Uber's COO Says It Is Getting Harder to Justify AI Tokenmaxxing Spend
Uber COO Andrew Macdonald said in a published interview this weekend that the company is struggling to justify its AI token spending. In a Business Insider Rapid Response interview, Macdonald said that after conversations with senior engineering leaders across Uber’s organization, the ROI on tokenmaxxing is not visible.
The statement is notable because it comes from the operations side of the house, not engineering. Uber’s CTO disclosed earlier in May that the company’s entire 2026 AI tooling budget was exhausted by April, with Claude Code costing individual engineers $500 to $2,000 per month across a 5,000-person engineering org. What was previously framed as a spend-management surprise is now, publicly, a cost-justification problem.
What Tokenmaxxing Costs at Uber’s Scale
Tokenmaxxing refers to saturating frontier model context windows and response generation on the theory that more tokens consumed produces higher-quality outputs. The practice accelerated in 2025 as enterprises opened AI tooling budgets and engineers discovered that frontier models respond well to large context payloads.
At Uber’s scale the economics are unforgiving. A senior engineering population of several thousand engineers, each consuming toward the upper end of the $500-$2,000 monthly range, puts annual AI tooling spend in the $30M to $120M range before infrastructure and orchestration layers. If that spend is not producing measurable acceleration in shipping velocity or defect reduction, the number has no natural ceiling and no internal defense.
Macdonald’s comment suggests Uber has not yet closed that loop. Engineers are consuming at capacity. The throughput metrics look healthy. The business outcomes are not yet traceable to the spend.
A Different Framing Than Google’s
At I/O 2026 this month, Google CEO Sundar Pichai cited 7x growth in AI product token consumption as a positive signal, using “tokenmaxxing” as shorthand for deep AI engagement. The implication was that tokens consumed correlate with value delivered.
Uber’s COO is questioning that assumption directly. Consumption does not equal productivity, and productivity does not equal business impact. A coding agent that writes more code, faster, is not generating value if that code is not shipping to production, not reducing operational incidents, or not closing product gaps that matter to revenue.
The distinction matters because enterprise AI renewal cycles are now approaching. Most large companies rolled out AI coding tools in late 2024 and early 2025 on one-year budget commitments. Those commitments are due for review.
What Comes After the First Wave
The first wave of enterprise AI adoption was driven by competitive fear and bottom-up developer demand. Procurement teams signed off on tools that engineers wanted access to, with productivity gains cited in qualitative terms: engineers feel faster, PRs close faster, onboarding is easier.
The second wave will need harder numbers. Uber’s COO is not announcing a cut. But a C-suite executive publicly framing token spend as hard to justify is a signal that the tools business cannot rely on the original buy-in rationale indefinitely.
The AI coding tools market, which includes Claude Code, GitHub Copilot, Cursor, and a growing set of agent frameworks, is priced on the assumption that productivity gains will be self-evident and self-reinforcing. Uber is a large-scale test of that assumption. The test is not going cleanly.