Discount Dynamics: Volume Pricing When Your Costs Refuse to Sit Still

Table of Contents

• The Assumption Volume Discounts Were Built On
• Why Variable COGS Breaks the Classic Discount Curve
• The Three Cost Layers a GaaS Discount Has to Survive
• Discount Structures That Actually Work
• Commit-Based Discounts vs. Realized-Volume Discounts
• Margin-Floored Tiers
• Discounting the Stable Layer, Not the Volatile One
• Modeling a Discount Before You Offer It
• The Negotiation Table: What Procurement Will Ask For
• Insights Most People Overlook
• References

The Assumption Volume Discounts Were Built On

Every volume discount in the history of commerce rests on one quiet idea: it gets cheaper to make the next unit. Print a million flyers and the per-flyer cost drops because the press setup, the plates, and the overhead are already paid for. Sell a million seats of software and the marginal cost of seat number 1,000,001 rounds to zero. That's the economic engine behind "buy more, pay less." The discount is a slice of the efficiency you gain at scale, handed back to the buyer to win the commitment.

SaaS leaned on this hard. Because software's marginal cost was effectively nothing, vendors could be wildly generous on volume and still print 80-plus-percent gross margins. A discount was a sales lever, not a margin decision. You could give away a logo and make it back on expansion.

Now look at an agentic service. When a customer runs 10,000 tasks instead of 1,000, what happens to your cost of goods? It goes up, roughly linearly with usage, and sometimes worse than linearly. Every task is a fresh stack of model calls, tool invocations, retries, and context tokens that you pay a frontier-model provider for in real money. There is no press setup to amortize. The marginal cost of task number 10,001 is essentially the same as task number 11, except when token prices spiked overnight, in which case it's higher.

This is the trap. GaaS founders inherit the instinct to offer volume discounts because that's what "real" software companies do. But they're offering a SaaS-shaped discount on a cost structure that behaves like cloud infrastructure or a manufacturing business. The mismatch doesn't show up on day one. It shows up three months later, when the biggest, most-discounted customer is also the one dragging your blended margin toward zero.

Why Variable COGS Breaks the Classic Discount Curve

Let's get specific about why the curve breaks, because the failure mode isn't obvious until you draw it.

A standard tiered discount looks like this: 0-1k tasks at $1.00 each, 1k-10k at $0.90, 10k+ at $0.80. The implicit promise is that your cost per task is flat (say $0.40), so even at the deepest tier you're keeping $0.40 of gross margin. Clean.

Two things blow that up in GaaS.

First, cost isn't flat, it's a distribution with a fat tail. Some tasks resolve in one model call. Others trigger the retry problem where a single task balloons into dozens of calls because the agent fails a tool invocation, re-plans, and tries again. Your $0.40 "average" is hiding tasks that cost $0.08 and tasks that cost $3.20. When a customer's mix shifts toward the expensive tail, which big enterprise workloads reliably do, because they hand agents the gnarly edge cases, your realized cost per task climbs at exactly the volume where you've promised your lowest price.

Second, the input price moves under you. Frontier-model pricing has trended down over time, but it doesn't move smoothly, and your own costs include a basket of models, tool-call fees, and egress that don't all move together. A reasoning-model price change, a context-window pricing tweak, or a shift in which model your router selects can swing your per-task cost meaningfully week to week. Andreessen Horowitz has written about how AI companies face structurally different margin pressure than classic SaaS precisely because compute sits inside COGS rather than outside it. A discount you signed in January, priced against January's inference rates, is still binding in June against June's rates.

Stack those two and the discount curve inverts. The cheapest price point coincides with the highest cost exposure. You've effectively written a call option to your largest customers on your own margin, and they exercise it by sending you their hardest, highest-volume work.

The Three Cost Layers a GaaS Discount Has to Survive

To discount intelligently, you have to separate your costs by how much they actually move. Treating COGS as one number is what gets founders killed. In practice there are three layers:

Layer 1, Truly fixed. Orchestration infrastructure, your control plane, the salaried humans reviewing edge cases, observability tooling. This behaves like SaaS. It amortizes over volume. This is the layer a volume discount is allowed to eat into, because more volume genuinely makes it cheaper per task.

Layer 2, Variable but predictable. Embedding calls, vector storage, caching infrastructure, the baseline model calls for a "happy path" task. You pay per use, but the per-unit cost is reasonably stable and you can forecast it within a tight band. Caching especially is a lever here, a well-designed cache can cut repeat-context costs dramatically, which is why prompt caching shows up in major providers' pricing docs as a first-class cost mechanism. Discount against this layer only if your caching and routing are mature enough to hold the cost steady.

Layer 3, Variable and volatile. Retry storms, fan-out into sub-agents, long-running "thinking" token consumption, and exposure to provider price changes. This is the fat tail. You must never anchor a discount to this layer. If your discount math assumes Layer 3 stays put, you're not pricing, you're gambling.

The discipline is simple to state and hard to hold: design discounts so they consume Layer 1, lightly touch Layer 2, and are structurally insulated from Layer 3. Most blown-up GaaS deals violate the third rule. They promised a flat per-task price, the customer's workload turned out to be Layer-3-heavy, and the vendor ate the variance.

Discount Structures That Actually Work

Knowing the layers, here are the structures that hold up.

Commit-Based Discounts vs. Realized-Volume Discounts

There's a meaningful difference between "you'll get a discount once you hit 10k tasks" (realized-volume) and "you commit to 10k tasks up front and pay a discounted committed rate" (commit-based). Cloud providers learned this lesson a decade ago and it's the same logic AWS uses with reserved instances versus on-demand.

Realized-volume discounts are the dangerous kind in GaaS, because the customer ramps usage and claims the lower price simultaneously, with no commitment protecting you. Commit-based discounts are far safer: the customer pre-pays or contractually commits to a floor, which gives you a forecastable revenue base to plan compute against. You can even pass some of your own reserved-compute savings through, because committed demand lets you buy compute on better terms. The discount becomes a genuine trade, predictability for price, rather than a one-sided giveaway.

If you offer realized-volume discounts at all, make them apply prospectively to the next billing period's rate, not retroactively to all usage. Retroactive volume rebates are how a single heavy month wipes out a quarter of margin.

Margin-Floored Tiers

Instead of fixing the price at each tier, fix the margin. A margin-floored tier says: "At this volume, your price is the greater of $0.70 or cost-plus-40%." If inference costs spike, the floor protects you; if they fall, the customer enjoys some of the savings. This is a usage-based echo of cost-plus contracting, and procurement teams understand it because it's transparent.

The catch: you need real-time cost attribution to make it work. If you can't tell what a specific customer's tasks actually cost you this month, including their share of retries and fan-out, you can't compute the floor. Margin-floored pricing forces you to build the cost-attribution and dashboarding muscle you needed anyway. Many vendors discover they can't do it, which is itself the finding: if you can't price a margin floor, you don't actually know your unit economics.

Discounting the Stable Layer, Not the Volatile One

The most elegant move is to unbundle. Charge a platform fee (covers Layer 1) and a usage fee (passes through Layer 2 and 3 with a markup). Then apply your volume discount only to the platform fee. The customer feels rewarded for scale, their fixed cost per task drops as they grow, while your variable, volatile costs stay protected behind a markup that floats with your actual spend.

This mirrors how mature cloud businesses structure pricing, and McKinsey has noted that the most durable AI economics come from separating what scales cheaply from what scales expensively. It's less "clever pricing" than honest accounting made visible to the buyer.

Modeling a Discount Before You Offer It

Before you put any discount in a contract, run it through three scenarios, not one. The single-scenario "if cost stays at $0.40" model is malpractice in this category.

1. Base case. Current cost mix, current model prices. Compute gross margin at the discounted rate. If it's already thin here, stop, there's no room.

2. Adverse mix. Assume the customer's task distribution shifts 20-30% toward the expensive tail (Layer 3 heavy). This is the likely outcome for enterprise volume, not the worst case. Does the discount still clear your margin floor?

3. Price shock. Hold the customer's behavior fixed but raise your input costs 30-50%, a plausible swing given how token and inference pricing has historically moved in jumps rather than smooth declines. Does the deal survive a bad quarter for model pricing?

A discount that only works in scenario one is not a discount; it's a liability with a delay timer. The deals worth signing clear all three, or they include explicit contractual valves, a repricing clause tied to published model rates, a cap on Layer-3 cost pass-through, an annual true-up. Build the valve in writing at signing. Trying to renegotiate after the margin caves is a far worse conversation, and customers rarely give back a price they've already banked.

One more modeling discipline: track blended margin across your whole book, not deal-by-deal. A deeply discounted anchor customer can be defensible if they're load-bearing for the rest of your funnel, but only if you've quantified that and you're watching the blend. If you can't see your blended number weekly, you're flying blind on exactly the metric a discount strategy moves.

The Negotiation Table: What Procurement Will Ask For

Enterprise procurement is good at this. They will ask for the structures that maximize their optionality and minimize yours, and they'll frame each as "standard." Know what you're being asked for:

• Retroactive volume rebates, the most expensive thing they can extract. Resist; offer prospective tiers or commit-based pricing instead.
• Price locks with no repricing clause, fine on Layer 1, dangerous on Layer 3. Lock the platform fee, float the usage fee.
• Unlimited / all-you-can-eat at a flat rate, only safe if you've modeled the adverse-mix scenario and can cap Layer 3, or if you quietly cap autonomy to bound per-task cost. Plenty of vendors are already doing the latter to protect margin.
• Most-favored-nation clauses, these turn one bad discount into every customer's discount. Avoid, or scope them narrowly.

The healthy stance is to walk in with a discount menu tied to commitments, not a single number you negotiate down from. When the discount is structurally explained, "deeper commit, longer term, or platform-fee-only, pick your trade", you keep the conversation on terms you can survive. When it's just a percentage, you're negotiating against your own margin with a counterparty who does this for a living.

Insights Most People Overlook

1. The deepest discount should go to the most predictable customer, not the biggest one. Volume is the wrong variable to reward in GaaS. Predictability is what actually lowers your cost, it lets you reserve compute, tune caching, and forecast spend. A medium-sized customer with a steady, Layer-2-heavy workload is cheaper to serve than a giant whose usage spikes and skews Layer-3. Price the variance, not the volume. Almost nobody does this, because "biggest customer gets the best price" is a reflex inherited from a different cost structure.

2. Falling model prices are a reason to discount less, not more. Counterintuitive, but: when input costs drop, your margin on existing contracts improves automatically. Founders see the cost drop and feel generous, passing it through as discounts. That's giving away the one tailwind the category has. The disciplined move is to let cost declines repair the margin damage from your previous over-generous deals before you hand the next one away.

3. A volume discount can increase your cost per task, by changing customer behavior. Cheaper tasks get used more carelessly. Drop the price and customers stop pruning their own usage; they fire the agent at marginal, low-value work and let retries run. The discount you offered to win loyalty can raise the average cost and lower the average value of every task they send. Pricing is a behavior lever, not just a revenue one.

4. "We'll just pass through model costs" is a discount you didn't realize you gave. Pure pass-through pricing feels safe, but it caps your upside at your markup while leaving you fully exposed to the customer's worst tasks. It's effectively a 0% discount with 100% of the cost risk, the inverse of what you want. The margin trap here is real enough that it deserves its own analysis, but the short version: pass-through is not the conservative choice it looks like.

5. The contract clause matters more than the discount percentage. Two deals at the same headline rate can have wildly different outcomes depending on whether there's a repricing trigger, a Layer-3 cap, and a commitment floor. Founders obsess over the number on the slide and skim the terms. In a variable-cost business, the terms are the price. A 20% discount with a model-rate repricing clause is safer than a 5% discount locked flat for three years.

References

• The Economic Case for Generative AI and Foundation Models, Andreessen Horowitz
• The State of AI, McKinsey & Company (QuantumBlack)
• Prompt Caching, Anthropic Documentation