Goldman Sachs projects agentic AI could increase token demand by 24 times in the next few years, forcing major enterprises like Microsoft and Uber to restructure AI spending as costs outpace tangible returns.
Token costs strain enterprise budgets
Uber’s operations chief, Andrew Macdonald, told Business Insider the company blew through its entire 2026 AI budget in months, with no clear correlation between token usage and consumer feature development. Over 80% of Uber’s software engineers use agentic AI, and more than 60% of code is AI-generated, yet executives cannot link higher token consumption to measurable product improvements.
Microsoft is revoking developer access to Claude Code, moving teams to its internal Copilot CLI tool by June 30. The shift, coinciding with the end of Microsoft’s fiscal year, is widely seen as a cost-cutting measure. Microsoft also switched GitHub Copilot to token-based billing after operational costs ballooned.
Agentic AI multiplies token consumption
Agentic AI workflows can consume over 1,000 times the tokens of a single chatbot interaction. Goldman Sachs estimates this class of AI will drive a 24-fold increase in token use within a few years, creating a growing gap between AI ambitions and budget realities.
OpenAI employee Peter Steinberger reported his three-person team spent $1.3 million on tokens in a single month running agentic tools. Such costs raise questions about the economic logic of replacing human workers with AI, especially when token bills exceed salaries.
Hardware gains cannot close the gap soon enough
Nvidia’s upcoming Vera Rubin platform promises up to 10 times the performance per watt over current Blackwell designs. However, over 50% of data center projects announced for Blackwell have been canceled or delayed, and Google, Oracle, and Microsoft now plan to run hardware for six years rather than upgrade annually.
Even with next-generation inferencing chips, the pace of token demand growth outstrips near-term hardware efficiency improvements. Token costs are falling, but not fast enough to offset the explosion in agentic AI usage.
The path forward
Enterprises face a fundamental tension: AI adoption is accelerating, but the cost of token-based billing is rising faster than the value it delivers. If companies like Microsoft and Uber cannot sustainably scale AI without budget overruns, smaller organizations will struggle even more. The industry’s ability to align AI spending with measurable business outcomes—not just code volume—will determine whether the current investment cycle yields profit or correction.
