LLM API prices / Self-hosting

EmbeddingGemma 2 self-host cost

Google's October 6, 2026 developer guide describes EmbeddingGemma 2 as an open multimodal embedding model with modular encoders. This page estimates rental cost only. A defensible cost per million embedded items needs your measured throughput, batching, utilization and total input length. There is no published universal token throughput or cloud rental requirement to insert into the formula.

Rental arithmetic

For a rented GPU, monthly compute rent equals hourly rate × billable hours. At 24 hours a day for 30 days, that is hourly rate × 720. The GPU cloud table contains sourced hourly offers and last verified dates. For example, its lowest listed hourly offer was $0.09/hour when checked on 2026-09-26, so 720 paid hours would be $64.80 before storage, CPU, network and taxes. Official GPU price source. This is a rental example, not an EmbeddingGemma 2 performance claim.

Measure before comparing with an API

Benchmark the actual model variant and input mix on the target GPU. Count items per second at the batch size and latency target your application needs. Then monthly capacity equals measured items per second × 3,600 × paid hours × utilization. Divide total rental and supporting infrastructure cost by that capacity to obtain cost per item. If your service runs part-time, use actual paid hours; if it requires availability around the clock, include idle hours.

A text-only index and an image, audio or video index have different preprocessing and memory demands. Embedding generation also needs storage for vectors, an index service, updates and recovery. API pricing may include some of those operational costs while raw GPU rent does not. Compare like workloads and verify an API provider's current embedding rate separately before claiming savings.

Last GPU price check: 2026-09-26. Use the VRAM planning tool to check a candidate accelerator and the throughput calculator with measurements. The LLM API token calculator covers generation model token charges, which are a different workload from embeddings.