The pitch behind replacing engineers with AI coding agents has always been straightforward. Salaries are expensive, software is cheap, and the math eventually works in management’s favour. A series of new disclosures this week suggests the math, for now, is doing nothing of the kind. Bryan Catanzaro, vice president of applied deep learning at Nvidia, told Axios that for his team “the cost of compute is far beyond the costs of the employees.” That is a remarkable sentence from inside the company selling the picks and shovels.
The mechanism is simple enough. Most AI systems bill by token, the chunks of text that pass through a large language model as it reads, reasons and writes. A single developer asking a coding assistant to refactor a file costs very little. A team of developers running fleets of autonomous agents in the background, each one chewing through codebases, debugging in parallel and reasoning out loud at length, is something else entirely. According to The New York Times, the heaviest individual users now run monthly bills above $150,000. One Stockholm based engineer, Max Linder, told the paper he probably spends more on Claude than he earns.
The corporate version of that confession appeared in a report from The Information this week: Uber’s engineers, using Anthropic’s Claude Code, have already burned through the company’s entire AI budget for 2026. The year, for the record, is not yet half over. Developers inside the industry have a word for the behaviour driving these bills. They call it tokenmaxxing: deliberately consuming as much compute as you can extract from your employer’s account, on the theory that more tokens equals more productivity, or at least more interesting experiments.
Some of this is enthusiastic adoption of genuinely useful tools. Boris Cherny, who runs Claude Code at Anthropic, claimed earlier this year that roughly all of Anthropic’s own code is now AI generated. Microsoft and Google executives have said AI now contributes about a quarter of their code output. Meta has reportedly begun factoring AI usage into performance reviews. The story enterprises were sold, that AI would let smaller teams ship faster, is at least partly real.
The story they were not sold is that token costs scale with how aggressively you use the tools, and the tools reward aggression. Agentic workflows multiply usage by design: an agent that plans, then writes, then tests, then explains its reasoning consumes many times more tokens than a chat assistant answering a single question. Anthropic has raised prices on some services this year. Microsoft has shifted GitHub Copilot from request based to usage based billing. Gartner’s latest forecast pegs worldwide AI spending in 2026 at $2.59 trillion, with model consumption alone growing 110 percent year on year. That is a polite way of saying the meter is running faster.
The pattern this creates is awkward for the AI replacement narrative. A McKinsey style productivity argument that compares one engineer’s salary against one AI subscription is no longer the right comparison. The right comparison is one engineer’s salary against a fleet of agents that engineer can now run in parallel, plus the time spent reviewing what those agents produced, minus the hallucinations and security issues they introduced. Several recent studies have warned that forcing AI use can increase complexity rather than reduce it, particularly when humans end up validating outputs they did not ask for.
None of this means AI agents will stop being adopted. Compute prices have historically fallen, and there is no reason to think they will not fall again as inference efficiency improves and competition between Anthropic, OpenAI and the chip vendors intensifies. Jensen Huang has even floated the idea of allocating engineers a personal token budget worth around half their base salary as a recruiting perk, which is itself a sign of how the labour math is being redrawn. But the comfortable assumption that AI is automatically cheaper than the workers it displaces has now been disproved in public by some of the companies most loudly pushing the technology. The cost of intelligence, at scale, is still a real number. For 2026 it is bigger than the salary it replaced.