PatakoLabs · publication

AI Cost Lab

How teams actually cut inference spend — not just watch the invoice.

Issue 01 · September 2026

The bill is written in the harness.

Cloud FinOps can still negotiate reservations and tag accounts. Generative AI spend mostly moves when engineers change models, prompts, caches, retries, and agent loops. These notes split those levers — with extra weight on Amazon Bedrock.

Articles

  1. Bedrock · FinOps + harness Amazon Bedrock spend playbook A sequenced playbook for Bedrock: cache economics, client-error waste, routing, cross-region, batch, reserved capacity, distillation, and how FinOps should read CUR. 2026-09-01
  2. FinOps × engineering Who owns the AI bill: FinOps levers vs engineering levers Which generative AI cost moves a central FinOps team can actually pull, and which ones only exist in prompts, routing, and the agent harness. 2026-09-01