Uber burned through its 2026 AI coding budget in four months. Microsoft canceled most of its Claude Code licenses six months after rolling t…

Uber burned through its 2026 AI coding budget in four months. Microsoft canceled most of its Claude Code licenses six months after rolling them out. The mechanics are simple: per-token cost keeps falling, consumption explodes faster. Cheap tokens, giant bill. The reaction from boardrooms is almost always the same: we know we pushed you to use tools, even tokenmaxxing god forbid, but it's time to show how these are useful. That's actually the right question. Most orgs just can't answer it. Not because the tools aren't valuable. Because the layer that would tell you which tools produced which value was never built. For decades, seat-based software was predictable. You buy 500 licenses, you pay for 500 licenses. Finance understood it. Procurement budgeted it. And attributing outcomes was crude but doable: "this team adopted the tool, this team didn't, look at the delta." Agentic AI doesn't fit that model. One developer can trigger 50 model calls before lunch. An agent running in the background can consume tokens continuously, with no human initiating anything. The cost isn't linear or predictable, and neither is the output. Which is why "show us it's useful" is so hard to answer. Which agent workflows produced which merged PRs? Which team's tokens correlated with fewer production incidents? Which invocations were catching real bugs versus just producing noise? Without a governance layer that tracks all of this, none of those questions have real answers. You get vibes on both sides: "it feels productive" from the users, "it feels expensive" from the CFO. We have a generation governance layer: PR reviews, CI/CD, code ownership. It took the industry 20 years to build and normalize. Nobody really questioned whether we needed it. We just built it, slowly, because the consequences of skipping it were obvious. We don't have an equivalent layer for AI usage and value attribution. Not because nobody could build it. Because we deployed agents before we thought we needed to measure them. Uber didn't have a cost problem. They had an attribution problem that surfaced as a cost problem. The boardroom asking for ROI is asking the right question. It's just landing on teams that don't yet have the infrastructure to answer it. The orgs that get this right won't be the ones that find cheaper models. They'll be the ones that build the governance and attribution layer first, and then have the value story ready when the board asks.

Source: Itamar Friedman (X) | 2026-08-06

Related