ve dealt with GPU selection, failed jobs, or infra headaches I’d really like to hear how you’re solving it today.