Commitment Instruments and Break-Even Utilization

A commitment instrument is a contract in which a buyer agrees to pay for a quantity of capacity, or a rate of spend, for a fixed term, in exchange for a price below the on-demand rate. The discount is real and often large, and it is not free. The buyer has converted a variable cost into a fixed one, and the conversion is only favourable if the capacity is used for enough of the term. The relevant question is not whether the discount is worth taking but what utilization the commitment requires before it beats paying list price.

Two shapes of commitment

The two common shapes differ in what is being committed. A reserved-capacity instrument commits to a specific instance family, size, and region for a term, and in exchange the buyer receives either a capacity reservation or a billing discount on matching usage. A spend-based instrument commits to an hourly expenditure for a term and applies a discount to whatever eligible usage occurs, up to the committed amount, without naming an instance type.

The distinction matters because the instruments fail differently. A capacity commitment that does not match the instances actually launched provides no benefit; the buyer pays for a reservation of a type that nothing runs on. A spend-based commitment applies to any eligible usage, so it survives changes in instance shape but not changes in total spend. A workload that shrinks, or moves to a service the commitment does not cover, leaves the buyer paying for a discount it cannot apply.

Both shapes have the same economics underneath: the buyer prepays or commits to a stream of payments, and receives a lower marginal price for eligible usage. Whether the exchange is favourable is a function of utilization and term.

The break-even is a ratio of two rates

Consider a commitment that costs a fixed amount per hour for a term and entitles the buyer to a discounted rate on usage. If the usage that would otherwise occur at list price is large enough that the discount on it exceeds the commitment’s cost, the commitment pays. The threshold is a quantity of usage per hour, and expressing it as a utilization fraction of the committed quantity makes it comparable across instruments.

Let the on-demand rate be 1, the committed rate be d (a fraction of list), and the commitment cover a capacity of C units. The buyer pays d·C per hour whether or not the capacity is used. Usage of u units at list would have cost u. With the commitment, the cost is d·C for the covered portion and 1·(u − min(u, C)) for any excess. Break-even occurs where the two are equal:

def break_even_utilization(committed_rate):
    """Fraction of the commitment that must be consumed to beat on-demand.

    committed_rate is the committed price as a fraction of list (0.7 for a 30% discount).
    Paying committed_rate * C beats paying 1.0 * u once u exceeds committed_rate * C.
    """
    return committed_rate

def hourly_cost(units_used, cover, discounted_rate, ondemand_rate=1.0):
    covered = min(units_used, cover)
    excess = max(0.0, units_used - cover)
    return cover * discounted_rate + excess * ondemand_rate

The result is obvious once written down and easy to invert by mistake: the break-even utilization equals the committed rate as a fraction of list. A thirty percent discount sets the committed rate at 0.7 of list, so the commitment beats on-demand once seventy percent of the committed quantity is consumed. The remaining thirty percent is margin, and the margin is what makes commitments attractive on stable workloads. A commitment taken against a workload that will consume only half of it loses money, and it loses exactly the difference between the two rates times the unused portion.

The function ignores the term, which is a mistake in practice. The commitment’s cost is incurred over the whole term, so the relevant utilization is the average across the term, not the utilization in the month the decision was made. A workload that consumes the full committed quantity for six months and nothing for the other six averages fifty percent, which is below the seventy percent threshold, so the commitment loses money despite having been fully utilized for half its life. A workload that grows into the commitment is better than one that shrinks out of it, because the break-even is a floor and growth adds margin.

The commitment is sunk, so marginal decisions change

Once the commitment is purchased, its cost is fixed for the term. The marginal price of additional eligible usage drops to the discounted rate for the covered quantity and returns to list beyond it. This changes the answer to questions that were previously straightforward.

A workload that would be uneconomic at list price may be economic while commitment coverage is available, because the alternative is leaving paid-for capacity idle. Running a batch job on capacity already committed costs the discounted rate and nothing else, so the correct comparison is against the discounted rate rather than against list. Conversely, a workload that consumes the last of the coverage has a marginal cost of list, and adding it may be more expensive than it appears from an average rate.

The same logic applies in reverse when a commitment is being sized. Sizing to the peak of a workload means paying for capacity that is idle most of the year. Sizing to the floor, and paying list for the peaks, is usually better because the discount applies to the portion that is always present and the variable portion keeps its flexibility. The floor is the quantity that never goes away, and it is the only quantity a commitment should cover.

Utilization is measured against the wrong denominator

Commitment coverage is commonly reported as spend covered divided by total spend. That figure is easy to compute and answers a different question. What matters is whether each hour’s eligible usage exceeded the committed quantity, and whether the committed quantity was fully consumed. A high coverage percentage can coexist with substantial waste if the commitment is concentrated in hours when usage is low.

The useful measures are the number of hours in the term during which the commitment was not fully consumed, and the total unconsumed quantity across those hours. The first measures how often the commitment was idle; the second measures how much was wasted. A commitment that is fully consumed in ninety percent of hours and idle in ten percent has a utilization of roughly ninety percent, and the waste is the unused portion in those ten percent of hours. Reporting the hourly series rather than a monthly average is what makes the waste visible, because a monthly average hides the alternation between saturation and idleness.

The hourly series also has to be built from usage records rather than from the invoice, because the invoice reports a period total in which a fully consumed hour and an idle hour are already summed; the ordering of discounts described in billing units and why a cloud bill is an estimate means the total cannot be decomposed back into hours after the fact.

A second denominator problem is scope. Discounts apply to eligible services and regions, and usage outside that scope does not consume the commitment. A buyer who commits against compute and then shifts work to a managed service sees the commitment go unconsumed while the managed service bills at list. The commitment was correct when sized; the workload changed shape underneath it. This is the structural risk of any term commitment, and it is the reason commitments are best taken against the most stable and least substitutable part of the estate.

Term length trades discount against irreversibility

Longer terms carry larger discounts and greater exposure. The trade is asymmetric because the buyer cannot know future usage, and the seller prices the discount partly as compensation for that uncertainty. A one-year term at a moderate discount and a three-year term at a larger one differ in the value of the option to stop paying, and that option is worth more when the workload is less predictable.

A reasonable construction is to cover the floor with the longest term the organization can commit to and to leave the variable portion uncovered. The floor is the quantity that survives replatforming, reorganization, and changes in traffic, so it is the part least exposed to the irreversibility. Covering the peak with a long commitment is the error that produces the classic outcome of paying for capacity that a migration made unnecessary.

There is also a timing consideration that is easy to miss. Commitments are purchased at a moment, and the price list at that moment may be superseded. A commitment priced against a list that later falls delivers a discount relative to the old list and possibly a premium relative to the new one. The exposure is real but small in ordinary cases; it becomes large when a provider reprices a family or introduces a cheaper generation, because the committed rate does not follow.

What a commitment does not buy

A commitment buys a price, not a guarantee of availability. The capacity-reservation form is the exception, and even there the reservation holds capacity in a specific zone, so a zonal failure removes access to the reserved capacity along with everything else. The spend-based form makes no availability promise at all.

Nor does a commitment reduce consumption. The total quantity of capacity used is determined by the workload and its provisioning policy, and the commitment only changes the rate at which that quantity is priced. A cost programme that begins with commitments and never addresses provisioning has optimized the numerator of a ratio while leaving the denominator to grow. The denominator is set by the policies described in autoscaling policies and the cost of headroom, and the interaction between a fixed commitment and a variable workload is where the two mechanisms can work against each other: an autoscaler that scales down below the committed quantity wastes the discount it was purchased for.