When the Bill Arrived
Paul Krugman published a short video note this morning reacting to the semiconductor selloff. The Philadelphia Semiconductor Index dropped nearly 8% on Tuesday. The KOSPI fell close to 10%. NASDAQ was off 2.2%. He’s appropriately cautious about reading too much into it; the Philly Semi was up 157% over the prior year, so some perspective is warranted. But he flags something more interesting than the price action.
A few months ago, Corporate America was mandating AI adoption. Not recommending it. Scoring employees on token usage like a KPI. The message was clear: use AI or fall behind, whether or not you think it’s actually valuable. Companies were calling this “tokenmaxxing.” Managers were evaluating performance on it.
Then compute got scarce. Chip prices spiked. The big AI companies started charging real money for real usage. And the same organizations that had been grading workers on AI engagement reversed course almost overnight: stop, economize, reduce your token consumption.
That’s a remarkable thing to have had to write in a company memo.
The Signal in the Reversal
If mandated AI adoption had been producing meaningful output that mattered to the business, the order to cut back would have been painful. You’d be walking back productivity gains. Teams would push back. Someone would build a cost justification for the continued spend.
Instead, it was easy. Companies turned the dial down without much apparent friction. That tells you what you need to know about how much of that mandated usage was solving actual problems versus performing innovation for the benefit of the quarterly earnings call.
The gap between “we have an AI strategy” and “AI is solving a specific, measurable problem for us” has been visible for a while to anyone building in this space. What changed is that the price signal finally forced the reckoning that the productivity evidence hadn’t managed to trigger. Studies have been accumulating that show AI lets people produce more volume, more code, more copy, more output across every category, while the actual business payoff from that volume is much smaller than the volume itself would suggest. The evidence was there. It mostly got ignored because the pressure to look like you were doing something about AI was stronger than the incentive to measure whether it was working.
Cost is harder to ignore than research findings.
The Compute Question
Krugman highlights something Satya Nadella said recently: Microsoft shouldn’t be handing over all this money to the big AI companies, and cheaper models deserve more serious consideration. He hinted at DeepSeek, a Chinese model that’s less comprehensive but dramatically less compute-intensive and therefore dramatically cheaper.
That reframing matters if you’re a builder or operator thinking through AI tooling. The question isn’t “which frontier model is most capable.” The question is “what’s the minimum capability that actually solves the problem in front of me?”
For most business use cases, the answer probably isn’t the most expensive option on the table. The race to the frontier made sense when compute costs were being absorbed by investor enthusiasm and by AI companies subsidizing adoption to build market share. When every token carries a real marginal cost, the calculus changes. You start asking whether the capability differential between a $20-per-million-token model and a $2 one is worth ten times the price for this specific workflow. Often it isn’t.
This is part of why model selection is getting more interesting than it has been. Commoditization is happening faster than the AI companies want to acknowledge. The gap between the frontier and the tier below it is closing on the tasks that most businesses actually need done. Anyone who has vibecoded a working prototype on a mid-tier model knows you don’t need frontier-level reasoning to move the needle on most real problems. The frontier is impressive. For a lot of work, it’s also overkill.
The corollary for vendor negotiations: the threat of switching, or building, is more credible now than it was eighteen months ago. That changes leverage.
What This Changes
Not much, if you’ve been approaching this right. It never made sense to mandate AI usage without a problem statement. It never made sense to score employees on token consumption as a proxy for value creation. That was always a metric searching for a strategy.
What’s shifted is that the market correction, in stock prices and in corporate rhetoric, has made the underlying reality harder to dismiss. The AI-or-else posture was built on FOMO, not ROI. Krugman calls it a “quasi-bubble” in corporate behavior rather than asset prices. The framing is useful. It was social pressure and competitive anxiety, not evidence, driving a lot of the adoption.
The companies that identified specific problems, selected tools that actually addressed them, and measured outcomes rather than usage are in a better position now. Not because they were contrarian. Because they were disciplined.
Cheaper models, better cost discipline, and a market pricing in some skepticism: this is a more honest environment to build in than the one where everyone was required to tokenmaxx.
The noise is going down. That helps.