That’s because the amount of AI investment looks like orders of magnitude more than anyone is willing to pay for AI products.

They were happy about spending billions, but now the bubble is into the trillions and there isn’t tens of trillions in market demand to make an investment that big pay off.

That assumes demand is fixed. Historically, the biggest technology shifts created entirely new markets that barely existed beforehand. Almost nobody predicted today’s cloud or app economy from early internet revenue. AI spending could still prove excessive, but current demand is a poor ceiling for what future demand might become.

“If you build it, they will come” refers to a homemade baseball field in a work of fiction. It’s not a valid argument for spending trillions of dollars to support demand that doesn’t exist just on the off chance that some day it might.

Current buyers are looking to cut back on spending, rather than dramatically increase. And that’s based on current prices which are being subsidised by burning investor money, not the prices they need to charge to make a profit.

Cloud and apps didn’t need trillions in investment to still not get off the ground. There isn’t really any scenario in which current AI spending doesn’t turn out to be excessive.

Maybe, but »excessive investment« and »failed technology« aren’t the same thing: Railroads, fiber optics and the dot com era all burned absurd amounts of capital, yet the infrastructure outlived the investors. AI could follow the same pattern: terrible returns for today’s shareholders, enormous value for tomorrow’s economy.

That would certainly be the good ending, I won’t hold my breath for it

Yeah, but the AI infrastructure is made out of compute hardware that’s going to need replaced in 6 years. Not rails that will still be useable in 50 years.

We aren’t building lasting infrastructure other than the actual buildings.

I’d argue the models and software are the real long term assets. GPUs depreciate like any other hardware, but a better training pipeline, inference stack, proprietary data, and a model with millions of paying users can survive multiple hardware generations. The chips are replaceable. The ecosystem and customer relationships are much harder to replicate.

Yeah, there’s a lot of innovation going on to run powerful models on modest hardware and the state of the art for running local models is changing all the time. If the trend continues, models hosted in a data center will only be for niche use cases.