AI定价双曲线：1%用户占75%成本，用限速与分级定价止损

Ulrik Lehrskov-Schmidt has worked on pricing full time for about ten years. He’s the founder and CEO of Willingness to Pay, his team has run hundreds of pricing projects, and he’s written several books on B2B pricing. He put off writing one on AI pricing because he didn’t yet have a model that explained what he was seeing in client work: agent pricing, MCP pricing, and gross margins dropping from 85-90% down to 55%.

He says he got there in April. He calls it the double hockey stick, and he walked through it with Amelia at SaaStr, using Higgsfield, Figma, and Clay as examples.

One Customer Can Be 30% of an Entire Base

Every product delivers different value to different customers. A bottle of water is worth nothing when you toss it at airport security and a lot when you’re lost in the desert. The same security software can be optional for one company and critical for another.

The spread is wider than most founders assume. It follows a power law, so one customer can get 1,000x the value another gets.

Willingness to Pay maps usage and value distributions on every project. A typical client at a couple hundred million in ARR has 10,000 to 20,000 customers, and a single one of them can account for 30% of all users or 30% of all infrastructure. The top 20% of customers get 80% of the value, the top 20% of that group get 80% of their share, and so on.

With flat pricing at $5K, $50K, or $500K, you miss the customer who would pay $5M. Ulrik’s team is working with a company in France right now on price increases of 30,000%, about 300x, for some customers who sit at the top of a steep value curve and pay the same flat price as everyone else.

Flat Costs Hid the Problem in Pre-AI B2B

Traditional B2B companies could leave that value uncaptured because every customer cost about the same to serve. The missed revenue was an opportunity cost that never appeared in the P&L, so there was no pressure to go after it.

AI added a second curve.

1% of Users, 50-75% of AI Costs

AI costs follow a power law too. Across a couple dozen high-growth AI pricing projects, Ulrik’s team regularly sees:

• About 50% of users with zero or near-zero AI consumption
• 1% of users driving 50% to 75% of total AI costs
• In extreme cases, 0.1% of users driving 75% or more

In a product with 1 million users, that last case means 1,000 people generate three quarters of the inference bill.

One client’s least profitable customer was spending 380x more in tokens than it paid in revenue. That client had flat pricing and spiky costs, and it was losing money.

A 380x Customer Next to a $20K-on-$500K Customer

A common reaction is to treat AI cost as something to pass through to customers. Ulrik’s point is that cost and value are independent. A product can turn $1 of tokens into $0.10 of customer value, and no pricing change will fix that. The 380x customer was in a category where the value wasn’t there to charge for. A price increase would have produced immediate churn.

It also goes the other way. Another client with roughly $500K ACVs was worried about AI spend. Their largest enterprise accounts were running about $20K a year in AI costs. Ulrik told them to keep doing exactly that: $20K of AI cost supporting $500K of ACV is a good business.

The ratio between the two curves also changes over time. Model capabilities improve, model prices move in both directions, and customers find new ways to use the product. Ulrik calls this “the sticks dance.”

His summary of the problem has five parts:

• There are now two hockey sticks, and the cost one shows up on the bill.
• Costs arrive automatically. Capturing value requires a pricing model built to do it.
• High cost doesn’t imply high willingness to pay.
• The relationship between cost and value keeps shifting.
• Very few companies have worked out how to handle this yet.

Four Cost-Value Combinations

If each curve is either spiky or flat, there are four combinations:

Netflix has flat value and flat, scalable costs. If it started generating custom video for each subscriber, that would change.

Spiky value calls for better pricing design: understanding customer outcomes, choosing the metric that tracks them, and pricing to that metric. Spiky cost calls for guardrails such as fair usage limits that throttle consumption, whether or not you charge more for it.

Higgsfield’s $19, $49, and $99 Plans Include Unlimited AI

A few days before the session, Higgsfield’s pricing page showed tiers at $19, $49, and $99, priced for prosumers. Several features on those plans are listed as unlimited.

Under that structure, Higgsfield collects a fixed fee and gives users uncapped access to something that costs it money on every use. Revenue per user has a ceiling and cost per user doesn’t, which is how a company ends up with a 380x customer.

Most Higgsfield users get similar value from the product, with a handful of heavy users at the top. Costs are much more concentrated. The goal is to charge a broad base flat prices while controlling what the heaviest users cost. Ulrik didn’t know whether Higgsfield has fair usage limits today, and said it should if it doesn’t.

Kill Points, Brakes, and Moving Them Gradually

Ulrik laid out the process his team uses:

Segment users. New users might get more headroom in their first month so they activate before they hit limits. Fraudulent users get handled separately. Everyone else is a third group.

Set a kill point for each segment. This is the usage level at which the product shuts off, for example a million actions per second.

Set one or more brakes. A brake reduces utility at a threshold: halve the speed after 1,000 edits a minute, cap concurrent sessions, or add a short wait before the next job starts.

Decide how users hear about it. An email, an in-product message, or a notice through the MCP.

Move the brake in slowly. Ulrik calls this boiling the frog. The first brake sits at the kill point and might touch 0.01% of users. Then it moves to 0.1%. Users who hit it do one of four things: keep going, use less, churn, or upgrade. Track the mix. If it’s acceptable, keep moving the brake. At around 4% of users affected, a company might have addressed 60-80% of its cost problem. Then add a second, stronger brake at the kill point and move it in behind the first until margins reach the target.

Ulrik’s team has worked with companies that went from profitable to unprofitable after adding AI and recovered up to 22 points of gross margin within weeks, using throttles that affected a few percent of users and no new pricing.

Claude’s usage limits work the same way.

Figma: 92% to 82% Gross Margins, Faster Growth Past $1B ARR

Figma faced obvious exposure to AI in design. It chose to give AI to everyone, but built fair usage limits before launch. After launch, usage was heavy. Figma then turned on credit counting, then enforcement, then monetization, in that order.

By Ulrik’s account, gross margins fell from 92% to 82% over that stretch, so the AI costs were real. Revenue grew faster than those costs, and the revenue distribution became more concentrated at the top.

Ulrik’s view is that Figma always had a value curve it wasn’t capturing. Seat pricing never fit the product well, and pay-per-design or consumption pricing would have fit better. AI costs gave the board and leadership a reason to change the model. After Figma added credits, growth accelerated at more than $1B in ARR.

He sees the same thing at other established B2B companies. They add pricing meant to cover AI costs and end up charging for value customers were already getting for free.

Clay Added Action Credits on Top of Data Credits

Clay has spiky value and costs that are mostly flat.