Uber's 2026 AI budget lasted four months across roughly 5,000 engineers

A seat count is a poor predictor of the bill, and the reason is in how an agent talks to a model. Each tool call re-sends the full accumulated conversation, so an agent making 20 calls during one task sends its context 20 times ^[8]. If the context grows at every step, the tokens billed for that task rise with the square of the step count ^[6]. The post cites an audit that attributed 62 percent of total agentic cost to re-sent context alone ^[9]. Strip all of it and 38 percent of the bill remains, about 2.6 times cheaper per task ^[2].

Uber's number gives a run rate. Claude Code went to roughly 5,000 engineers in December 2025, and the full-year budget was gone by April 2026 ^[1]. Assume the budget year is the calendar year: four months of spend covered twelve months of plan, a run rate near three times forecast ^[1]. The post says Uber's CTO has since confirmed the figure publicly ^[3].

Unit price is moving the other way. Per-token inference cost fell roughly 75 percent year over year ^[4], while enterprise generative AI spend went from $11.5 billion in 2024 to $37 billion in 2025 ^[6]. That is spend up 3.2 times against a price at a quarter of its level, which implies token volume up about 13 times ^[3][4]. Take the 13 as an upper bound: reasoning-optimised models passed 50 percent of all enterprise tokens by late 2025 and cost more per token than their predecessors ^[10], so the average price actually paid fell by less than the headline figure.

Jellyfish's 18.6x per-developer increase over nine months is a measurement of the population Jellyfish observes, driven almost entirely by agentic coding features ^[5]. It transfers to another engineering org only if developers there switched on agentic features on a similar schedule and the agents run similar tool-call depth. Average prompt length grew about 4x since early 2024, from around 1,500 tokens per request to over 6,000 ^[7]. The two windows do not line up, so treat the division as indicative: 18.6 over 4 implies roughly 4.7 times more model requests per developer ^[5].

Attribution has to be engineered before rollout. An LLM API bill arrives as a single line item, with no breakdown by workflow or feature and no customer-level view ^[11]. The post puts the miss rate of traditional FinOps tooling at 70 to 90 percent of real AI costs ^[12], and cites the FinOps Foundation finding that 98 percent of organisations now actively manage AI spend, up from 31 percent two years earlier ^[13].

One caution on the evidence. All of this sits in one post on dev.to, whose figures are marked with bracketed reference numbers that the supplied text does not resolve ^[15]. The $500 million single-month Claude bill appears there as a report of one unnamed company that deployed access without usage caps ^[14]. Goldman Sachs' 24x global token growth by 2030 is a projection ^[16].