Committed Spend and Volume Discounts: When to Sign

A committed-spend agreement trades a discount for a promise. The discount is easy to see and the promise is easy to underweight, which is why the arithmetic here is worth doing before the call rather than after it.

The shape of the agreement

Terms vary between providers and none of the specifics below are quoted from anyone — this is the general structure such agreements take, and the point is to reduce whatever you are offered to these variables before valuing it.

Valuing it against an uncertain forecast

Let S be your actual spend at list prices over the term — a random variable, because you do not know it. Without a commitment you pay S. With one, you pay the greater of the commitment and your discounted consumption:

pay_committed = max( X,  S * (1 - d) )
saving        = S - max( X, S * (1 - d) )

which splits into two regimes:

  S*(1-d) >= X    you consumed the commitment:  saving = S * d
  S*(1-d) <  X    you underconsumed:            saving = S - X   (negative)

The asymmetry is the whole story. The upside is capped at d times your spend. The downside is uncapped in the sense that it grows one-for-one with the shortfall — every dollar you fail to consume is a dollar lost, not a fraction of one. A 15% discount is being paid for with a 100% exposure on the miss.

Value it with a two-point forecast, which is all the precision anyone honestly has. Two points is not a shortcut for lack of data; it is an honest representation of what you know. Pick a pessimistic case you would be genuinely disappointed by — a flat quarter, a churned customer, a feature that slips — and an optimistic case you would be pleased with, then assign them rough probabilities. Anyone objecting that the probabilities are guesses is correct and has missed the point: the decision is not sensitive to whether the split is 60/40 or 65/35, and it is extremely sensitive to whether the low case consumes the commitment. Assume a one-year commitment of X = $120,000 at d = 0.15, forfeited if unused, and a forecast of a 60% chance of $150,000 of spend and a 40% chance of $70,000 — all figures illustrative.

high case  S = 150,000
  S*(1-d) = 127,500 >= 120,000   -> saving = 150,000 * 0.15 = +22,500

low case   S = 70,000
  S*(1-d) =  59,500 <  120,000   -> saving =  70,000 - 120,000 = -50,000

E[saving] = 0.6*22,500 + 0.4*(-50,000)
          = 13,500 - 20,000
          = -6,500        do not sign this one

Commit to your pessimistic case

The instinct after that result is to walk away. The correct move is to commit less. Keep the same discount and set the commitment so that even the low case consumes it: X ≤ S_low × (1 − d), which here is 70,000 × 0.85 = $59,500.

X = 59,500, d = 0.15

high case  S = 150,000 -> 127,500 >= 59,500 -> +22,500
low case   S =  70,000 ->  59,500 >= 59,500 -> +10,500

E[saving] = 0.6*22,500 + 0.4*10,500 = 13,500 + 4,200 = +17,700

Half the commitment is worth $24,200 more in expectation than the full one, on identical discount terms. The reason is that the discount applies to consumption, not to the commitment, so committing beyond what you will certainly consume buys nothing and risks everything. This generalises to a rule you can apply without rebuilding the model each time:

Commit at your p20 forecast, not your p50. The p50 is the level at which you have a coin-flip chance of eating a shortfall. The p20 is the level you are fairly confident of exceeding, and above it the marginal discount is being bought with pure downside. If the provider will only offer the discount at a higher tier, the question to price is whether the incremental d covers the incremental risk — which is exactly the expected-value calculation above, run twice.

Tiered structures need the same treatment run twice rather than a different method. If a larger commitment buys a better rate — say d improving from 15% to 20% at double the commitment — compute the expected saving at each tier and compare the two numbers. The extra five points apply only to consumption you were going to have anyway, so the incremental gain in the high case is small, while the incremental exposure in the low case is the whole of the additional commitment. Tiers almost always look better than they are, for exactly the reason the base case did.

Two structural notes. If unused commitment rolls over, the low-case loss shrinks to a timing problem rather than a write-off and you can commit far more aggressively. And if the agreement has a true-up rather than a forfeit — you pay list on anything above the commitment and simply do not benefit — the downside disappears and the analysis becomes trivial. Establish which of the three you are being offered before discussing the number.

The term risk nobody prices

Per-token prices for a given capability level have historically fallen, sometimes sharply, as models are superseded and providers compete. A committed price is only a discount for as long as it stays below the list price it was measured against.

your effective price over the term:  (1 - d) * P0
what an uncommitted buyer pays:     (1 - g) * P0

  g  the fraction by which list price falls during the term

You are ahead only while  d > g.

Which turns a 20% discount on a 24-month term into a bet that list prices will not fall more than 20% in two years. That is not an obviously winning bet in this market, and it is not a reason never to commit — it is a reason to prefer shorter terms, and to ask whether the agreement re-prices if list prices drop. Some do. Asking costs nothing.

The clauses worth more than the discount

At the point where you are negotiating, several things are on the table that are worth more than a few percent off tokens, and they are frequently easier to get.

• Rate limits and capacity. A committed rate limit, or reserved capacity during your peak, is worth more than a discount if throughput is what constrains your product.
• Data handling and retention. Written terms about training, retention and regional processing are what unblock enterprise customers. A procurement blocker removed is worth more than any d.
• Model lifecycle notice. How much warning before a model you depend on is deprecated. This is the term that determines whether a deprecation is a sprint or an emergency.
• Scope across models. A discount that follows you to the successor model protects you against your own roadmap.
• An exit. What happens if you need to leave, and whether the commitment survives. Ask before signing, not after.

And a prerequisite for all of it: the analysis above is only as good as the forecast underneath it. If you cannot produce a defensible spend forecast with a low case, you are not ready to commit, because the low case is the only input that matters.