Intelligence got cheaper. Running it got more expensive.

Put the number in its place first: $500 billion is a statement of intent, not a signed cheque. In announcements like this the big figure exists for the headline. The news is the calendar, not the sum. Memory for a facility opening in 2027 is being allocated in July 2026.

Where this touches a small company is not the server room, it is the invoice. High bandwidth memory and ordinary memory come out of the same plants and compete for the same line capacity. When a memory maker commits a year of output to one buyer, everyone else draws from a smaller pool. Memory costs are pointing up across the board, from laptops to servers. If you need hardware, waiting for the price to fall may cost you this time.

The pattern of the last two years held again: the price of the best model halves before the model is obsolete. The practical consequence is that budgeting AI cost at today's price is meaningless. If you plan six months out with today's token price as your floor, you will build an over cautious budget.

But watch what is actually getting cheaper. The model is getting cheaper. The deployment is not. Even if tokens went to zero, deciding which job to hand an agent, measuring it, and catching it when it gets things wrong costs the same effort it always did. Plenty of people read a price list and conclude AI is cheap now. The next story is that effort with a price tag on it.

The clearest signal in this story is the missing price list. The company with the best model distribution on earth is not selling agent deployment as software you install yourself. It sends engineers. The reason is not technical. The hard part is not connecting the model, it is writing the rules of the job: when approval is required, when a human takes over, how success is measured.

For a small studio that is good news. Anything that can be self installed eventually becomes free. The part that cannot is the work we do. Read the 75 percent figure for what it is, too: OpenAI reporting on its own line, with its own product, using its own measurement. That is a vendor number, not independent verification.

Nine people, no commercial customers, no product, and €12 million. What makes this a story is not the size of the round but that European capital put that much into pre product research. The familiar European pattern runs the other way: revenue first, investment after.

The claim itself is worth noting. Today's models match patterns rather than reason about cause, which is why they stumble on situations they have not seen. That is exactly where kausable is playing. Whether it works is unknown, and the company says itself that it is still at the research stage. Still, at a moment when European AI usually means wrapping an American model in a new interface, a team raising this much to change the underlying approach is news on its own.