Why I Won't Sell "Unlimited AI"

Traditional SaaS has a comforting property: once it's built, one more user costs almost nothing. AI products throw that out. Every interaction has a real marginal cost. A database query costs a fraction of a cent; a capable-model API call runs $0.03–$0.10 per thousand tokens, and a full conversation with reasoning and context can cost anywhere from $0.50 to $5.00 in compute — per session.

That single fact reshapes how you're allowed to price.

"Unlimited" is one of three things

Any product promising unlimited AI access on a flat monthly fee is doing one of the following:

1. Quietly serving you weaker, cheaper models than the marketing implied.
2. Imposing rate limits after the fact, once your usage becomes inconvenient.
3. Burning investor cash to paper over the gap, until it can't.

There is no fourth option where the math just works. The clearest proof came in mid-2025, when Anthropic introduced weekly rate limits on its $100–$200/month Max plans after disclosing that a tiny minority of users had consumed "tens of thousands of dollars" in model usage on a $200 plan. If the company training the model can't offer sustainable unlimited, neither can an app built on top of it.

Why AI costs scale faster than AI revenue

Four forces push inference cost up faster than subscription revenue comes in:

• Power-user abuse. A small slice of users consume exponentially more than the average.
• Agentic usage. An AI agent running autonomously can burn orders of magnitude more compute than a human clicking through a session.
• Context costs. Long conversations and big documents or code repos passed as context multiply the token bill.
• Better models cost more. "Unlimited access to all models" is a liability that grows every time a smarter, pricier model ships.

Rate limits, in that light, aren't hostility toward users. They're honesty about physics — plus fairness (one whale shouldn't degrade everyone else's service) and abuse prevention.

How to price an AI feature without lying

The architecture I'd defend for any AI product in 2026:

1. Never price "unlimited" unless you have a concrete cost model that survives your top 1% of users consuming 100x the average.
2. Use token- or credit-based pricing for anything API-shaped — it aligns cost and revenue almost perfectly.
3. Define a value metric users actually understand — documents analyzed, images generated, summaries produced — not raw token counts. People can reason about "50 reports a month." Nobody budgets in tokens.
4. Build in overages instead of hard walls. Let people buy more credits when they hit the limit; a hard block at the value moment turns a paying fan into a churned one.
5. Tier by model capability, not just volume. "Fast model vs. smart model" is more intuitive than "1M vs. 5M tokens," and it lets heavy-but-cheap users coexist with light-but-demanding ones.

This is the same lesson as the rest of my AI work

I've argued before that AI builders should chase primitives, not bigger models — that reliability and tight constraints are the product, not raw capability. Pricing is the same argument wearing a different hat. The founder who promises "unlimited" is really promising to either degrade quality or go broke, and letting the customer discover which one later. A usage metric the user can see and understand is the honest version of that same product.

The lesson: in AI, "unlimited" is a marketing word for an accounting problem. Price the value, show people what they've consumed, cap the runaway cases — and you'll still be in business when the ones who promised infinity aren't.

Part 6 of From Free to Premium. Next: the trust cost of subscription fatigue and dark patterns.