The Bill arrives for AI

SYNOPSIS

The fact is obvious that the age of unlimited AI could not continue forever. The illusion of plenty was, in reality, a subsidy, with AI firms internalizing cost to promote usage, companies giving out usage as a perk, and engineers striving to rise to the top of internal leader boards through the use of compute instead of discernment. But here's the twist: Constraint spurs good engineering. In every era, engineers build better products when they need to optimize something like speed, memory, or bandwidth. Tokens are only the latest optimization problem on this list. Once the actual costs of compute became clear, engineers stopped asking themselves whether they could do something and began asking whether they should. It turns out this was the crucial question to ask all along.

For two years, AI felt infinite. Tokens were plentiful, usage was encouraged, and waste was ignored. The age of infinite AI is over and engineering is the better for it.

Artificial Intelligence has felt unlimited for the last couple of years. Tokens used to be very cheap, often subsidised, and the unambiguous message from every AI company was: use more, build more, don't worry about the economy. That era is subtly ending.

A major tech firm spent all its AI budget for 2026 early in the year. It was not a seismic autonomous system that has gone rogue. Internal leaderboards were established to recognize engineers for their AI usage. In what will not come as any surprise, the budget for the Afterburner Project has been massively exceeded. The Afterburner project allows engineers to burn native tokens to put themselves on a scoreboard to receive R&D funding or job offers. The company's CTO called it "back to the drawing board", something that did begin quietly beginning talks with a competitor's AI product not long after, meaning it was just as much a negotiating tactic.

Another big software company's account is contrasting but just as instructive. This enabled teams to carry out a comparison between the two coding tools. When the exercise was completed, it pulled the licenses at the timing of the financial year end. The engineers were not pleased, but the decision was never about the tools.

Beneath these headlines a structural pricing shift is actually happening. Leading platforms are shifting from flat-rate to usage-based billing. Companies that sell AI services that will go public, will no longer be able to subsidise customer usage to grow. The cost of AI will rise significantly before it becomes more affordable.

This is positive news for all. The main purpose of engineering is optimizing subject to constraints. Engineers have always developed superior systems under constraints with either speed, latency, memory or bandwidth. Tokens are just the next constraint to be added to that list. The leaderboard era produced waste. Teams will make deliberate decisions about when to deploy agents, when to write code by hand and what the tradeoff actually is when compute has a real cost. The age of metering will bring judgment.

Making this trade is worthwhile.

Aaradhya Gupta