Uber's 2026 AI budget ran out in April

Someone asks for a summary, gets something unusable, pastes in two more documents, restates the question and runs it again. That fourth attempt is where Fast Company locates the money: unclear requests, failed outputs, repeated tries, and conversations still hauling irrelevant context forward ^[7]. The same piece notes research finding that the cluttered conversation degrades the answer even when the relevant material is still sitting in it ^[8]. So attempt four costs more than attempt one and returns less, and at no point does the person making it see a price.

The worked example is a $2.25 workflow run twice a day, reaching $99 a month for one employee ^[2]. Back the arithmetic out: $99 divided by $4.50 a day is 22 working days, so the figure assumes an office calendar and no weekend agents ^[3]. Then $99 times 12 is $1,188 a head, and a thousand heads is $1,188,000 a year, which is the article's total, rounded ^[4]. The exposure tracks how many people you handed access to, not how ambitious the roadmap is.

The energy half is shakier, and the author is open about the method: roughly 200 joules per token, backed out of published Nvidia chip specifications, standard power usage effectiveness benchmarks and average estimates of how fast models handle data ^[9]. A million-token multi-agent run therefore comes to 55.6 kWh ^[18], and the "nearly two days of household electricity" comparison implies a home drawing about 28 kWh a day ^[19]. Apply the same constant to Google's 3.2 quadrillion tokens a month and you get 177.8 TWh ^[5], a number the article does not reconcile with anything. One joules-per-token constant cannot cover a batch embedding job and a multi-agent reasoning run at once, and a label that ships with one will be wrong in a direction the buyer cannot check.

Worth separating the pitch from the practice here. The nutrition label is a prediction in the article, not a product anyone is shown shipping ^[14]. Its supporting evidence is a study, unnamed in the piece, in which organisations with full visibility into AI operating costs were five times likelier to report established ROI ^[6][10], plus OpenAI's enterprise lead reporting that customers now ask whether they can audit a model's efficiency ^[1]. That correlation travels with finance discipline in general: teams that can see per-prompt cost tend to be teams that already tag spend and own the line it lands on.

The person this is for is whoever holds the AI invoice and cannot currently state the cost of one completed piece of work. Two axes decide whether metering does anything. Does the user see cost before committing, and does that cost land on a budget the user's team answers for. Visible cost without ownership produces a flicker of curiosity and nothing else. Ownership without visibility produces hoarding and quota gaming, which is roughly what a token cap buys, since a cap limits volume without saying whether the compute matched the outcome ^[11]. Leaderboards supply neither, and leaderboards are what a lot of firms installed last year ^[12]. Only the corner with both changes attempt four.

The version of this available without waiting on a vendor: take your five heaviest AI workflows and try to write cost per resolved ticket or per merged change beside each. Where the cell stays empty, the label would have been decoration.