Why AI is affordable today

Many Mittelstand owners carry a number in their heads that is long out of date. It comes from the time when AI projects sounded like big-consultancy work: six figures, months long, with an uncertain outcome. Anyone who measures AI against that old number reaches the sensible conclusion to wait. Only the number is no longer right.

What has changed is no small thing. It is one of the fastest price drops any technology has ever seen.

The floor has fallen out

The most solid analysis of this comes from the research institute Epoch AI. Across six common performance tests, the price to reach a given level of AI has fallen by a median of roughly fiftyfold per year. Depending on the task, the drop lies between ninefold and nine-hundredfold. Fifty per year is the middle value.

In concrete terms: the performance you still paid around 30 dollars per million tokens for when GPT-4 launched in spring 2023 costs a fraction of that today. What used to justify only the most expensive applications can now sit in almost every everyday step of work. The question has flipped. It is no longer: can we afford AI here? But: is there a reason not to use it here?

The old question was whether AI is worth it. The new question is why not.

What has got cheaper, and what hasn't

Here is the honest framing you rarely hear. What has got cheaper is the model, the raw material. What has not automatically got cheaper is everything around it that is needed for it to actually run in the business.

That matches what companies report. A third of the businesses using AI say it ended up more expensive than expected. That is no contradiction to the price drop. It just shows where the cost sits today: no longer in the model, but in the integration into existing workflows, in data upkeep, and in the fact that in the end someone carries the responsibility. Anyone who does the maths from the providers' shop-window prices misjudges it. Anyone who does it honestly sees: the expensive part is the implementation, not the technology.

What that means for getting started

That is exactly why no mega-project is needed today. When the raw material is cheap, the lever shifts to finding the right two or three use cases and implementing them cleanly, instead of buying an expensive all-round transformation.

That allows an entry that wasn't possible before: with a fixed, small price that is set before you start. No open day rates. No concept that ends in a drawer. A working result in weeks, not months.

And it allows something the Mittelstand cares about: what gets built runs in your own house and stays with the company. Vendor-neutral, with the best tools for each task. No lock-in, no dependency on a single provider whose prices you don't control.

The price drop has scrapped the old excuse. AI is no longer the expensive gamble it once was. The only question left is where in your own business it pays off first.