PRODUCTION SCENARIO
An organization decides to self-host an open-weight model so it controls the deployment. A sponsor assumes that having the weights removes the ongoing cost.
Which cost does self-hosting still carry?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Accelerator capacity and the staff to keep it running
Open weights remove the licence and the per-call bill; they do not remove the service. Self-hosting means paying for the accelerators, keeping the endpoint up, monitoring quality and patching it - the operating work a managed provider was doing on your behalf. A per-token licence fee is the most common wrong expectation and describes proprietary licensing, not open weights. A support contract is something you can buy for a self-hosted model, not a condition of using one. Retraining on every Google release is the other operational misconception: open weights are a snapshot you control, so moving to a newer Gemma version is a choice you schedule, not an obligation that arrives with the model.
Official Source: https://ai.google.dev/gemma/docs ↗