Serving Gemma 4 2B on a Single TPU v5e Chip with MCP and Antigravity CLI

This article is the v5e follow-on to the v6e-1 debugging guide. Same MCP tooling, same Antigravity CLI driver, smaller and cheaper silicon — and a different set of failure modes. This time the model is google/gemma-4-E2B-it on a single Cloud TPU v5e chip, with measured throughput, measured latency, and a cost breakdown that actually uses the numbers from the sweep. What this project is trying to do Same brief as before: a DevOps/SRE assistant whose brain is a self-hosted Gemma 4 model. The MCP server provisions the TPU, deploys the vLLM container, discovers the endpoint, and then uses that endpoint to analyze Cloud Logging output. 31 tools, one server.py, stdio transport. What changed is the target: TPU v6e-1 → TPU v5e-1, and Gemma 4 4B → Gemma 4 2B. The interesting question this article answers: v5e is roughly half the price of v6e per chip-hour. Is it half the machine, or worse? The chip, on paper Spec (per chip) TPU v5e (v5litepod) TPU v6e (Trilliu