Key Takeaways

  • Enterprise AI adoption has taken off, but companies like Uber, Meta, and AWS are grappling with massive, often unproven, spending on AI. Builders need to prove ROI fast.
  • "Token maxing" is a major issue: employees, incentivized by internal dashboards, are reportedly leaving AI tasks running overnight without clear productivity gains, driving up costs.
  • This behavior is a textbook case of Goodhart's Law: when AI spend or token count becomes a target, it ceases to be a good measure of actual value or productivity.
  • Even as the cost per AI task drops (the bull case), Jevons Paradox suggests overall consumption could skyrocket, negating expected savings and keeping total bills high.
  • Corporate America is now rationing AI, forcing a hard look at immediate costs rather than relying solely on long-term cost deflation predictions.

The Scramble for AI ROI

Everyone wants AI, and the adoption curve for enterprise AI has been steep. John Coogan, one of the hosts, observed, “There's been a complete fast takeoff in enterprise AI adoption.” But this rapid deployment comes with a nasty hangover: no one's sure what they're actually getting for their money. Companies as big as Uber, Meta, and AWS are reportedly wrestling with vast new budgets dedicated to AI, and the immediate return on investment (ROI) remains fuzzy. Executives are scrambling to track these returns, as Coogan put it, “as the bill for massive computing needs comes due.” It's a classic founder trap: chasing the hot new thing without first defining the measurable impact it needs to deliver.

The Toxic Incentives of "Token Maxing"

Beyond the raw cost, there's a more insidious problem: "token maxing." This is when employees, perhaps chasing internal metrics or just experimenting without guardrails, consume AI resources without generating real value. Coogan shared a stark example: "Token maxing dashboards have been reportedly led to potentially ROI negative AI use in at name brand companies like Meta... people leaving things running overnight that weren't necessarily productive just because they want to rank up on the dashboard."

This isn't just waste; it's a direct consequence of Goodhart's Law: “When a measure becomes a target, it ceases to be a good measure.” If your internal systems reward AI usage or token count, regardless of outcome, you're practically inviting your teams to run up the bill for no reason. For a founder, this highlights the danger of tracking vanity metrics that don't directly tie to business value.

Jevons Paradox and the Deflationary Trap

The bull case for AI costs is compelling: the cost per completed task will fall dramatically over time. Coogan noted, “The bull case is that the cost per task completed by AI will decrease very quickly. So even if you're spending half a billion per month on AI or tokens today, uh the same output will be available next year maybe for a tenth of the cost.” While individual task costs drop, there's a lurking danger: Jevons Paradox. This economic principle suggests that as a resource becomes more efficient or cheaper to use, total consumption can actually increase, sometimes drastically.

Think about it: if AI becomes incredibly cheap, your team might find endless new, slightly useful (or even useless) ways to apply it, pushing total spend higher, not lower. So even with deflationary unit costs, your overall AI bill could still skyrocket, leading to corporate rationing and a renewed focus on immediate, measurable returns.

What to Do With This

Stop tracking AI spend as a standalone metric. Immediately audit your AI usage by individual team or project, and for each, define the specific business outcome (e.g., higher conversion, faster code deployment, fewer support tickets) it must impact. Then, implement real-time dashboards that show cost per outcome achieved, not just total tokens or dollars spent. If a project's cost per outcome isn't improving week-over-week, kill it, or force your team to justify its existence with a clear path to value.