The tokenmaxxing party is over, and the CFOs want to see receipts.
The Summary
- Uber COO Andrew Macdonald told the market he can't draw a line between AI token spending and measurable improvements, kicking off a broader conversation about AI ROI
- Sam Altman says the question he hears most from enterprise customers now is "Where is the revenue?", a stark shift from the tokenmaxxing era when companies raced to max out AI budgets
- CIOs are hitting budget ceilings with nothing tangible to show finance teams, forcing a reckoning between AI hype and actual business outcomes
The Signal
The bill for the agent economy is coming due, and companies are squinting at the line items. Uber's COO admitted publicly what many enterprise leaders are whispering in budget meetings: they're spending millions on AI tokens and seeing productivity gains that feel more like rounding errors than transformation. This matters because Uber isn't some cautious laggard. They deployed agents early. They leaned in. If they can't make the math work, most companies won't either.
The irony is sharp. Just months ago, "tokenmaxxing" was a badge of honor. Companies posted leaderboards showing how many tokens their agents burned through, treating API costs like a scoreboard for innovation. The viral moment has passed, but the costs kept climbing. Most enterprises never adopted the public leaderboards, but they absorbed the underlying assumption: spend now, justify later. That bill of goods is expiring.
"The question CIOs are asking isn't whether AI works. It's whether it works at this price point."
OpenAI's Sam Altman is hearing it directly. When he talks to companies, the conversation has shifted from capabilities to cash flow. "Where is the revenue?" isn't a hypothetical. It's the question finance teams are asking IT leaders in quarterly reviews. Google's Sundar Pichai is hearing the same feedback loop. The Big Tech CEOs who built the models are now getting real-time market research on whether enterprises can actually afford to use them at scale.
This is the gap between demo and deployment, between proof of concept and profit and loss. Agents can write code, answer support tickets, and summarize meetings. What they can't do yet is show up on a P&L in a way that makes CFOs nod instead of wince. The problem isn't that AI doesn't work. The problem is that it works just well enough to keep spending, but not well enough to justify the spend.
Key factors driving the ROI crunch:
- Token costs scale linearly with usage, but productivity gains don't
- Most agent deployments are still in "assistant" mode, not "replacement" mode
- The gap between what AI can do in a demo and what it can do in production remains wide
The Implication
Expect a wave of AI budget cuts disguised as "optimization." Companies that went all-in on agents without clear KPIs will quietly throttle usage and rebrand the pivot as strategic discipline. The survivors will be the ones who treat AI like capital equipment, not magic, with hard ROI thresholds and kill switches for underperforming deployments.
For builders in the agent space, this is the filter. Products that reduce token costs, prove ROI in weeks instead of quarters, or deliver step-function improvements in specific workflows will separate from the pack. The era of "AI for AI's sake" spending is closing. What opens next is harder to fundraise for but easier to sell: tools that actually pay for themselves.