Size a commit: the maths of
over- and under-committing.

A committed-volume contract is a bet placed twelve months forward: promise the vendor a floor, get a better rate, and pay for the promise whether traffic shows up or not. Sizing it is genuinely just maths — two asymmetric penalty functions applied to your forecast distribution — and doing that maths beats every rule of thumb, because the right answer moves with your overage rate, your growth, and your forecast quality. Here is the calculation, worked.

What a commit really prices

Strip the structure to its variables: a commit volume C at a committed rate r_c, with usage above C billed at an overage rate r_o (typically above r_c, sometimes steeply), and usage below C billed as if it were C — the shortfall is paid for at r_c and delivers nothing. The vendor is selling rate in exchange for certainty; you are buying rate with forecast risk. Two ratios govern everything: how much cheaper r_c gets as C rises (the tier curve — the shape explained in our tiers piece), and the overage premium r_o/r_c. Get both in writing before any sizing, along with the measurement definition (monthly? annual pooled? per-region?), because a commit measured monthly punishes seasonality that an annual pool forgives — a terms difference worth more than most rate differences.

The cost of each mistake, in formulas

For a month with actual usage A: undercommitting (A > C) costs the overage premium on the excess — (A − C) × (r_o − r_c) more than a perfectly sized commit. Overcommitting (A < C) costs the empty floor — (C − A) × r_c paid for nothing. Note the asymmetry: each undercommitted GB costs the premium (often 20–50% of r_c), while each overcommitted GB costs the full rate — so a GB of overcommit typically hurts two to five times more than a GB of undercommit. That asymmetry is the entire intuition: rational commits sit meaningfully below the expected usage, exactly how far below depending on the premium ratio — the smaller the overage premium, the lower you should commit; and if r_o/r_c is large, the vendor is charging heavily for forecast risk, which is itself a negotiation finding.

Sizing against the forecast distribution

Now bring the forecast band, because the optimal C is a property of the whole distribution, not the central line. The clean way to think about it: for each candidate C, expected annual cost = Σ over months of [C×r_c + expected overage at r_o], computed against your monthly forecast distribution (central, low, high — or better, the residual distribution from your forecasting history). In practice a spreadsheet over your twelve monthly scenarios × five candidate commits answers it in an hour, and the answer almost always lands in a recognizable place: C around the low-to-central edge of the forecast — roughly the level you are 70–85% confident of exceeding — higher confidence required when the overage premium is small, lower when it is punitive. One more force pushes C down: growth. A commit sized to December’s traffic overpays from January to November; against a growth curve, size to the year’s weighted expectation or negotiate a ramp (section five), never to the exit rate.

A worked example

Concrete numbers, kept small so the method stays visible — multiply by your scale. Forecast: central case 550 TB/month, low case 470 TB, and your forecasting history says you beat the low case four years in five. Two quotes: commit 400 TB/month at $4.00/TB with overage at $5.20 (premium ratio 1.3), or commit 550 TB/month at $3.70 with the same overage rate. Central month on the 400 commit: 400×$4.00 + 150×$5.20 = $2,380. Central month on the 550 commit: 550×$3.70 = $2,035 — the bigger commit saves $345 a month. Now the downside test, low case at 470 TB: the 400 commit costs 400×$4.00 + 70×$5.20 = $1,964, while the 550 commit still pays its full floor, $2,035 — the bigger commit loses $71 a month. Weight the scenarios (80% central-or-better, 20% low) and the bigger commit wins on expectation by roughly $260 a month while risking only $71 in the bad year — sign it. The point is not these particular numbers: flip the overage premium to 2× or shrink the tier discount to $0.05 and the same spreadsheet flips its answer. An hour with your own distribution beats every heuristic in the industry.

Contract terms that change the maths

Before signing whatever the spreadsheet says, negotiate the terms that reshape both penalty functions. Annual pooling or quarterly true-ups forgive seasonality that monthly measurement punishes. Ramp schedules (commit rising quarterly along your growth curve) remove the size-to-December problem directly. Burst or event carve-outs price your known spikes at committed rates instead of overage — bring the event calendar to the table. A mid-term upgrade clause (raise the commit later at the better tier, no penalty) makes undercommitting nearly free, since growth converts into a renegotiation rather than an overage bill — the single most valuable clause in this family. And in a multi-CDN estate, remember commits interact: two commits must jointly sit inside the split strategy’s movable range, or the split loses its leverage. The commit file — the spreadsheet, the chosen C, the reasoning — goes into the renewal folder, where next year’s negotiation will thank you for it.