The wrong million tokens
Uber burned its entire 2026 AI budget in four months. The real problem wasn't the tokens—it was the missing denominator.
Why it matters
Enterprise AI spending caps are a symptom of unmeasured value, not a solution. The industry controls token volume while ignoring cost-per-task—the metric that lets you defend the bill to finance. For practitioners, this is a blueprint for measuring agentic workloads before they spiral.
The key facts
9 to knowUber deployed Claude Code in December 2025; 84% of engineers adopted it by March 2026, exhausting full 2026 AI budget in four months
Uber's response: hard $1,500/employee/month cap per tool (reported by Bloomberg, June 2026)
Uber COO Andrew Macdonald: 'very hard to draw a line' between token spend and actual feature output—measurement failure, not cost failure
Token cost compounds quadratically in naive agent loops: 20-step loop with 1K tokens/step yields 210K billed input tokens, not 20K
Prompt caching reduces 210K to 44K effective tokens (79% savings) but leaves 2.2x the naive estimate; cache invalidation on prefix changes can push back to 90K
Liu et al. (TACL 2024) 'Lost in the Middle': accuracy follows U-shaped curve; drops with input size even on long-context models
Du et al. (EMNLP 2025 Findings): holding retrieval perfect and growing input anyway reduced accuracy 13.9–85%; damage concentrated in first 7K tokens
Six cost-reduction techniques identified: metadata filtering, reranking, semantic chunking, transcript summarization, deduplication, stop-point definition
Core metric missing: cost per task (resolved ticket, merged PR, closed claim), not cost per token or per seat
Go to the source
CIOcio.com
Publisher excerpt: Last month I made the case for an ROI exchange rate: the formula you negotiate with finance before deployment that converts KPI movement into dollars. One point of first-call resolution equals this many dollars. One hour of engineering time recovered equals that many. A common objection was a…