AI Cost Concentration Turns 1% of Users Into 75% of Expenses, Says SaaS Pricing Veteran

Ulrik Lehrskov-Schmidt, founder of Willingness to Pay, coined the “double hockey stick” after observing a stark shift in AI-driven SaaS economics. In a recent SaaStr interview he walked through the pattern with Amelia, citing Figma, Clay and Higgsfield as live examples of firms grappling with the new cost curve The Double Hockey Stick – SaaStr.

One Percent, Up to 75% of AI Spend

Lehrskov-Schmidt’s data, drawn from “a couple dozen high-growth AI pricing projects,” shows a power-law distribution of AI consumption. Roughly half of all users consume little or nothing, while the top 1% of users account for 50%-75% of total AI costs; in extreme cases the top 0.1% generate more than three-quarters of the inference bill. In a product with one million users, that extreme case translates to 1,000 heavy users shouldering the bulk of the expense. One client discovered a “least profitable customer” that spent 380 times more on token usage than it contributed in revenue, a loss hidden by flat pricing.

Flat Pricing Hides a Dual Curve Problem

Traditional B2B SaaS could ignore such variance because serving each customer cost roughly the same. AI introduces a second, independent curve: cost now scales with usage while value may remain flat. The mismatch means that flat fees—whether $5 K, $50 K or $500 K—miss high-willingness-to-pay customers and expose firms to unprofitable heavy users. Lehrskov-Schmidt notes that a French client is negotiating price hikes of up to 30,000 % (about 300 ×) for its top-value accounts, illustrating the scale of the untapped revenue curve.

Four Cost-Value Configurations, Three Real-World Cases

Lehrskov-Schmidt maps the interaction of cost (flat vs. spiky) and value (flat vs. spiky) into four quadrants. The SaaS landscape currently exhibits three of them:

• Spiky value, flat cost – Clay: The data-enrichment platform sees uneven value across accounts but relatively stable AI costs. It introduced “action credits” on top of existing data credits, turning its heaviest users into higher-paying customers without altering the underlying cost structure.

• Spiky value, spiky cost – Figma. Gross margins fell from 92 % to 82 % as AI spend rose, but revenue growth outpaced cost, pushing the company past $1 B ARR.

• Flat value, spiky cost – Higgsfield: Its tiered plans ($19, $49, $99) include “unlimited AI,” creating a ceiling on revenue per user while cost per user can climb indefinitely. Lehrskov-Schmidt suggests implementing fair-usage limits to prevent 380× customers from eroding margins.

The fourth quadrant—flat cost, flat value—is exemplified by Netflix, which currently enjoys scalable costs aligned with a uniform subscription price.

Throttles, Kill Points and Gradual Brakes

Lehrskov-Schmidt outlines a pragmatic mitigation framework. First, segment users (new, fraudulent, and the bulk). Next, define a “kill point”—the usage threshold where the service shuts off. Then add “brakes” that reduce utility incrementally (e.g., halving speed after a certain edit rate). Communication channels (email, in-product notice, MCP) are chosen, and brakes are introduced slowly, akin to “boiling the frog.” In practice, affecting roughly 4 % of users can address 60-80 % of the cost problem, allowing firms to recover up to 22 gross-margin points within weeks, according to Lehrskov-Schmidt’s case work.

Outcomes from the Three Companies

• Figma: After layering usage caps and credit systems, the company’s gross margin dipped but revenue grew faster, leading to accelerated growth beyond $1 B ARR. The shift also concentrated revenue among top-value customers, aligning cost and value curves.

• Clay: By adding action credits, the platform captured more of the existing value curve. Heavy users now pay several times their prior fees, narrowing the gap between AI cost and revenue without a wholesale pricing overhaul.

• Higgsfield: The firm’s unlimited-AI tiers risk a 380× cost-to-revenue ratio. Lehrskov-Schmidt recommends fair-usage limits to protect margins, though he does not confirm whether such limits are already in place.

Our read: The “double hockey stick” frames AI adoption as a structural pricing challenge rather than a marginal cost issue. The evidence—power-law usage, 380× loss cases, and margin swings of up to 22 points—shows that flat pricing models are ill-suited for AI-intensive products. The suggested throttling approach offers a low-friction path to re-balance cost and value, but it relies on careful segmentation and incremental enforcement, which may be operationally complex for fast-growing SaaS firms.