All-you-can-eat AI is about to end

What happens when AI companies stop subsidizing their users and charge what it really costs to build and operate their datacenters? Will people keep loving AI like they do now?

We should know soon, because AI companies are running out of people to fire. Lay off employees → Show Wall Street you have “fiscal discipline” → Boost your stock price → Plow more cash into your datacenters. That’s the current flow chart.

But it can’t go on forever. Eventually, users will have to be charged more. Already, $200 a month “pro” tiers are emerging. This is the price at which an average individual is profitable for the AI company.

And what about an enterprise user? Well, factor in all the costs--power, water, GPUs and, crucially, the rapid depreciation of those GPUs (most are obsolete in 1-3 years)--and enterprise users are looking at a bill of $120–$150 per month per employee.

If a company has 1,000 employees in AI seats, it’s suddenly facing a $150,000 a month bill for its AI subscription.

That’s a lot. And that’s when CFOs start asking which tasks are worth the AI spend. Will writing be one of them?

When AI is cheap (subsidized), companies treat it like a slot machine: they keep pulling the lever till something usable comes out. But as 2026 data shows, 80% of companies are already missing their AI cost forecasts by more than 25%.

They’re feeding more into the slot machine than they planned. They’re also winning less--and less often.

As LLMs train more and more on content generated by other LLMs, the outputs become more and more mushy. Getting usable content now means pulling the lever five, 10, 20 times, with each pull costing more tokens and the outputs taking more time to clean up.

At what point do companies decide the juice ain’t worth the squeeze? When do they reinvest in humans, an asset that doesn’t depreciate?

The Apollo program ran from 1960 to 1973 and cost around $280 billion in today’s dollars.

In 2026 alone, the combined capex of the “big four”--Microsoft, Meta, Google, Amazon--is projected to be $650 billion-$690 billion.

That’s two Apollo programs per year.

Unsustainable is an understatement. Very soon, AI companies will have to start charging users the full cost of keeping their datacenters up and running.

What happens then? Wait and see. It won’t be long.