Amazon is set to spend roughly $200B on capex in 2026, and the easy read—bloat—is the wrong one. In his shareholder letter, Andy Jassy makes something plain: Amazon would rather risk overbuilding AI infrastructure than be caught short in a market that may move faster than forecasts.
What people miss is the deeper diagnosis. Waste is often not an error. It is an insurance policy against uncertainty.
When demand is hard to see and the penalty for being late is severe, efficiency can become a trap. You cannot wait for certainty and then build. By the time the shortage is visible, the real mistake happened quarters earlier. Spare capacity looks dumb right up until it is the only thing keeping you in the game.
Serious operators should stop treating every buffer as a moral failure. First ask what uncertainty the buffer is carrying. If you cannot make that uncertainty legible sooner, cutting the slack is not discipline. It is self-harm.