Tech giants spent years telling staff to lean on AI to up productivity. Now those same firms are running into an ugly, expensive surprise: neural-net bills have leapt — fast and far beyond initial estimates.
The Verge says Microsoft is actively shifting engineers away from Claude Code toward its own Copilot CLI. Official explanations cite a move to in-house tooling; insiders, however, point to a simpler culprit — Claude’s token costs become prohibitive at scale (i.e., when usage ramps up across teams).
Fortune notes similar headaches at other big firms. Agent-style AI is especially brutal on budgets: these agents don’t answer a single prompt and stop. They chain steps, call things back and forth, and in practice can eat hundreds to thousands of times more tokens than a plain chatbot request (e.g., automated workflows, multi-step data pulls, etc.).
A striking data point: OpenClaw developer Peter Steinberger reported his team burned through more than $1.3 million on tokens in a single month. Yes — the number feels surreal.
The industry has even coined a term for part of the problem: "tokenmaxing." Staffers sometimes route trivial or vanity tasks through models just to bump internal metrics. Conversations about such excesses are happening at Amazon, Microsoft, and other IT giants — companies that remain among the largest backers of AI development.