Our Vertex AI bill was $8K/month. Here’s the architecture change that cut it by 71%

Four second query latency. Eight thousand dollars a month. We thought it was the model. It wasn’t. It was orchestration — and the way Google’s default setup is quietly expensive by design. We rebuilt the request pipeline, moved non-critical tasks to async, switched model tiers where quality tolerance allowed it. The result: 71% cost reduction, latency down to under a second. Google doesn’t document this path clearly. This is the breakdown we wish we’d had:…