The AI Usage Limit Wars Have Begun

The chilling effect of unlimited AI no longer being unlimited for businesses, and how to solve it

Yesterday, I had coffee with a founder friend of mine who left his successful startup to start a company entirely built with AI.

Headline of that coffee meeting: He’s got 20 agents running, and he has named them. With names like “Steve.”

OK, I’d name them too.

But even more damning is the fact that what he’s been doing with those agents has been working. Until now. He called me because a big old crack just appeared in the foundation of his successful AI-built business.

I told him I’d help him with his nightmare if he let me write about it.

Letting AI Build and Run a Business Until It Breaks

My friend had originally reached out to me in late 2023, about six months before he left his former startup, and explained his grand plans in detail. It’s him and two other guys, all former winners in the tech startup game. He’s the business mind, he’s got a sales guru, and a tech resource who is barely an adult, but with some impressive coding cred.

They decided they would build and fund a tech startup, from scratch, using AI to lead the way and do the work.

The business idea and model? AI came up with it. The go-to-market plan? AI came up with it. The MVP and eventual build? AI coded it. Marketing and sales?

You get the idea.

It sure sounds fun.

Look, I don’t want to keep being the guy that rains all over this robot parade. I helped build some of the early versions of generative AI, and I’ve come to the conclusion that I need to accept modern AI like I accept Taylor Swift. She’s inevitable and I guess the music is foundationally OK, so I need to stop making fun of the fandom or my niece will stop talking to me.

Anyway, they launched about a year ago, and about six later, one of my friend’s feature sets stopped working. Not just a busted form, but an entire workflow barfed when his customers tried to use it. He found out about this from his customers, not his “AI team,” and that was concerning. But the tech coder kid jumped in and, three hours later, found this issue and pluckily had his AI bots plug the hole in the code.

“The bug was a silly mistake,” my friend said. “The AI did something it shouldn’t have. When one user got clever and went a little rogue with a use case, the AI added code back into the code base to handle it.”

“That’s not good,” I said.

“I know,” he replied. “But it comes with the territory. It has to learn like any new developer”

“Fair enough.”

A month later, it happened again.

Broken Code Is Not the Real Problem

The next time it happened, it wasn’t bad code that broke the app but a misinterpretation that deleted a few rows from the database. No problem, they had backups, and the data was restored in an hour.

But the preventative fix took a couple days. No data was lost while they were debugging and patching, but the specter of their entire database being wiped out was hanging over their heads. That was too much for my friend.

“I had [coder kid] go through the codebase and refactor and clean up all the technical debt,” he said. “This took a few days but it was worth it. So we made it part of the process, and even used our new AI agents to do most of the work. But only [coder kid] approved the final changes.’

“Good,” I said. “Smart.”

This actually worked out quite well. Fridays became “Black Friday,” when coder kid and his army of agents would sift through the code to make sure everything was on the up-and-up.

The business continued to grow quickly and to the levels that my friend was dreaming about when he first got the idea back in 2023. The codebase got bigger and did a lot more stuff.

Then, about a week ago, insert loud cracking noise.

Coder kid slacked my friend to tell him that, while trying to source and fix a new broken code issue, the AI provider had stopped him out. They had hit usage limits.

“What usage limits?” My friend responded.

“Unlimited” AI is Not Unlimited

But this is still not the real problem.

In mid-July, Anthropic, and other AI providers, got more serious about the limits they were setting on customer usage at all tiers, from the free tier to Unlimited Extra-Super-Cool-Guy-Pro tier. Anthropic set the limits without telling anyone, which caused confusion and anger among their users.

I can speak to Anthropic because it’s a documented case and I use Claude Code. I don’t overindex on AI, for sure, but when I went down a rabbit hole on a new product I was building, I noticed these limits too. For me it was a couple hours, sometimes less than an hour, and I used that time to go get a sandwich.

The lack of communication on the new limits was a problem, but only a one-time problem. The first time it happened to me, I said to myself, “Good thing I wasn’t in the middle of anything crazy.”

Unfortunately, my friend could not say the same thing.

Apparently, “Steve” and them boys needed some “me time.” During the workday.

A broken production code base with nothing you can do about it for five hours in the middle of the day is a waking nightmare.

Here’s the Real Problem

I don’t think there’s anything Anthropic and the other providers can do about it.

See, the real problem is not that Dario and Sam and “them boys” got more greedy — say what you will. It’s that the economics of the processing for all this AI magic are starting to become an unflinching and unavoidable truth.

Read between the lines on the July series of fiascos and you may come to the conclusion that the providers might not be 100 percent sure of how to measure the costs of customer usage properly, but they know for sure they don’t have the machines to handle what they’re seeing. And while they’re building new data centers at an alarming clip, the massive energy requirements would still like a word.

Unlimited is never, ever unlimited.

And so yeah, while they’re all claiming to be setting these limits to “restrict the 5 percent” of users who beat the holy hell out of those servers on the unlimited plan, that does not bode well for my friend, who isn’t even close to the 5 percent until he is, which only happens when things go horribly, horribly wrong. Or horribly right, like in the case of a happy surge in customer acquisition.

Either way, there’s no good mitigation strategy for that. You just take your lumps. Maybe “Steve” can handle all those angry support calls.

Oh yeah. He’s in “me time.”

Here’s the Solution

My friend can dig his business out of this hole right now. It’s just going to cost a lot, because now it’s going to cost what it really costs.

Lesson painfully learned.

We all just went through this with SaaS and the freemium model not five years ago, when every platform was free or cheap until it wasn’t. AI suffers from the same problem on steroids and the impacts are automated.

The AI usage cost and limit wars have begun. Now that Anthropic dropped the hammer on its massive customer base, the others can and will follow because the economics of unlimited were never going to work.

I told my friend to take his next “Black Friday” and however many days needed after that to refactor the code base completely so it can be managed by humans. He’s got one tech resource, so hire another. In fact, hire as many as it takes on a contract basis to turn his nightmare back into a dream.

Then, for the sake of all that is human, go back to at least some of the basics. Hire developers and product engineers and designers and proper sales and marketing people and let them use “Steve” and them boys as the tools they are supposed to be. Figure out what your real costs are, and adjust your offering and pricing accordingly.

There’s nothing wrong with automation, speed, and looking for an edge in the margins to separate you from the competition. But if you’re going to jump into the pool face first, you need to understand that sometimes they say it’s eight feet deep but it’s really only three feet deep.

Don’t smash your face on the bottom because you got lured in by a loss leader during a gold rush.

Hey, if I can help you create an ounce of prevention to avoid a pounding to your face, join my email list and get a heads up when bad things happen to people who aren’t you. And for the record, the pool thing really happened to another friend of mine. He’s OK. Now.

canonical=https://www.inc.com/joe-procopio/the-ai-usage-limit-wars-have-begun/91221027