Uber president says AI spend is ‘getting harder to justify’ as token-versus-headcount maths bites
TL;DR:
- Uber president and COO Andrew Macdonald told Rapid Response the company cannot yet draw a clear connection between rising Claude Code token consumption and additional useful consumer features being shipped.
- Uber reportedly burned through its annual AI budget within four months of 2026, after spending $3.4 billion (£2.5 billion) on R&D in 2025, up 9% year-on-year, with CEO Dara Khosrowshahi flagging earlier this month that hiring restraint is the offset.
- Macdonald framed the open question as a straight token-versus-headcount trade: “if you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping, that trade becomes harder to justify”.
The comments are notable because Uber has been one of the loudest enterprise endorsers of Claude Code, and the COO’s framing matches a broader pattern of late-2025 enterprise AI buyers asking vendors for evidence rather than promises — a shift Target signalled earlier this week and one Resultsense covered in yesterday’s piece on usage-pricing pressure.
Context and Background
The structural problem Macdonald describes is increasingly common: developer-productivity metrics (acceptance rates, token throughput, lines of generated code) are “trending in a really astronomical direction” but do not translate cleanly into shipped feature counts that finance functions can put on a P&L line. UK CFOs running similar pilots — including HSBC, Standard Chartered and several Tier-1 banks — face the same accounting challenge as token-based usage pricing replaces seat-licence economics.
Khosrowshahi’s hiring-restraint framing is the candid version of what most enterprises are doing without saying so: the AI bill is being paid for by not filling headcount that would otherwise have been added. That works at portfolio level but does not, as Macdonald notes, give engineering leaders a defensible per-team ROI story. The Verge piece offers no resolution; Uber is openly admitting it does not yet have one.
For UK business readers, the relevant frame is that one of the most aggressive AI-adopting enterprises in the market — with a US balance sheet and engineering culture to match — is publicly conceding the ROI link is unproven. UK SMEs evaluating Claude, Copilot or Cursor deployments will read this as permission to demand the same feature-shipping evidence from their vendors that Macdonald says Uber cannot produce internally.
Looking Forward
Expect a wave of CFO-led AI-spend reviews through the second half of 2026 as 2025-vintage pilots come up for renewal. The metrics question Macdonald has surfaced — token economics versus feature throughput — will increasingly anchor those conversations. Anthropic and OpenAI’s enterprise teams have moved to address this with delivery-engineering offers and Forward Deployed Engineer hires, but the substantive proof points are still pending. The next two earnings cycles will tell whether the productivity claims survive contact with finance.