- Decision - the call, with the reasoning
- Roadmap - what would have to change, and the triggers to revisit
Do not migrate jina-embeddings-v5 to GCP TPU now. The software path to serve embedding/pooling models on TPU does not exist for 3 of 4 models (encoder / multimodal). For the 4th, v5-text-small, decoder embedding serving shipped on 2026-05-05 (tpu-inference#2420), so its blocker is now integration + economics rather than capability - and TPU's lack of Cloud Run scale-to-zero still makes the economics unattractive for the current low-utilization, per-task topology. Stay on L4 / Cloud Run. Track the remaining software steps and one workload trigger; re-evaluate only when both align.