1% of Users Drive 75% of AI Costs — and Most SaaS Pricing Can't Handle It

One percent of users generate 50% to 75% of total AI costs in typical AI products, and in extreme cases 0.1% drive three quarters of the inference bill. That's the core data point pricing consultant Ulrik Lehrskov-Schmidt brought to SaaStr, drawn from a couple dozen high-growth AI pricing projects run by his firm, Willingness to Pay.

Lehrskov-Schmidt, who has worked on pricing for about ten years and founded Willingness to Pay, calls the pattern the "double hockey stick." Two power-law curves now run through every AI product: a value curve that always existed, and a new cost curve that shows up directly on the bill. AI-era gross margins at his clients have dropped from 85-90% to 55%.

The value curve is steep. A typical client at a couple hundred million in ARR has 10,000 to 20,000 customers, and a single one can account for 30% of all users or infrastructure. His team is currently working with a French company on price increases of 30,000% — roughly 300x — for customers sitting at the top of that curve who pay the same flat price as everyone else.

What does the cost curve look like in practice?

Across his projects, the usage distribution is consistent:

• About 50% of users show zero or near-zero AI consumption
• 1% of users drive 50-75% of total AI costs
• In extreme cases, 0.1% of users drive 75% or more — 1,000 people out of a million

His most striking case: one client's least profitable customer spent 380x more in tokens than it paid in revenue. Cost and value are independent curves, he argued — a product can turn $1 of tokens into $0.10 of customer value, and no pricing change fixes that. The 380x customer would have churned at any price increase.

The reverse also holds. Another client with roughly $500K ACVs worried about AI spend; its largest accounts ran about $20K a year in AI costs. His advice: keep doing exactly that. A $20K bill on a $500K account is a healthy ratio.

Which combination are you in?

Lehrskov-Schmidt maps every AI business onto four combinations of flat or spiky cost and value curves:

• Flat value, flat cost (Netflix): everyone pays ~$10/month; no problem until costs spike
• Spiky value, flat cost (Clay): classic B2B; needs better value pricing
• Spiky value, spiky cost (Figma): needs both value pricing and cost guardrails
• Flat value, spiky cost (Higgsfield): needs cost guardrails, urgently

Higgsfield sells prosumer tiers at $19, $49, and $99 with several features listed as unlimited. Revenue per user has a ceiling; cost per user doesn't. He didn't know whether Higgsfield has fair usage limits, and said it should if it doesn't.

How do you actually fix it?

His team's method: segment users (new, fraudulent, everyone else), set a kill point where the product shuts off, then set brakes that reduce utility at thresholds — halve speed after 1,000 edits a minute, cap concurrent sessions. Then move the brake in slowly, what he calls boiling the frog. The first brake might touch 0.01% of users, then 0.1%. Users who hit it keep going, use less, churn, or upgrade — track the mix. At around 4% of users affected, a company may have addressed 60-80% of its cost problem.

The results are measurable. Companies that went from profitable to unprofitable after adding AI recovered up to 22 points of gross margin within weeks using throttles that affected a few percent of users — with no new pricing.

What did Figma and Clay do?

Figma gave AI to everyone but built fair usage limits before launch, then turned on credit counting, enforcement, and monetization in that order. Gross margins fell from 92% to 82%, but revenue grew faster than costs. After adding credits, growth accelerated past $1B ARR. Clay added action credits as a second pool on top of its data credits; its heaviest users now pay several times what they paid before.

On agents, he was blunt: per-agent seats "worked for about five minutes." Agent output has no natural range the way human output does, so price on what the agents do, not how many exist.

His book, The AI Pricing Roadmap, comes out in about five weeks, and he'll go deeper on agent pricing at SaaStr AI next May.