KI Blase

I've had some money sitting on the sidelines for a while. Not a lot. But enough to now seriously ask: where is it better off than in a bank account?

AI companies were an obvious candidate. I work with AI every day and see what this technology does inside my businesses. Of course I thought about it.

So I did the research. Not for an article. For myself.

What I found was an answer I hadn't expected to be this clear.

Two markets moving in opposite directions

Everywhere you look, you read the same story right now: rate limits are being cut. Frontier models are getting more expensive. First cracks in the AI bubble. Anyone who actually knows the numbers sees something different.

March 2023 · GPT-4 Launch

Around $30 per million tokens via the API. That's what companies pay when they embed AI into their own workflows. Automated. In the background. For hundreds or thousands of processes simultaneously.

May 2026 · Today

GPT-4-comparable quality via models like DeepSeek V4: around $0.50 per million tokens. A factor of 50 to 100 cheaper. In three years.

At the same time, the absolute frontier models have indeed gotten more expensive. GPT-5.5, released at the end of April 2026, costs $5 input and $30 output per million tokens. Twice as much as its predecessor two months earlier.

"Two markets, one technology: at the top it's getting more expensive, at the bottom it's getting dramatically cheaper."

What businesses actually need

No mid-sized business needs the most powerful model on the market for its daily processes. What's needed is reliable GPT-4-comparable performance for routine tasks: customer communication, internal documentation, analysis, translation.

Exactly this performance class has become 50 to 100 times cheaper in three years. Not incrementally. Not moderately. Factor 50.

every 2 mo.
AI costs at equal performance halve, according to Epoch AI
3.3×
global AI compute capacity grows per year since 2022. Faster than Moore's Law.
50–100×
cheaper is GPT-4-comparable quality today vs. launch in March 2023

Anyone talking about a bursting bubble when looking at these numbers is watching the wrong indicator.

What AI companies' losses actually mean

OpenAI: approx. $14B loss in 2026. Anthropic: approx. $11B. That sounds dramatic. Until you understand what the money is being spent on: infrastructure. Compute capacity. Model training. This isn't a sign of weakness. This is the largest infrastructure build program technology history has ever seen. And it's running at full speed right now.

For the investment question, this means: whoever isn't positioned now is waiting for a moment that won't come. The structures of this industry are being decided right now. Whoever wants to enter in two years, when the losses are gone, will be buying in at a higher price.

The two groups that are going to get burned

📉
The waiting investors
They're waiting for the moment when the industry is "proven." That moment doesn't exist. There's a market consolidating right now. Whose bet pays off isn't decided in two years. It's being decided now.
⏳
The waiting businesses
They're waiting until AI is "more mature." The performance available today for $0.50 was unaffordable for any business three years ago. What will be available at that price in three years is something we can barely imagine today.

I've been running AI processes across three companies for over three years. The businesses that waited back then with "too expensive" or "not mature enough" are already far behind. They have less experience and less established processes. In 18 months, that gap will be even larger.

The bubble isn't bursting. But the train keeps moving.

With or without you.

"What is concretely stopping your business from starting today? That's the question I have to ask far more often than I should."