Enterprise AI ROI Reckoning: Token Spend Meets Hard Returns
Enterprise AI ROI accountability is arriving faster than hyperscaler stock prices acknowledge. Within a single month: Uber burned through its entire 2026 AI budget, Microsoft cancelled most Claude Code licenses on cost grounds, and a Fortune 20 CEO ordered token spending "dramatically slashed" after $200M in spend produced only modest savings. The Financial Times calculates negative implied returns on AI capex for four of five major hyperscalers — even under zero-cost assumptions.
What the Source Actually Says
@ThierryBorgeat's viral X thread catalogs the receipts from the last 30 days. Microsoft cancelled most of its Claude Code licenses citing cost; Uber exhausted its entire 2026 AI budget in four months, with the COO publicly stating costs are "harder to justify." A Vivek Garipalli account details a Fortune 20 CEO who commissioned $1 billion in AI-driven opex savings — the team spent $200M on tokens and delivered modest customer service improvements and slightly reduced engineering hiring. Token spending has since been ordered "dramatically slashed." A separate Axios-sourced anecdote reports one company spending $500 million in a single month on Claude after failing to set employee usage limits — a figure Gary Marcus and Ethan Mollick each publicly called implausible at any reasonable headcount.
On the supply side, H200 GPU rental prices crashed from $7/hour to $4/hour in three weeks, a 43% collapse signalling softening infrastructure demand. The FT's Panmure Liberum analysis frames the structural picture: modelling hyperscaler AI investment against projected 2025–2030 revenue under zero-cost assumptions, only Amazon clears positive (+7.2%). Microsoft sits at –9.2%, Alphabet at –15.7%, Meta at –28.8%, Oracle at –35.6%. Real returns — factoring GPU depreciation, power costs, and headcount — are worse. Stocks remain at all-time highs.
The YouTube batch adds a second independent data point: Lev Selector's weekly AI roundup notes Microsoft is already redirecting engineers from direct Claude Code licenses to GitHub Copilot CLI specifically to manage spend, while Goldman Sachs projects a 24× surge in total token consumption as agent workloads scale — meaning per-token price cuts will not prevent total bill growth.
Strategic Take
The question is no longer whether AI produces results — it often does — but whether it produces them fast enough to justify current token economics. Builders should implement hard usage caps, per-agent model routing (frontier for planning, cheap models for execution), and explicit ROI gates before workloads scale. Uber's lesson: budget exhausted, returns modest, executive patience gone.
