Uber Questions Whether AI Spending Is Delivering Real Results

Four months. That's how long it took Uber to consume its entire annual AI budget in 2026. The figure is revealing on its own, but it provides the backdrop for some striking statements: Andrew Macdonald, Uber's president and COO, has admitted in an interview with Rapid Response, reported by The Verge, that spending on artificial intelligence is becoming "difficult to justify". The specific reason: a sustained increase in Claude Code token consumption that isn't producing a corresponding improvement in productivity.

This isn't a minor complaint from a frustrated executive. It's a signal that the narrative of "more usage equals more value" is starting to crack in some of the corporate environments most exposed to these tools.

More Tokens, More Productivity?

The problem Macdonald describes is one of broken correlation. Companies that have adopted Claude Code at scale, Anthropic's official CLI with support for skills, subagents, hooks, and MCP servers, are discovering that the volume of model calls grows organically and almost inevitably as engineering teams integrate the assistant into more parts of their workflow. But measurable output, code in production, closed bugs, shortened release cycles, doesn't scale at the same pace.

This doesn't necessarily mean Claude Code doesn't work. It means that the metric "tokens consumed" is a poor proxy for value generated, and that many organizations still lack dashboards that let them separate productive use from noise: redundant queries, poorly designed contexts, iterations that lead nowhere.

Why It Matters That Uber Says It

Uber isn't a startup exploring AI for the first time. It's a company with thousands of engineers, its own infrastructure, and dedicated teams focused on adopting development tools. If they're struggling to draw a clear line between investment and return, it's reasonable to assume the problem is structural and widespread, not an isolated case of poor implementation.

The moment a COO of this caliber publicly verbalizes that spending "is difficult to justify" has practical consequences: it reinforces the position of finance teams who have been asking for stronger metrics before approving new budgets, and puts pressure on vendors, in this case Anthropic, to demonstrate value in more granular ways.

Anthropic has bet heavily on Claude Code as a vector for penetrating enterprise environments. The tool has gained real traction, with an ecosystem of plugins and MCP servers that grows week by week. But the conversation emerging in companies like Uber suggests the next battle isn't about adoption, it's about justification.

Who This Debate Affects

This kind of friction matters to several different profiles:

• Engineering teams that use Claude Code daily and need to argue internally why their budget shouldn't be cut.
• Technology procurement managers who are evaluating whether to scale licenses or consolidate existing usage.
• Independent developers and small studios, who paradoxically may come out stronger: their more controlled scale makes it easier to measure ROI directly, something large organizations find harder to do.
• The integrator ecosystem itself, such as the work we do at ElephantPink with agents and MCP servers, where the design of the consumption architecture can make the difference between justified spending and costs that spiral out of control.

The underlying problem isn't that Claude Code is ineffective. It's that large organizations tend to deploy AI tools without redesigning the processes around them, which generates inflated consumption with diluted value.

---

EP View: The fact that Uber is saying this out loud is, in a sense, useful for the ecosystem. The conversation about value metrics in AI development tools has been happening too long in private; having it on the table, with real numbers, is a better starting point than continuing to measure success by the volume of tokens billed.