When AI spend starts to hurt, the provider’s account manager shows up with a reasonable proposal: commit to an annual volume and the unit price drops. On paper it is free money. In practice it is a contract that ties you to a level of consumption at the worst possible moment, precisely when optimization should be bringing it down.
At a glance
| What you are offered | A lower unit price in exchange for committed volume |
| What it locks in | Your consumption, not just your price |
| Risk | Optimizing stops paying off, or you pay for what you do not use |
| Rule | Commit only the floor you would consume anyway |
The three shapes of commitment
The volume discount is the most common: a better price if you commit to spending a given amount. Miss it and you either pay the difference or lose the discount.
Reserved or provisioned capacity is a different animal: you reserve compute for a period and pay to have it available, used or not. It buys stable latency and a predictable price, and in exchange it charges you for the peak while you live in the trough.
Prepaid credits are the soft version: you pay up front at a better price, with an expiry date. The risk is not the price, it is the calendar.
The underlying conflict
A volume contract and an optimization project pull in opposite directions. If you have signed for an annual consumption and then cut spend by 40% with per-task routing and prompt caching, the saving does not reach your account: it reaches the committed minimum.
That is why order matters. Measure and optimize first, then negotiate on the real, already optimized consumption. Signing first is buying yourself the right not to improve.
How to decide with numbers
Take your consumption over the last six months and separate two figures: the floor, what you consume even in the weakest month, and the peak. Commit at most the floor, and only the part you know you will not be able to optimize away within the term of the contract.
Then look at duration. A twelve-month commitment in a market where list prices move several times a year can leave you paying above the public price halfway through. It is worth asking for a clause that drops you to list price if list price falls, and a review at six months.
And always translate the discount into cost per task, not into a percentage off list. A 20% discount on a model you should not be using is a bad deal that looks good.
Frequently asked questions
Should you never commit volume? You should, sometimes: when consumption is stable, already optimized, and the commitment covers only the floor. There the discount is real money.
What about reserved capacity for latency? That is a product decision, not a cost decision. If you need guaranteed latency, compare it against what the same service would cost on demand, including the quiet hours you will pay for empty.
Can I negotiate without committing? Often yes, especially if your volume is growing. It is worth asking for a price before accepting the commitment, and comparing against the public rates of the major providers.
At e-ficient we measure the cost per task of companies that already run AI in production and tell you how much you would save with each lever at your real volume. The first audit is free if you start a plan with a three-month commitment and returns a diagnosis within 72 hours.