The Cracks are Starting To Show

The dominos are falling

The cracks are starting to show.

AI companies are running out of ways to hide how expensive this all is.

And the last two weeks made that very obvious.

What Anthropic just did

Anthropic ran a painted door test.

Simple concept. You show a different price on your page to a small percentage of visitors. See how many still convert. Figure out how much more money you can extract without losing too many customers.

Anthropic’s version: they quietly removed Claude Code from the $20 plan for some users. Showed them a $100 plan instead.

No warning. No explanation. Just — Claude Code isn’t available at your price tier.

Some people paid the $100.

That’s the test. See how many will.

Why they had to do it

Every time someone uses Claude, Anthropic loses money on the inference cost.

Not because they’re generous. Because the model training costs haven’t been recovered yet.

Opus 4.5 cost a fortune to build. If the inference revenue from Opus 4.5 didn’t cover that cost before Opus 4.6 replaced it — Opus 4.5 was just a loss. Pure and simple.

OpenAI just raised $120 billion. That’s enough to run roughly 18 to 24 months at their current burn rate.

5 to 7 billion dollars a month in the hole.

That’s the actual situation. Which means price tests aren’t optional anymore. They’re survival.

Then Microsoft did the same thing

GitHub Copilot changed their pricing model.

Before: pay a flat rate, get a set number of actions.

Now: token-based usage. Because not every model costs the same.

A fast cheap model costs almost nothing per call. An expensive frontier model costs 20 times more.

Charging everyone the same flat rate while they used wildly different models was always mathematically broken. They just couldn’t admit it until now.

Google is the only one laughing

Google is pouring $100 billion plus per year into AI.

And after all of that they still make money.

They don’t have to do painted door tests. They don’t have to terror-market their users about job replacement to juice fundraising rounds. They don’t have to answer to nervous investors every quarter wondering if the business model works.

They invented the transformer architecture that made all of this possible. Somehow weren’t first to market anyway. And now they’re watching competitors burn through investor cash while Google just quietly keeps building.

That’s also why you don’t hear the same apocalyptic hype from Google. They don’t need it. Dario and Sam need to raise billions constantly. Google just needs to ship.

The Uber situation explains everything

Uber told every employee to use AI maximally. Made AI usage a performance metric. Judged people on how much they were using it.

Then burned through their entire annual AI budget in four months.

Shocked face.

When you tell thousands of employees that their performance review depends on AI usage and then hand them access to frontier models — you should not be surprised when they use frontier models constantly.

The economics of that decision were never run. They just assumed more AI usage equals more productivity equals worth it.

It didn’t work out that way.

What’s actually happening

Nobody is going back to hand coding everything. Nobody is abandoning AI.

But the era of unlimited access at flat rates is quietly ending.

Token budgets. Usage limits. Tiered pricing. Model restrictions.

It was always going to come to this. You cannot spend 5 billion a month indefinitely hoping inference revenue catches up eventually.

The hype was always partly fundraising marketing. Dario telling you your job is gone in three years isn’t just a prediction. It’s a pitch to investors. Fear moves capital.

Google doesn’t need to scare anyone. They already have the capital.

Everyone else is running a very expensive experiment with other people’s money. And the bills are starting to come due.

Want more stuff like this? Even more detailed? I go into deep dives every single day on my 📌Substack📌

At the price of a coffee I can give you enough knowledge to replace a bachelor’s degree in computer science or maybe even a masters, who knows.

💜 If you like my work and would like to support to me financially. Even a small donation would help a lot!

My PayPal — Please Support