Uber's CTO calls time on tokenmaxxing after quadrupling its frontier AI users

Bain's two figures multiply out to a clear direction for the sector bill: 5.5 times the volume bought at half the December 2024 unit price is about 2.75 times the spend ^[12][17]. The Silicon Data Token Expenditure Index runs the same way, with a token price down more than 90% since 2023 and large language model spending doubled since late last year ^[11].

"As tokens get cheaper, companies don't spend less," Apollo chief economist Torsten Slok wrote in a recent blog post. Instead they "run more AI agents, automate more workflows, and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses" ^[13].

What Naga listed is unit-price work: better prompt caching, a changed default model setting, evaluation of new models for efficiency, and letting engineers see their own usage and cost per hour ^[7]. "You might expect costs to rise as adoption accelerates," he wrote. "We've seen the opposite. Not because we've restricted access, but because we've treated efficiency as an engineering problem rather than a budget problem" ^[8].

Read the post as a claim about the total bill and a per-head number follows. Four times the users on a smaller total puts spend per user below a quarter of its earlier level ^[6][18]. Take it as a claim about cost per token alone, and the invoice can still be larger than it was in March.

Fortune reported that many companies backed off usage incentives after finding they did not deliver returns to justify the spending, without putting a figure on how many ^[4]. Uber's own account of the returns was thin in May. President and chief operating officer Andrew Macdonald said on the Rapid Response podcast, "That link is not there yet," adding that it was "very hard to draw a line between one of those stats and 'Okay, now we're actually producing like 25% more useful consumer features'" ^[16]. Deutsche Bank Research Institute's Jim Reid warned last month that AI productivity gains were still years away ^[14]. Between the first quarters of 2023 and 2026 the Magnificent Seven's profit margins went from 15% to 25%, a rise of ten points, while the rest of the S&P 500 saw only 10% margin growth ^[15][19].

I think the tokenmaxxing that ended at Uber is the leaderboard. Whether the compute bill kept growing is a separate question: the ranking scheme went, the number of people on the tools quadrupled, and the published figure stops at cost per token ^[3][6][5]. If aggregate spend really has fallen while four times as many employees use frontier tools, that is evidence enterprise inference pricing is deflating faster than usage grows. The companies selling tokens by the million have a revenue problem arriving. Uber's total AI spend for this quarter, set against the budget it went through in the first months of the year, would decide between the two ^[1].