This is a phrase which Andrew Tindall uses in this Drum piece to describe the pricing shift already underway in language models. Most consumers are still insulated from a change which is currently directed at enterprise customers and ‘power users’ of Claude:
Microsoft’s decision to wind down Claude Code licens in parts of the business is the canary in the blood-diamond mine. A company with enough infrastructure and cloud power to make rain nervous is pushing engineers away from a popular coding tool and toward its own stuff. The official line is convergence, but you cannot ignore that the news lands on Microsoft’s fiscal year-end. Meanwhile, GitHub is moving Copilot itself to usage-based billing. Again, presumably after getting jealous of how much cash Claude has been raking in.
What all AI leaders tried to avoid is the cold reality of the models we have fired people for and plugged into every work process: they are real, useful and expensive AF. It is a much trickier reflection on AI than “AI is fake and useless.” Fake and useless get adopted slowly. Useful and expensive get governed and rationed. Terrible for profits. Meanwhile, every head of tech has independently selected their favorite AI provider, usually with zero business case or P&L analysis.
It’s something that Justin Pickard has just published about here, based on a contribution to our workshop during the summer. I love this description in particular of the development work that might be prompted by this sudden shift into token scarcity:
By 2027, a second layer had grown around model use. Balance widgets sat
in browser windows; cache timers counted down beside reset timers. ‘Ask Later’
buttons queued prompts, like pet feeders for work the user did not trust themselves to start at the right time. Spreadsheets recorded token burn, task type, cheaper windows. All of it read from figures the platform had not thought to round. The chat window stayed where it was. The counters multiplied around it.4
Mark Carrigan 