CODE IN TOKYO

The Weight of Compute Resources, and What Comes After Endless Evolution

Claude Fable 5’s pricing has been announced, and as expected, it has sparked discussion everywhere.

To be honest, I froze for a moment when I first saw the pricing table.

“Hmm, this is expensive if I want to use it heavily every day.”

That was my honest reaction.

It is currently in a trial period, but it will become paid starting June 22.

Based on how it feels to use right now, I probably will not go as far as paying extra to use it as my daily model.

AI Is Not Just Software Anymore

Still, when I step back and think about it calmly, I can understand why the price is set this way.

AI looks like it is running inside a screen, but in reality it runs on top of data centers filled with massive numbers of GPUs.

It needs land.

It needs electricity.

It needs cooling systems.

It needs networks.

And above all, it needs GPUs.

Recent models are no longer just “smart software.” They are closer to enormous factories.

The Industry Is Fighting With Compute

The AI industry today is competing by throwing compute resources at the problem.

Of course, algorithms are improving too.

Efficiency is also getting better.

But in the end, it becomes an extremely physical question:

“How many GPUs do you have?”

OpenAI, Google, Anthropic, and others are all building data centers around the world.

But land has limits.

Electricity has limits.

Semiconductor manufacturing capacity has limits.

This is not a world that can scale infinitely just because someone has enough money.

The Current Intelligence Race Will Hit a Wall

That is why I think the current extension of the “intelligence race” will eventually reach a ceiling.

Of course, models will keep improving next year.

They will improve the year after that too.

But I do not think the explosive growth we have seen over the past year can continue for ten straight years.

At some point, humanity will need to look for the next move.

Maybe it will be a new algorithm.

Maybe it will be a new computing principle that replaces GPUs themselves.

Maybe it will be quantum computing.

Or maybe it will be some surprisingly plain optimization technique.

I do not know what will arrive.

But the current direction of “bigger and more massive” seems to have limits.

The Bill From Physics

Seen this way, the new pricing feels somewhat symbolic.

The improvement of AI intelligence has finally begun sending us the bill from the laws of physics.

Until now, companies had been covering that bill while accepting losses.

But recently, that bill has slowly started to reach users too.

“Is this intelligence really worth paying that price?”

That is the question being asked.

That is why I have recently become more interested in how models are used than in the models themselves.

If you run an expensive model endlessly, you can get smart answers.

That is obvious.

What is truly interesting is:

This is now an engineering problem.

Engineering Has Always Been About Limited Resources

In the past, programs became faster when CPU clock speeds went up.

Then came the multicore era.

Then came the cloud era.

And now we are in the AI era.

In the end, what engineers do has not really changed.

We are simply thinking about how to use limited resources.

Of course, while we seriously think about cost optimization, there is also a very real chance that a smarter model will appear next month and overturn everything.

Watching the AI industry feels less like making a future prediction and more like checking the weather forecast.

So my conclusion is this:

Do not force yourself to chase premium models.

Borrow them only when necessary.

And use the money you save to eat a bowl of gyudon, Japan’s classic beef rice bowl.

At least for now, no AI has surpassed the cost performance of that bowl.