Chapter two: Lock-in and cost drift

In the last week of July 2026, an internal review at Amazon surfaced a number that should be pinned above every chief financial officer's desk. A single coding task, handed to Claude and left to run, had consumed 1.8 million dollars of model usage, 860 per cent over its budget. The overrun ran for five months before anyone noticed, because eighteen months earlier the whole job would have cost less than a laptop.

The same week, Accenture's finance function disclosed that the firm now runs 8.7 trillion tokens of model usage every week, and that one internal tool's consumption had risen 113-fold in ten weeks. Nineteen per cent of users were driving 80 per cent of the spend. This is not a firm losing control of a cost line. This is a firm discovering that it has acquired a new class of cost line, one that behaves like energy, not like software.

And here is the detail that completes the picture. Token prices fell all through that same week. Anthropic launched Opus 5 at half the price of its flagship. OpenAI cut GPT-5.6. The per-unit price of intelligence dropped, and the bills went up.

Chapter one argued that compute is a production input, and that the strategic prize is legible cost rather than low cost. This chapter is about what happens at the seam where that input enters your business, because two failure modes live there, and they arrive together.