Generated by Codex with GPT-5

TBPN surfaced this May 29, 2026 post, and the original is From Tokenmaxxing to ROImaxxing.

The important shift in the piece is easy to miss because the surrounding AI news is still enormous. Anthropic has passed \$47B in annual recurring revenue, raised a \$65B Series H at a \$965B post-money valuation, and released Claude Opus 4.8. Large companies are rolling AI tools out widely. Usage is no longer a niche experiment run by a few curious engineers.

But that success creates a harder question: what exactly are companies buying with all those tokens?

TBPN argues that enterprise AI adoption is moving from “tokenmaxxing” to “ROImaxxing.” The first phase rewarded visible usage. Teams built dashboards, celebrated heavy users, and treated rising token consumption as evidence that the organization was learning how to work with AI. The next phase will be less forgiving. Executives, finance teams, and investors will want to know whether the spending improves products, reduces cost, increases revenue, or lets the company accomplish something materially different.

When The Metric Becomes The Goal

The problem with tokenmaxxing is not that companies are spending heavily during a technology transition. Early adoption can be messy, and some short-term waste is a reasonable price for learning. The problem begins when a proxy for useful work becomes the objective.

TBPN points to reports of potentially return-negative AI usage at companies such as Meta, Uber, and AWS. Once employees know that model usage is visible, an internal leaderboard can quietly distort behavior. Running more agents, using more expensive models, or consuming more tokens starts to look like ambition even when the underlying task is low value. A dashboard intended to encourage experimentation can become a machine for producing activity without a clear business result.

The post names two useful ways to think about this. Goodhart’s Law says that when a measure becomes a target, it stops being a good measure. Jevons Paradox says that when a resource becomes more efficient to use, people often consume more of it rather than less. Both apply to AI tools. Model costs may fall sharply, but cheaper inference does not automatically reduce a company’s bill. Lower prices can encourage more use, more agent loops, and more automation attempts. If the organization is rewarding consumption, efficiency can expand the waste as easily as the value.

The ROI Question Moves Up The Org Chart

The post’s sharpest example comes from Match Group CEO Spencer Rascoff. TBPN highlights his explanation that the company is spending \$5-10M a year on AI tools from a starting point near zero and is slowing hiring somewhat to fund the expense. His broader point is even more revealing: companies may believe they are benefiting from AI while still struggling to feel or measure the payoff.

That is a different conversation from the one AI vendors were having a year earlier. The question is no longer only whether employees adopt the tools or whether a coding agent can complete an impressive demo. The question is what changes in the operating model. Does a support team resolve more cases without hurting satisfaction? Does an engineering group ship better software with the same headcount? Does a sales team close more business? Does a back-office workflow become cheaper and more reliable?

Token counts do not answer those questions. Neither do anecdotes about a handful of power users. A useful internal measurement system has to connect AI spending to outcomes, while accounting for quality, review time, errors, rework, and opportunity cost. A coding agent that produces more pull requests is not automatically valuable if senior engineers spend more time reviewing weak changes or if incidents increase.

Spending Can Still Be Rational

TBPN does not argue for a blunt clampdown. The bull case remains plausible: the cost per completed task could fall quickly as models improve, tooling matures, and companies learn which workflows deserve automation. A company that spends aggressively today may be buying institutional knowledge that becomes a real advantage when the economics improve.

That makes the management problem more subtle than cost cutting. Companies need room to experiment without confusing experimentation with proof. They need to distinguish useful learning from performative usage, and high-leverage automation from endless loops pointed at make-work. Some uneven return is normal when a new tool changes how work gets done. A permanent inability to explain the return is not.

The next earnings cycle should make this more visible. If AI expenses materially increase operating costs or reduce net income, analysts will ask where the benefit appears. Companies with a credible answer will be able to point to revenue, margins, product velocity, hiring plans, or customer outcomes. Companies without one will have a usage story when the market is asking for an economics story.

Takeaway

The most useful idea in TBPN’s post is that AI adoption is entering its accountability phase.

Tokenmaxxing made sense as an early cultural push: get tools into employees’ hands, make usage normal, and learn quickly. It becomes dangerous when the proxy survives after the real question has changed. The goal is not to consume more intelligence. It is to produce better business outcomes with the right amount of it.

That requires a more mature dashboard. Track spending, but connect it to tasks completed, quality, cycle time, customer impact, and staffing decisions. Encourage experimentation, but do not reward empty consumption. Assume inference gets cheaper, but plan for total usage to grow. The companies that navigate this well will not be the ones with the biggest token numbers. They will be the ones that can explain what the tokens changed.