Most AI companies today are making a hidden bet: that their primary cost won't move against them. If it does — if token prices double — a surprising number of "profitable" AI businesses stop working overnight.
I saw this play out recently over lunch with a former Travelers executive and a retired New York City attorney who had joined an AI legal startup. Their pitch: revolutionize legal services through flat billing and subscription pricing. No more billable hours. Affordable, predictable legal help for the people who've always been priced out of it.
That's a genuinely important problem. Real people — small business owners, families navigating disputes, employees without HR departments — are making consequential decisions without legal guidance because they can't afford it.
I asked one question.
What happens to your financial model when the cost of tokens goes up?
The table got quiet. That question is the whole ballgame for every AI business built on flat or subscription pricing. And most of them haven't answered it yet.
The Spirit Airlines Warning
Spirit Airlines shut down on May 2, 2026. Their restructuring plan assumed jet fuel at roughly $2.24 a gallon. By late April, prices had climbed to $4.51 — nearly double — driven by oil market disruptions. Spirit's CEO cited the sudden and sustained rise in fuel prices as the reason the airline had no alternative but to cease operations.
Spirit was already struggling before fuel prices doubled. They had filed for bankruptcy twice in less than two years. The doubled fuel cost didn't create a broken business model. It made an already-thin model impossible to survive.
Ultra-low-cost models leave no room for input cost volatility. When your entire value proposition is built on razor-thin margins — cheap fares, cheap legal fees, cheap anything — a doubling of your primary input cost doesn't hurt you. It ends you.
Now apply that to subscription AI.
Your Margin Lives on Someone Else's Pricing Page
Every AI startup selling flat-fee or subscription services is building on someone else's fuel. Anthropic, OpenAI, Google — they set the price per token, not you. Your business plan assumes token costs stay in a certain range. Maybe they go down — token prices have been falling. But you don't control the direction.
When you're chatting with AI casually, you pay $20 a month and never think about it. Under the hood, your words get broken into small units called tokens — roughly three-quarters of a word each. Every token you send costs money. Every token the AI sends back costs more, typically four to five times more. When a company is running a business on top of AI — processing thousands of client documents, answering thousands of legal questions — the token bill compounds fast.
NVIDIA framed the real risk well in their analysis on AI infrastructure economics. Most companies focus on the wrong metric: the cost of the GPU, the cost of the compute. What actually determines whether an AI business is viable is the cost per token delivered — the all-in cost to produce each unit of intelligence your business runs on. That's the number that determines whether you can scale profitably. Everything else is a proxy.
The billable hour in legal services exists for a reason. It isn't just greed. It's a hedge against variable cost. Flat-fee AI legal services can't do that. The predictable price is the product. And that's the risk.
What's Actually Happening Inside Companies Right Now
Gergely Orosz, who writes the most widely read software engineering newsletter on Substack, recently surveyed 15 companies about their AI token spend. Token usage has increased roughly 10x in the last six months at many companies, with no signs of slowing.
At a late-stage fintech with around 5,000 employees, some developers are spending $500 a day on AI coding tools alone. At a seed-stage AI infrastructure startup with just 15 people, spend went from $200 per developer per month to $3,000 in six months — a 15x increase. A VP of AI at a mid-sized finance company described what happens when you set a $100-per-user monthly limit: people exhaust it in three to five working days.
Then there's "tokenmaxxing" — a phenomenon where developers deliberately run agents at maximum cost to inflate their usage statistics, because companies have introduced AI usage leaderboards in performance reviews. The incentive to use AI aggressively has become decoupled from the incentive to use it efficiently. People are burning tokens to look good on a dashboard.
Agents Are 1,000 Times More Expensive Than You Think
Most of the AI hype right now isn't about chatbots. It's about AI agents — systems that don't just answer a question once, but take sequences of actions, loop back, and work through complex multi-step tasks autonomously. Every law firm AI demo, every "AI paralegal" pitch, every claims-processing automation — these are agent-based systems.
A recent Stanford Digital Economy Lab study found that agentic tasks consume roughly 1,000 times more tokens than simple question-and-answer interactions. When an agent reads a task and takes an action, it has to re-read the original task plus that action before the next step — then re-read all of that plus the next action. Every step builds on all previous steps. The context window keeps growing. The token bill compounds with every move.
The same researchers found that token costs in agentic tasks varied by up to 30 times for the exact same task. Think about that in the context of a flat-billing legal services startup. You've promised a fixed monthly fee to process contracts, answer compliance questions, handle routine legal work. Your AI agent's cost variance is 30x. You find out the actual cost after the work is done.
That's not a business model. That's a lottery ticket.
The Full Circle Nobody Saw Coming
The industry's response to runaway token costs isn't to raise subscription prices. It's to hire engineers. Not to replace AI. To optimize it.
Prompt compression rewrites instructions to convey the same meaning in fewer words — reducing token usage 30 to 50 percent. Model routing sends simple queries to cheaper models and reserves powerful ones for complex tasks; one company reduced costs 30 percent overnight just by changing the default model. Semantic caching reuses answers when thousands of users ask essentially the same question instead of paying the model thousands of times. RAG retrieves only relevant context instead of feeding massive amounts into every call. And self-hosted models give you ownership of your economics entirely — when you run your own infrastructure, someone else's pricing page is no longer your problem.
We spent three years telling developers that AI was coming for their jobs. Now we're hiring them back to make AI affordable. Full circle.
What Winning Actually Looks Like
The businesses that get this right will have done one thing the others didn't: they'll have modeled their token exposure before they signed the contracts, not after. They'll know their cost floor, their cost ceiling, and what happens when their primary input cost moves against them. They'll have built governance — measurement, routing, optimization — before the bill arrived.
They won't be Spirit Airlines discovering that $4.51 fuel was incompatible with $29 fares when it was already too late.
Mark Cuban put the right frame on it: "Know how your company will make money and how you will actually make sales." Not just how you'll make the sale. How you'll make the money. Most AI businesses right now know how to make the sale.
What do your token costs look like if they double — and have you stress-tested your pricing model against that scenario?
If you're a managing partner at a law firm, healthcare system, or financial services company evaluating AI subscriptions, that's the question your vendor should be able to answer before you sign — not six months from now when the invoice arrives.