EmbeddingGemma 2 on Cloud Run: Smallest Viable Specs and When an L4 GPU Is Cheaper

This guide sizes EmbeddingGemma 2 text embeddings on Google Cloud Run, testing the smallest CPU instance the model runs on. It also evaluates an int4 weight repack and compares the cost of a document burst on a 1 vCPU service against an NVIDIA L4 GPU. Findings show a 1 vCPU, 2 GiB int8 service handles queries and small batches cheaply, while 1 GiB is too small for longer documents. The L4 becomes cheaper only for bursts above roughly 5,300 documents because it bills idle time, and the int4 repack proved slower on CPU despite being 28% smaller. The author recommends serving routine traffic from the CPU service and deploying an L4 only for large batch jobs that scale to zero afterward.