How-to/ 5 min read Inference is not training: sizing for serving, not for the run Training is throughput-bound and batch-friendly. Serving is latency-bound and bursty. Sizing one like the other is how GPU bills get strange. Sep 2, 2026 compliance/ 5 min read Fine-tuning an open-weight model in one EU region Keeping a fine-tune in one jurisdiction is an operations problem, not a legal one. The seven places training data actually leaves — and how to close them. Sep 2, 2026 How-to/ 6 min read Sizing a GPU instance: what to put around the accelerator A GPU is only as fast as the CPU, memory and disk feeding it. How to size the rest of the instance so the accelerator is the bottleneck. Sep 2, 2026 Cost/ 6 min read Moving a training pipeline off a hyperscaler without stopping it You cannot pause research for a migration. A sequenced move that keeps runs going: data first, then one reproduced run, then a shadow period. Sep 2, 2026 Cost/ 5 min read What a training run actually costs The hourly rate is rarely more than half the bill. The five line items that make up the rest, and how to measure cost per completed run. Sep 2, 2026