The all-you-can-eat era of AI is over, and the invoices are landing.

Some companies are hitting their entire annual AI budget in three months. Others are watching their bills double or triple. After two years of “spend freely and experiment,” Corporate America is quietly starting to ration AI.

That shift, reported by The Wall Street Journal in late May 2026, doesn’t signal an AI winter. It signals something healthier: the end of a subsidy, and the start of an honest conversation about value.

Why the bills suddenly hurt

For a while, the math was upside down. Flat-rate subscriptions quietly subsidized the heaviest users, and providers often lost money on them. The unspoken deal: get people hooked now, fix the economics later.

Then pricing went usage-based. Every query started running a meter, a few fractions of a cent each time. Multiply that across thousands of employees using the most powerful models all day, and the bill stops looking like a rounding error.

So the brakes came on. Uber blew through its budget for autonomous AI by March. Microsoft cut access to an outside coding tool in favor of an internal one. Salesforce built a system to check whether spending actually maps to results.

Using a Ferrari to buy groceries

Some of the overspend was just habit. When usage looked free and “being AI-forward” was a status signal, people ran top-tier models to answer trivial questions, or just to make small talk.

Factory’s CEO put it well: if your kid needs help with algebra, you can find someone cheaper than Albert Einstein. Using a frontier model to reformat a list is the same mistake. Now that the meter is visible, that waste has a price tag.

Usage is not the same as impact

This is the line that matters most, from Meta’s CTO Andrew Bosworth:

“All motion is not progress, and token usage alone is not a measure of impact of any kind.”

For two years, usage was treated as a stand-in for value. More tokens, more transformation, great dashboards. But a team that triples its spend hasn’t necessarily shipped anything new.

The proof is stark. Across 2,000+ companies, only about 18% of spending on advanced AI coding tools turned into products that reached real users. The rest disappeared into experiments, dead ends, and the unglamorous work of debugging and rewriting AI-generated code.

But demand is still exploding

It would be easy to read this as a pullback. It isn’t.

Google now processes more than 3.2 quadrillion tokens a month, about seven times a year ago. Sales and usage keep outrunning forecasts. As one investor put it, “We’re still in the pretty early innings.”

Both things are true at once. Adoption is real and accelerating, and the era of spending without measuring is ending. Rationing isn’t retreat. It’s what happens when a technology has to start justifying itself.

What to actually do

If you run a team or a budget, the lesson isn’t “use less AI.” It’s “match the tool to the task, and measure what comes out”:

• Right-size the model. Save frontier models for hard problems. Cheaper ones handle most daily work fine.
• Route your queries. Send simple tasks to cheap models, hard ones to expensive models, automatically where you can.
• Track outcomes, not activity. The number worth watching is the share of AI spend that turns into something a customer uses.

My take