Credit breakage is the revenue a vendor keeps from credits a customer bought but never consumed. The customer prepaid for capacity, the capacity expired or was forfeited under no-refund terms, and the vendor recognized the money without running the LLM inference behind it. Every prepaid model produces some breakage; gift cards are built on it. Credit-based AI software pricing manufactures it structurally, through contract terms that rarely appear on the pricing page.
The seller-side verdict belongs up front: breakage is brittle revenue. It books today and gets clawed back at renewal, because procurement arrives holding the consumption report.
What Credit Breakage Is (and Where It Hides on the Pricing Page)
A credit is a surrogate unit: a vendor-defined billing unit that folds several underlying value metrics into one number, with vendor-controlled conversion ratios between them. As we covered in what a credit in AI pricing actually is, the conversion table is the vendor’s margin lever. Re-rating it at renewal is the subtle setting of the lever. Breakage is the bluntest: the vendor keeps the entire balance and delivers nothing against it.
The economics never appear where the buying decision happens. Pricing pages publish the credit allocation per plan and the price per pack; expiration windows, reset schedules, refund treatment, and minimum purchase sizes live in terms-of-service documents and billing FAQs. A vendor that expected these terms to survive an informed renewal conversation would publish them next to the list price.
The Four Mechanics of AI Credit Expiration and Breakage
Breakage requires no single aggressive term; it accumulates from four mechanics that compound quietly.
Expiration windows
Purchased credits carry a shelf life; promotional credits carry a shorter one. GitHub’s promotional credit top-ups for Business and Enterprise Copilot plans expire on September 1, 2026, when allotments return to standard levels. At least one major platform expires purchased credit packs a year after purchase. The industry’s standing defense is accounting: credits must eventually expire or the deferred revenue never resolves. The defense does not survive contact with the accounting. Revenue recognition handles balances that never expire, recognizing expected breakage in proportion to redemptions, or when redemption becomes remote; gift cards resolve this way every day. Whether an expiration boundary exists at all is a commercial choice.
Monthly resets with no rollover
The allowance version of the same mechanic. Credits allocate monthly and vanish monthly; one AI vendor’s published terms grant a one-month rollover, so unused credits expire at the end of the following billing period, the same forfeiture on a short delay. Because usage is skewed, a monthly reset guarantees forfeiture for the light months every real customer has. The vendor prices the allowance as if consumption were level. Consumption metrics never are: usage swings month to month in ways count metrics like users, sites, or locations do not. A level allowance wrapped around a variable metric manufactures forfeiture by design.
No-refund terms
The backstop that converts forfeiture into recognized revenue. The same published terms state that no refunds are issued for unused or expired credits on any plan. No-refund language is close to universal in credit-based AI pricing, and it separates breakage from float: without it, unspent credits are a liability the customer can unwind; with it, they are margin the moment the window closes.
Minimum purchase blocks sized above consumption
The quietest of the four. Packs and allowances are sized on average usage, and averages in skewed distributions describe nobody: a pack sized for the mean customer is oversized for most, so the forfeiture is designed in before the first credit is spent. Heavy users hit overage; light users fund the breakage pool. A pack size that looks like a packaging convenience is doing undisclosed pricing work.
Two billing platforms serving this market have named breakage outright as a built-in consequence of prepaid credit models, a signal that the terms have hardened into a category norm, not an edge case.
Which of the Four Breakage Mechanics Is Quietly Compounding in Your Deals?
Expiration windows, forfeiture terms, and rollover restrictions each hit your licensing, packaging, and pricing differently. Find out which mechanic your architecture is most exposed to.
Why Breakage Is Brittle Revenue
Breakage looks like margin in the quarter it books. It behaves differently across the customer lifecycle, in two ways.
The first is the renewal clawback. Procurement pulls the consumption report before renewal, and a forfeiture line is one it will not miss: capacity paid for and never delivered, quantified to the credit. The predictable outcome is a harder renewal that hands back the forfeited value one way or another. That concession is not sales reps giving ground; it is the cost of an architecture gap, because a commitment sized above consumption is a licensing-model defect, and no negotiation posture fixes a defect.
The second costs more than the clawback, and it is a pattern we have watched play out in client engagements: the customer who saw credits expire in March starts rationing in April, and the product’s surface area shrinks to the workflows someone can pre-justify. We have written about how variable pricing suppresses the exploration vendors need while AI use cases are still being discovered, and what happens when rationing hardens into a token diet. Breakage is the sharpest trigger because the customer is watching money already spent evaporate. A pricing model that trains customers to use the product less, in the exact phase when the vendor needs usage to prove value, works against its own expansion revenue.
Breakage revenue does not recur, and it damages both the renewal and the adoption curve: booked once, paid for twice.
The Architecture That Replaces Breakage
Replacing breakage means sizing and terms that keep the commitment connected to consumption, which protects the renewal instead of mortgaging it. These are licensing-model decisions: the value metric and the terms wrapped around it (expiration windows, rollover, refund treatment, purchase sizing) sit ahead of packaging and ahead of the pricing model.
The test is the consumption report. Put your own terms next to the report your customer’s procurement team will pull before renewal. Every line that reads as capacity paid for and never delivered is funding next year’s discount. An architecture that survives the report has three properties, and each is a design outcome, not a clause to copy: commitments sized against what customers actually consume rather than what makes the deal look bigger; one coherent term length rather than an annual commitment with monthly confiscation running inside it; and overrun treatment that does not punish the customer for succeeding with the product.
Getting those properties into an order form is where the real work lives, because each one interacts with the value metric, the conversion table, and the Customer Group it applies to. The direction, though, is not ambiguous: the margin story moves from forfeiture, which procurement can see and attack, to a well-chosen value metric and a defensible conversion table, both far harder to negotiate against. If you have not measured your own breakage exposure, the pricing architecture assessment will surface it in the licensing and packaging lenses before your customers do.
If you are seeing forfeiture lines in your billing data, or renewal conversations that open with a consumption report, describe what you are seeing to a pricing expert. A pricing architecture expert replies, and the pattern is often diagnosable from the terms alone.
