The Bill Always Comes Due

The seat-based pricing model struggling as it measured logins instead of value. The challenge is to get to unit economics that work before the VC money cushion runs dry.

Last week I wrote about why the seat-based SaaS pricing model has largely collapsed. The short version: the seat was always a proxy for value, not a measure of it, and AI agents blew that proxy apart by doing real work without a clear user model. If you haven’t read it, start there. This piece picks up where that one left off.

Matt Barrie, CEO of Freelancer, published a remarkable piece last week that goes deeper than anything I’ve seen on the economics underneath the pricing crisis. It’s long and technically dense in places, but worth the read if you want the full picture. What he’s really asking is that even if the industry figures out how to price AI correctly, can the underlying economics ever actually work?

His answer, at least for now, is no. And the reasoning is worth understanding even if you’ve never looked at a margin calculation in your life.

Here’s the simplest version. Traditional software is almost free to copy. Once a company writes the code, serving the next customer costs almost nothing extra. That’s why software companies can keep 70 to 85 cents of every dollar they bring in. AI is different. Every time you ask it a question, somewhere a server burns electricity, processes your request, and generates a response. In otherwords, runtime execution is doing net-new work. That cost is real, it recurs every single time, and it doesn’t get cheaper just because you have more customers. As Barrie puts it, the unit economics of AI are “at their core, the unit economics of compute, not software. And compute has never commanded software margins. It never will.”

He’s right that the numbers are ugly right now. OpenAI spent $8.4 billion on inference alone in 2025 and lost money on its most expensive subscription tier. Sam Altman has said publicly that the $200 per month plan loses money. Anthropic has spent roughly two dollars for every dollar it’s earned since it was founded. These are not rounding errors. They are structural deficits being covered, for now, by venture capital.

Which raises the obvious question: how long does that last?

Probably longer than skeptics expect, and shorter than optimists hope. The cost of running AI is falling fast, far faster than any previous technology. Per-token inference costs have dropped roughly 280 times in 18 months. Anthropic’s gross margins have moved from deeply negative in 2023 to around 50 percent today, with a credible internal projection of 77 percent by 2028. If that trajectory holds, it starts to look like software economics. The companies actually making money right now, like NVIDIA selling the chips that run everything and cloud providers renting the infrastructure, are running at margins that would make any SaaS CFO envious. The problem is concentrated at the model layer, not across the whole ecosystem.

What optimistic takes, which are taking shape miss is that even as the cost per query falls, the number of queries is exploding. Agentic AI workflows use, often, 5x to 20x times more compute per task than a simple chatbot. Every efficiency gain creates demand for more AI, not less. The total bill keeps rising even as the per-unit cost falls. Meanwhile, open-source models are closing the quality gap fast, which puts a ceiling on what anyone can charge, and that ceiling is dropping every year.

So where does that leave the industry? The seat was a bad proxy for value because it measured logins instead of outcomes. Consumption pricing is also an imperfect proxy because it measures compute instead of results. Outcome-based pricing, charging only when something actually gets resolved or completed, is directionally correct, but only about 10 percent of vendors have managed to implement it at scale.

The companies that survive this are the ones building toward that model, not defending what came before.

Barrie frames the whole thing better than I can: the AI industry’s current answer to “who pays for all this?” is venture capitalists and debt markets, paying on the promise that someday the unit economics will fix themselves.

Someday is doing a lot of work in that sentence.