Valuation Haircuts When Model Costs Compress Margins: How Investors Are Repricing Agent Companies

Table of Contents

• The Core Tension: Why Cheaper Models Can Hurt Your Valuation
• How the Haircut Math Actually Works
• Who Gets Hit Hardest
• Per-Outcome Pricers With Thin Buffers
• Single-Model Dependents
• The "We'll Make It Up on Volume" Crowd
• What Investors Actually Diligence Now
• How Founders Defend the Multiple
• The Counterintuitive Case: When Cheaper Models Help Valuation
• Insights Most People Overlook
• References

The Core Tension: Why Cheaper Models Can Hurt Your Valuation

There's a comfortable assumption floating around agent fundraising decks: model costs are falling roughly an order of magnitude a year, so today's brutal gross margins will fix themselves on a timeline. Just wait for the cost curve. It's the AI-era version of "we'll grow into the multiple."

The problem is that falling model costs are not a private windfall for the agent company sitting on top. They're a public good, available to every competitor on the same day the price drop ships. And the savings rarely flow cleanly to the bottom line, because three other parties have a claim on them: customers who negotiated price-downs tied to your cost, competitors who'll undercut you the moment their COGS falls too, and the model provider itself, which can raise prices, deprecate the cheap tier, or change rate limits whenever it likes.

So when a Series C investor or an acquirer looks at a GaaS company whose gross margin is, say, 45% on a revenue base that looks SaaS-like, they don't value it like SaaS. They ask a sharper question: if the model price moves against you, where does that 45% go? If the honest answer is "down to 25%," the revenue multiple gets a haircut. This is the heart of the revenue-quality question that runs through the whole GaaS funding conversation, usage revenue carrying volatile COGS is structurally worth less than the same dollar of subscription revenue with stable COGS.

The word "haircut" matters here. Nobody is saying the company is worthless. They're saying a dollar of this revenue should be capitalized at a lower multiple than the comps suggest, because the margin underneath it is exposed.

How the Haircut Math Actually Works

Let's make this concrete, because the abstraction hides the mechanism.

Take two companies, both doing $10M ARR, both growing 120% year over year. Company A is a vertical SaaS tool with 80% gross margin and inference that's a rounding error. Company B is a GaaS agent that bills per resolved ticket, runs heavy multi-step model calls per outcome, and lands at 50% gross margin.

A naive analyst multiplies both by the same revenue multiple. A serious one doesn't. The standard adjustment is to value on gross-margin-adjusted revenue, effectively, gross profit dollars scaled by an efficiency factor, rather than top-line ARR. Bessemer's long-running work on cloud and the broader move toward valuing software businesses on gross profit rather than raw revenue gives you the template. Company B's $10M at 50% margin throws off $5M of gross profit; Company A's throws off $8M. Right there, before anyone touches a multiple, B is worth roughly 40% less per ARR dollar.

Then comes the second, harsher adjustment: the variance discount. Company A's 80% is stable, it'll be 80% next year. Company B's 50% might be 60% if model prices fall and the company keeps the savings, or 35% if a competitor passes the savings to customers and B has to match. Investors price uncertainty by widening the discount rate or shaving the multiple again. A revenue stream whose margin could swing 25 points in either direction is, by definition, lower quality than one that won't.

Stack those two adjustments and you can easily get a 50-60% haircut versus the headline SaaS comp, even though revenue, growth, and retention all look fine on the dashboard. That's the part founders find maddening. Nothing in the metrics they obsess over went wrong. The haircut lives entirely in the COGS line and its volatility.

Who Gets Hit Hardest

Not every agent company eats the same haircut. The exposure clusters into a few recognizable profiles.

Per-Outcome Pricers With Thin Buffers

Per-outcome and per-task pricing is the celebrated GaaS innovation, you charge for a resolved support ticket, a booked meeting, a reconciled invoice, not for seats. It aligns price with value beautifully. It also creates a vicious squeeze when your price is fixed by contract but your cost per outcome is set by a model provider you don't control.

If you sold "$2 per resolved ticket" on the assumption that inference costs $0.40, you've got a comfortable buffer. But if a model upgrade tempts you into multi-agent reasoning loops that quintuple token consumption per ticket, and quality-conscious founders constantly face this temptation, your cost per outcome can drift toward your price faster than you'd think. Investors now model this directly. They want to see the cost-per-outcome trend line, not just the price-per-outcome.

Single-Model Dependents

A company that can only run on one provider's frontier model has handed its margin destiny to a vendor. If that vendor raises prices, deprecates the model, or throttles your rate limits during a capacity crunch, you have no recourse. This is concentration risk dressed up as a technical dependency, and it's why how VCs underwrite GaaS bets differently from SaaS increasingly includes a model-portability section in diligence. Companies with a tested abstraction layer, able to route to a cheaper or alternative model without a quality cliff, defend their margins, and their multiples, far better.

The "We'll Make It Up on Volume" Crowd

Some agent businesses run negative or near-zero gross margins deliberately, betting that scale and the cost curve will bail them out. Sometimes that's a defensible land-grab. More often it's a bet that the model providers will subsidize their growth indefinitely, and the providers are themselves under pressure to improve their own economics. When an acquirer sees a company structurally dependent on its supplier not optimizing for profit, that's not a moat. That's a countdown.

What Investors Actually Diligence Now

The diligence conversation has matured fast. A year ago, "what's your gross margin?" was the whole question. Now it's a battery of follow-ups designed to find the margin's true floor:

• Cost per outcome over time, not just margin today. A flat 50% margin that was achieved by quietly degrading model quality is worse than a 45% margin holding steady with the best model. Trend and method both matter.
• Contractual exposure. Which customer contracts have price-step-downs tied to your costs? Which have minimum-volume commitments that protect you? Investors map the contract book against model-price scenarios.
• Model portability proof, not promises. Can you actually fail over to a second model? Many decks claim it; few have run the fire drill in production. The difference shows up in the multiple.
• The pass-through clause. Do customer contracts let you raise prices if model costs rise? Almost nobody had these in 2023. The savvy operators added them, and it's becoming a quiet diligence checkbox, the same way enterprise SaaS contracts normalized indexation clauses.
• Burn sensitivity to a price shock. If your primary model's price doubled tomorrow, how many months of runway does that erase? This ties directly into the burn-rate problem that makes agents structurally expensive to run, and it's a question late-stage investors now ask explicitly.

McKinsey's broader analysis of how generative-AI economics flow through to enterprise margins is increasingly cited in these conversations, because it frames model cost as a variable input that behaves more like cloud compute or even raw materials than like the fixed R&D cost SaaS investors are used to.

How Founders Defend the Multiple

The good news for operators: the haircut is defensible, and the defenses are concrete rather than narrative.

Start with the contract structure. A pass-through clause, the right to adjust price if underlying model costs move materially, converts your most volatile input into a managed one. It's not glamorous, but it does more for your valuation than another 20 points of growth, because it directly addresses the variance discount.

Build and demonstrate model portability. Not a slide claiming you're "model-agnostic," but logs showing you ran the same workload across two providers last quarter with measured quality parity and a real cost delta. That evidence collapses the single-model concentration discount.

Move work off the frontier model where you can. The companies with the healthiest, most defensible margins are the ones doing aggressive routing, cheap or fine-tuned models for the easy 80% of tasks, the expensive frontier model reserved for the genuinely hard 20%. This is the operational discipline that separates a 65% gross margin from a 40% one, and it's a capability you can show, not just assert. It also connects to the broader capital-efficiency comeback among lean agent startups: the teams that treat inference as a managed cost center, not an afterthought, are exactly the ones raising on better terms.

Finally, instrument cost per outcome as a first-class metric and put its trend line in the deck. Investors discount what they can't see. A founder who walks in with a declining cost-per-outcome curve, a pass-through clause, and a portability fire-drill log has pre-empted most of the haircut before it's applied.

The Counterintuitive Case: When Cheaper Models Help Valuation

Here's the twist that the doom framing misses. Falling model costs are genuinely good for some agent companies, the ones positioned to keep the savings rather than compete them away.

If your moat is workflow depth, proprietary data, or a system of record you've become embedded in, not raw model access, then a price drop is pure margin expansion. Your customers aren't leaving over a model price they never saw on the invoice. Your competitors' cost falling doesn't matter, because they can't replicate your integration overnight. In that world, the cost curve is exactly the tailwind the decks promised, and the valuation should reflect expanding margins on a sticky base.

This is why the same news, "frontier model prices dropped 70%", can be a tailwind for one agent company and a margin threat for another. The differentiator isn't the model. It's whether the company built something the model can't easily commoditize. Investors who understand this don't haircut every agent business reflexively; they haircut the ones whose only buffer was the COGS line, and they pay up for the ones whose value lives above it.

Insights Most People Overlook

1. The haircut is really a bet against the founder's own pricing power. When an investor discounts your multiple over model-cost exposure, they're implicitly saying you won't be able to hold price when costs rise or competitors cut. The most effective rebuttal isn't a cost projection, it's evidence of pricing power: low churn after a price increase, customers who've never asked about your model. Pricing power neutralizes the haircut more cleanly than any margin forecast.

2. Cheaper models can compress your margin even when your absolute costs fall. The non-obvious mechanism: a price drop lowers the floor a competitor needs to undercut you. Your COGS dropping 50% is irrelevant if a rival's COGS also dropped 50% and they choose to pass it through. Margin is set by the competitive equilibrium, not by your cost sheet. Many founders model their own cost curve and forget the curve is industry-wide.

3. The deprecation risk is underpriced versus the price-hike risk. Everyone worries about a model provider raising prices. The sneakier threat is a provider deprecating the specific cheap model your unit economics depend on, forcing you onto a pricier tier. A price hike is a number you can renegotiate around; a deprecation is a forced migration with quality re-validation. Diligence is only now starting to ask about model-lifecycle exposure, not just price.

4. A "good" margin built on a degraded model is a hidden liability. Some companies protect margin by quietly routing to cheaper, weaker models, preserving the gross-margin number while eroding output quality. This shows up later as churn, not as a COGS problem, which makes it doubly dangerous in diligence. A sophisticated acquirer reads margin and quality together; a flat margin achieved by quality erosion is a worse signal than an honestly thin one.

5. The pass-through clause is becoming the GaaS equivalent of the SaaS auto-renewal. A boring contractual term that, in aggregate, reprices a whole category. The first wave of agent contracts had no cost-adjustment mechanism, leaving margin fully exposed. The operators adding pass-through and indexation language now are quietly building the most haircut-resistant revenue in the category, and it costs nothing but legal foresight.

References

• Bessemer Venture Partners, State of the Cloud 2024
• McKinsey, The Economic Potential of Generative AI: The Next Productivity Frontier