Big tech’s AI boom is burning budgets — and CTOs are scrambling

Runaway adoption, runaway bills

It has been reported that Uber deployed Anthropic’s Claude Code to engineering teams in December 2025 — and that the company’s CTO told management the annual AI tools budget was exhausted within the first four months of the year. Reportedly, about 70% of new commits in some repositories were AI‑generated, and per‑engineer AI tool costs were estimated at $500–$2,000 a month, driving millions in monthly spend. Those figures, if accurate, upend the usual corporate budgeting math: adoption rates and unit costs that used to be predictable now have no natural ceiling.

A new managerial problem, not just a technical one

Why is this different? Traditional enterprise software suffers from under‑use; companies spend big to get systems adopted. AI coding assistants flip that script. Engineers self‑service the tools, ramp instantly, and—because the marginal cost of asking the model another question is low—usage snowballs. Tokenmaxxing has even emerged as a social metric among engineers: “How many tokens did you burn this month?” As Axios reportedly noted, some teams’ compute bills have begun to eclipse payroll for deep‑learning work. Add strained GPU supply and geopolitical pressures on chip exports, and compute becomes both expensive and strategically fraught.

Misaligned incentives and blunt controls

The result is a three‑way incentive mismatch: vendors benefit from volume, engineers reward immediate productivity gains, and finance teams face an unpredictable line item. It has been reported that companies are already experimenting with blunt controls — monthly caps, mandatory requirements documents, routing simple tasks to smaller models — but those measures feel paradoxical: buying a productivity tool and then policing its use. Who pays when AI becomes “good enough”? That question is fast replacing “can AI replace engineers?” as the industry’s most urgent debate.

The Uber episode, if representative, signals a shift in how firms must value engineering labor and compute. Senior executives now need ROI models for continuous consumption, not one‑time seat licenses. And for regulators and competitors watching from abroad, rising compute demand — concentrated in a few chipmakers — raises fresh questions about resilience and strategic dependence. In short: AI has solved part of the productivity problem. It hasn’t yet solved the accounting one.