This week the price tags showed up. Uber capped employee AI spending at $1,500 a month — a number Simon Willison flagged as a real signal, not a curiosity. NPR asked whether a future where households actually pay for AI ends in free tools or one more subscription line. And Broadcom shed $300 billion in market value because its AI revenue merely met expectations instead of beating them.
For three years the defining question about AI was what it can do. That question is closing. The new one is what it costs, and who pays — and that shift is the most bullish thing to happen to the field, not the most worrying.
Capability was always the easy thing to demo and the hard thing to price. A demo costs nothing; a deployment costs tokens, and tokens cost money someone has to attribute to a line item. Uber's cap is a company discovering that a single engineer's agent habit is worth four figures a month, and deciding that's the ceiling. The household-subscription question is the same discovery pointed at consumers. Broadcom's $300 billion is the market doing the arithmetic in reverse: if revenue only meets the forecast, the forecast was the magic, and the magic is now a spreadsheet. Stratechery named it this week — in an AI buildout financed by borrowed capital, capital becomes the ultimate commodity.
Things only get priced when they're real enough to budget for. Nobody runs a cost-control meeting about a toy.
The obvious read is that this is the air leaving the balloon — repricing as disappointment, the comedown after the hype cycle. It isn't. Things only get priced when they're real enough to budget for. Nobody runs a cost-control meeting about a toy. The $1,500 cap exists precisely because $1,500 of monthly value was real enough to need a ceiling.
The exciting phase of a technology isn't when it's free. It's when someone finally writes down what it's worth. That happened three times this week.