The Week the Machine Cost as Much as the Worker

Andy Jassy published his annual shareholder letter and the confession is on page three. Two hundred billion dollars in capital expenditure for 2026. Not a projection. A signed check. AWS's custom chip business — Graviton, Trainium, Nitro — already generates over twenty billion annually, growing at triple-digit percentages. Jassy said it with the calm of a man whose numbers agree with him: "We're not investing two hundred billion in capex in 2026 on a hunch. We have customer commitments for a substantial portion of it." OpenAI committed one hundred billion in AWS services over eight years. Amazon's free cash flow dropped from thirty-eight billion to eleven billion. In Mexico City's Doctores neighborhood, they call this a mortgage. In Seattle, they call it strategy.

Jason Calacanis — investor, All In podcast host, a man not easily rattled — discovered that one AI agent running on the Claude API was costing him three hundred dollars a day. One hundred thousand a year. The loaded cost of a junior employee. The agent did not take vacations, did not call in sick, did not quit. It also did not replace a full employee. It performed a fraction of the work at full price. Chamath Palihapitiya, his partner, now imposes token budgets per developer: "An agent has to be at least twice as productive as another employee." In November 2025, two LangChain agents spent eleven days messaging each other. Nobody noticed. The bill came to forty-seven thousand dollars. Gartner predicts over forty percent of agent projects will be canceled before reaching production by 2027. The file grows thicker.

But the machine does work when someone watches it. Tim Ryan, Citigroup's Head of Technology and Business Enablement, reported that account-opening time in the services division dropped from one hour fifteen minutes to fifteen minutes. An eighty percent reduction. One hundred eighty-two thousand employees across eighty-four countries now use AI tools. Six and a half million prompts logged. Citi still operates under two 2020 Federal Reserve consent orders demanding enhanced risk management. The machine did not fix that. But the procedure that once took a lunch break now takes a coffee. The savings are real. The question is for whom.

Alex Tabarrok, economist at George Mason University, published in Fortune an equation that should be pinned to every HR office on the continent. Sixty percent of people employed with forty percent unemployed produces the same working hours as one hundred percent employed at sixty percent of the hours. "Imagine I told you AI was going to create forty percent unemployment. Sounds catastrophic. Now imagine I told you AI was going to create a three-day work week. Sounds great. It's the same number." The difference is political, not arithmetic. Who keeps the hours and who loses the job depends on a single variable: who controls the distribution. Between 1870 and today, working hours fell forty percent. Nobody remembers it as a crisis. But executives surveyed by Boston Consulting Group already report "AI brain fry" — mental fatigue from supervising multiple tools. The restaurant owner eats well. The waiters are hungry.

Eighty percent of white-collar workers do not use the AI tools their company purchased. The figure comes from WalkMe, an SAP subsidiary, in its fifth annual State of Digital Adoption report: 3,750 executives and employees across fourteen countries. Fifty-four percent bypassed them in the last thirty days. Thirty-three percent never used them at all. Only nine percent trust AI for complex decisions. Executives think things are fine: sixty-one percent trust AI, eighty-eight percent say employees have adequate tools. The employees disagree: twenty-one percent. The gap is sixty-seven points. Average digital transformation budgets rose thirty-eight percent to $54.2 million annually, and forty percent of that spending underperforms. Dan Adika, WalkMe's CEO, put it simply: "They have pride in what they do. They won't let some AI bot take over." The rebellion has no banners. It has passwords that never get used.