Allora (@AlloraNetwork): Goldman Sachs is betting that AI agents will turn tokens into the new fuel: token consumption driven by agents could swell 24x by 2030.

Goldman Sachs is betting that AI agents will turn tokens into the new fuel: token consumption driven by agents could swell 24x by 2030. And agents are quickly becoming the first real “can we afford this?” stress test of the AI boom. This week’s reporting suggests even Uber and Microsoft are stepping back to reassess how pricey heavy agent usage can get. A chatbot gives you a single reply. An agent does much more: it sketches a plan, invokes tools, validates outputs, patches errors, and runs the cycle again—sometimes multiple times. That iterative loop is where costs explode. One “simple” request can burn 10x, 50x, or orders of magnitude more tokens than a standard response. In Goldman’s optimistic scenario, monthly token demand climbs to roughly 120 quadrillion 2030, while inference prices per token keep dropping 60%–70% annually. So the battleground is set: do agents deliver enough productivity to justify the extra tokens, or are they just an expensive way to spin cycles? Microsoft’s recent move adds color. Earlier this month it started pulling developer access to Claude Code, aiming to shift users to its own Copilot Command Line Interface by June 30. Officially it’s “tool consolidation,” but landing at the end of the fiscal year makes it hard not to read it as a cost-control maneuver too.