The Hidden Token Tax: 5 Honest AI Margin Checks

A founder showed me her pricing page last month. Twenty dollars a month, unlimited AI chat, proud of it.

Then I asked one question. "What does one heavy user cost you to serve?"

She didn't know. Almost nobody knows. And that gap is where the money quietly leaks out.

Here's the thing nobody told her when she bolted an AI feature onto her app: the old software rules don't apply anymore. For twenty years, software was the best business on earth because once you built it, serving the next customer cost almost nothing. For two decades, software was a spectacular business because its marginal cost per user approached zero. Once the code was written, serving the next customer cost almost nothing. Gross margins of 75 to 85 percent followed automatically.

AI broke that. Quietly. Permanently.

The take: every AI request is a meter, and you're paying it

Add an AI feature, and every single use costs you real money. Not "more servers eventually" money. Right-now, per-click money.

AI broke that. Every time a user interacts with an AI feature, every query, every inference, every agent action, real compute gets consumed. Token costs, GPU time, and third-party API fees stack up with every user interaction. The software is no longer free to operate at scale; it has a cost that grows with usage.

That has a name now. People call it the token tax. And the numbers are not small.

ICONIQ Capital's 2026 data shows AI-native product gross margins averaging 52 percent, compared with 75 to 85 percent for mature software-as-a-service businesses, a gap of 23 to 33 percentage points driven almost entirely by inference spend.

Put it another way. For every $1 million in AI product revenue a SaaS company books in 2026, roughly $230,000 walks out the door as inference cost before a single engineer, AE, or marketer gets paid. That number comes from ICONIQ Growth's 2026 State of AI Bi-Annual Snapshot.

So if you're a non-technical founder who just added a chatbot or an "AI summary" button, this is for you. Not because AI is bad. It's great. But because the price you charge has to survive the bill you'll get.

Why flat pricing is the trap most founders fall into

The natural move is to copy the SaaS playbook you've seen everywhere. Pick a number. Twenty bucks. Unlimited. Simple.

That worked when serving a user cost nothing. It doesn't work when a single power user can burn through more compute than ten casual ones combined.

The proof is sitting in plain sight. Even GitHub, sitting on Microsoft's own data centers, gave up on flat AI pricing. GitHub announced that Copilot would move to usage-based billing as of June 1, 2026. Why? Because the variance in usage is enormous and the cost gap between users is widening.

And the smartest, fastest-growing AI company of the moment learned it the hard way too. For most of 2024 and into 2025, Cursor's cost of goods sold was dominated by inference fees paid to Anthropic and OpenAI. Every heavy user represented a direct pass-through cost that exceeded what Cursor was charging. TechCrunch reported in April 2026 that the company had operated at negative gross margins until recently, meaning it cost more to run the product than the company could collect.

Negative margins. On the fastest-growing software product in history. Let that land.

If your flat price doesn't cover your heaviest 10% of users, you are subsidizing them. The more they love your product, the more money you lose. That's a horrible trap to be in, and most founders don't see it until the bill arrives.

The number that decides everything: your real cost per use

Here's where it gets sneaky. The cost you can see is not the cost you actually pay.

A typical user interaction costs 200 to 300 tokens on the surface. The real number is often ten times that once you account for the full system. This is the Token Iceberg, and most operators only price for the visible tip.

Why ten times? Because a single thing a user clicks often fires off a whole chain behind the scenes. The model rereads context. It retries. An "agent" takes several steps to do one job. For agentic products, where a single request fans out into a chain of reasoning steps, tool calls, and retries, the cost of one session can multiply many times over.

So the rule is brutal but simple. Most AI operators underestimate their real token cost by two to five times and discover it after launch.

You do not want to discover yours after launch. You want to know it before you set a price.

The good news: this is fixable, and the world is moving your way

This is not a doom story. AI margins are an engineering and pricing problem, not a death sentence. The teams who treat it seriously are pulling away from the teams who don't.

Three things are working right now, and you don't have to be a coder to ask your team about them.

One. Use a cheap model for easy questions. Most requests are simple. You don't need the most expensive brain for "what's the weather." The default architecture across well-run AI SaaS products is now a tiered router: a small, cheap, fast model handles the 80 percent of queries that are simple, and a frontier model gets called only for the genuinely complex 20 percent. Red Hat's AI infrastructure team has documented enterprise deployments that cut compute costs by 70 percent while holding output quality steady.

Two. Cache the stuff that repeats. If your AI keeps re-reading the same instructions every time, you're paying for the same thing over and over. Both Anthropic and OpenAI now offer roughly 90 percent discounts on cached input tokens.

Three. Price for usage, not a flat seat. This is the big one, and the whole industry just voted on it. In January 2026, Stripe bought Metronome, a company whose entire job is billing for usage. "Metered pricing is the native business model for the AI era," Collison said. When the world's payments giant spends a reported billion dollars to do usage billing properly, that's a signal worth reading.

You don't have to go all-in. Most companies are adopting hybrid models rather than going pure outcome-based. A base subscription covers fixed costs and provides revenue predictability, while the outcome-based component captures upside when the AI delivers measurable results. This blended approach gives vendors a floor while still aligning their incentives with customer value.

If you're early and your AI feature is one piece of a bigger product, plain web app advice still applies. Build the thing first, measure the real cost, then price. If you're scoping that build now, our MVP development approach bakes cost-per-use into the plan instead of leaving it as a surprise.

When this take is wrong

If your AI feature is tiny, this matters less. There are three P&L profiles: AI-Augmented, with a target margin of 80%; AI-Enabled, with target margins of 60% to 79%; and AI-Native, with target margins of 50% to 60%. Most SaaS companies today are AI-Enabled. If AI is a sprinkle on top of a normal product, your margins won't crater. Don't panic and over-engineer.

And don't bet your whole plan on prices falling. They are falling per token, but people use more, so total bills climb. Don't model falling inference costs as guaranteed budget relief. LLMflation is a signal to build usage governance, not a signal to skip it. Without governance, efficiency gains get absorbed by consumption growth.

The artifact: 5 questions to ask before you price your AI feature

Copy these into a doc and walk through them with whoever builds your product. You don't need to write code to ask them.

1.     What does one average use actually cost us, fully loaded? Not the visible tokens. The whole chain, including retries and context the user never sees.

2.     What does our heaviest 10% of users cost us per month? If your flat price doesn't cover them, you're losing money on your biggest fans.

3.     Are we using a cheap model for easy requests and saving the expensive one for hard ones? If the answer is "we use the best model for everything," you're likely overpaying by half or more.

4.     Are we caching the parts of our prompt that repeat? A one-week job here can cut effective cost by a lot.

5.     Does our pricing pass usage back to the customer, even partly? A base fee plus a usage component protects you when someone falls in love with your product and uses it all day.

If you can't answer questions one and two with a real number, stop. That's the work to do first.

We've shipped a lot of AI features for founders, and the pattern is always the same: the ones who model cost-per-use before launch sleep fine, and the ones who guess get a scary bill in month two. If you've built something on Lovable, Bolt, or Cursor and you're about to put a price on it, send it our way. We'll tell you what one real user costs you to serve, and whether your price survives it. Book a 20-minute scoping call and bring your pricing page.

P.S. The deadline that should worry you isn't a platform deadline. It's your second invoice. The token tax doesn't show up on launch day. It shows up when usage climbs, four to eight weeks in, right when you're celebrating growth. Better to know the number now. See how we build production-grade AI features that hold under real load on our AI development page.