For twenty years, the scariest number on a startup’s books was payroll. It was the line that kept founders up at night, the one you triple-checked before every hire, the wall everyone eventually hit. Every dollar of it was a decision. A human sat down, weighed it, and signed.
Last month, a company spent half a billion dollars on AI, and no human signed anything.
According to a consultant who told Axios about it, one enterprise let its employees use Claude with no caps and no limits, and the bill came back at $500 million. For one month. Nobody approved that. Nobody chose it. It just happened, quietly, while everyone thought they were being productive. And if you’re building a startup right now, that story is not a funny enterprise blooper you scroll past. It’s a preview of how runways are going to die from here on out.
What actually happened
Let me give you the real numbers, because they matter more than the cartoon headline.
The $500 million wasn’t one reckless person. It was thousands of employees with unlimited access, all running long coding sessions and chained agent workflows at the same time, each one quietly burning tokens. No single action looked insane. The total was. The consultant’s point to Axios was blunt: this is systemic, not a freak accident.
And it isn’t isolated. Microsoft sharply cut its internal Claude Code licenses after costs climbed, with some engineers reportedly generating between $500 and $2,000 a month each, and pushed people toward cheaper tooling with tighter controls. Uber, per the same reporting, burned through its entire 2026 AI budget by April. Its COO admitted the costs were getting hard to justify. These are not scrappy startups with a junior dev who left a loop running. These are some of the most operationally disciplined companies on Earth, and the meter still got away from them.
Here’s the part that tells you it’s a real shift and not just a few bad weeks. On June 8, a startup called PointFive raised a $60 million Series B led by Accel for what it calls an “AI Efficiency OS,” software whose entire job is to watch your cloud and GPU spend and stop it from quietly eating you. When VCs write a check that size for a product that does nothing but tame AI bills, that’s the market telling you the bills have become a category of pain worth a company. The fire is real. Someone just raised money selling extinguishers.
Why your old burn math is broken
Payroll was terrifying, but it was honest. It was fixed, it was slow, and it could not surprise you. You knew on the first of the month exactly what was leaving your account. To spend more, a person had to do something: make an offer, sign a contract, hire a body. The cost had a face and a decision behind it.
Tokens have neither. Your AI bill is metered, it’s variable, and it can move while you’re asleep. An agent that hits an error and retries, then retries the retry, then decides to re-read your entire codebase to figure out why, is a spending decision being made hundreds of times a minute by something that doesn’t know what money is. You didn’t hire it. You can’t fire it. And it doesn’t go home at six.
The trap most founders are walking into is a mental model problem. We learned to think about software cost as a SaaS seat. Flat, predictable, one price per person per month. You could put it in a spreadsheet and forget it. But agentic AI doesn’t bill like a seat. It bills like electricity, except the appliance can turn itself on, run all night, and call friends. This is the same hidden dynamic playing out at the top of the market, where I wrote about how Big Tech now spends roughly five times its payroll on AI infrastructure, which means the layoffs were never really about saving money. The biggest companies already crossed the line where compute, not people, is the dominant cost. You’re going to cross it too, just with smaller, scarier numbers.
And here’s the cruel twist. The price of the underlying intelligence keeps falling. We watched DeepSeek cut the cost of frontier-grade models by something like ninety percent, and everyone assumed cheaper tokens meant smaller bills. The opposite happened. When something gets ten times cheaper, you don’t use the same amount and save money. You use a hundred times more, because now you can, and your bill goes up. Cheaper fuel didn’t make the car cost less to run. It just made you drive everywhere.
The part I’m not sure about
Let me be honest, because I could be overcorrecting. You might read this and think, Liran, I’m pre-seed, I spend forty dollars a month on API calls, this is an enterprise problem and you’re scaring me about a fire in a building I don’t live in.
Fair. Maybe. If you’re a solo founder hand-typing prompts into a chat window, your blast radius is small and this is a someday problem.
But the moment you ship anything agentic, the moment you wire a model into a loop or a background job or a feature that calls it on every user keystroke, the math stops being linear and you don’t get a warning. The enterprise hit $500 million because scale multiplied a small per-action cost into a catastrophe. A startup hits its version of that the first time a bug ships a function that calls a frontier model inside a while-loop over the weekend. Same mechanic. Three orders of magnitude smaller, and just as capable of eating the runway you raised to last you eighteen months. I’m not sure it’ll get you. I’m sure it can.
What to actually do about it
Stop treating AI like a perk and start treating it like cloud infrastructure, because that’s what it is. Nobody runs production servers without billing alerts and budget caps, and nobody should be running agents without them either. This is unglamorous plumbing, and it is the difference between a long life and a short one.
Set a hard spending cap before you set anything else. Most providers let you cap usage and get alerted at thresholds. Do it on day one, not after the invoice teaches you. A cap that pauses the party is annoying. An invoice that ends it is worse.
Know your cost per task, not just your cost per month. The monthly number hides everything. What you actually need to know is what one run of your agent costs, what one user session costs, what one feature costs per call. If you can’t answer that, you’re flying a plane with no fuel gauge, and the fact that there’s still gas in the tank today tells you nothing about Thursday.
Match the model to the job. You do not need the most expensive frontier model to format a date or tag an email. A huge share of spend is people firing a Ferrari at problems a bicycle would solve, because the Ferrari was right there and felt smarter. Cheap models for cheap jobs, expensive ones only where they earn it.
And while you’re auditing what you spend, audit what you get. There’s a brutal companion fact to all this, which is that around eighty percent of companies get nothing measurable back from their AI spend. So a lot of these bills aren’t even buying outcomes. They’re buying motion. Before you optimize the cost of a thing, make sure the thing is doing something. The cheapest token is the one you never needed to spend.
The meter is the new co-founder you didn’t pick
Here’s what’s actually new, and why this isn’t just another “watch your spending” lecture you’ve heard since you were broke.
Every previous threat to your runway had a human in the loop. You could see it coming, argue with it, slow it down. A bad hire, a bloated tool stack, an over-eager ad budget. They all moved at human speed. The runaway AI bill is the first one that moves at machine speed, makes its own decisions, and feels productive the entire time it’s bleeding you. It looks exactly like progress right up until you open the invoice.
So put a gauge on it. Cap it, watch it, and respect that you’ve added something to your company that spends money without asking. The $500 million company didn’t get greedy. It just forgot that it had quietly hired the most expensive, tireless, unsupervised employee in history, and never gave it a budget. Don’t make the small-startup version of that mistake. Yours won’t make the news. It’ll just make your runway disappear.
