Do You Even Tokenmaxx, Bro?

“Looksmaxxing is…the idea of maximizing physical attractiveness by any means in order to ascend.”

Yeah, I'm out there. I walk into boardrooms looking like Temu Sheldon Cooper and I am the most dangerous person in the room because I have a credit card, investors who don't fully understand what I'm doing, and a pathological need to use every single token available to me. I'm not here to be efficient. Efficiency is a middle-manager's religion. I am here to flood the context, run the reps, and become so deeply embedded in your tech stack that removing me would require a board vote. The model doesn't know what it's capable of until someone truly unhinged gets hold of it. That someone is me. I am tokenmaxxing to the top, and I will see you there, or I will see you in my context window.

—The Tokenizer

For the uninitiated, a token is a small piece of information, in text roughly three-quarters of a word, that a large language model processes as it reads and generates language. As such, they are the measure for the work (and billing!) that an AI model does. Tokens measure work, but not exactly. For example, you might estimate the cost of an article in this outlet by word count for us as authors, but you would not do the same for, say, Mya Angelou. Like so many metrics, Tokens are used because they are easily and agreeably quantifiable; we can convince ourselves that we know how to measure AI.

"What does the money machine eat? It eats youth, spontaneity, life, beauty and above all it eats creativity. It eats quality and shits out quantity." – William S. Burroughs

Similar leaderboards have reportedly appeared at Microsoft, Salesforce, and Shopify[ii]. It seems the industry is competing to prove AI adoption by the most expensive and least meaningful measure available. By itself, it is hard to know if this is ‘a lot’ or ‘a little.’  {Note – I started to convert this to the number of human workers effort this burn of tokens took, and it was depressing. I stopped.

For those of us who lived it, this feels like an AI version of the market inflation that happened in the earlier part of this decade, where mid-level leaders were incentivized to increase their headcount in order to show the ‘value’ of their business unit. The difference here is that the headcount has now become virtual; while it is tempting to try to convert the number of tokens and energy into an equivalent number of human workers, you quickly realize that human brains just don’t work like that. Left unchecked, this is contributing to the accelerating de-humanization of the workforce[iii], and will lead to disaster.

Cassie Kozyrkov, a prominent decision intelligence researcher, has a name for this: tokenmaxxing. And she has a law for why it was always going to go wrong: Goodhart's Law — "When a measure becomes a target, it ceases to be a good measure."[iv] Tell people to maximize tokens, and they will maximize tokens. They will not, necessarily, produce anything useful.

Some organizations have moved toward a better unit of measurement: the Agentic Work Unit (AWU), a discrete task completed by an AI agent. Outputs rather than inputs — a genuine improvement. But even AWUs are no cure-all. An agent pointed at the wrong task produces AWUs of zero value. The direction of the work matters far more than its volume. As Kozyrkov puts it: "If you're going in the wrong direction, sloth is a virtue."

Ironically, perhaps, Kozyrkov’s sentiment is more-or-less reflected in the Defense Establishment’s policies, specifically the fact that software releases typically lag industry by approximately a year. The idea is that weaknesses in software can be discovered – as they frequently are – in the commercial sector before being exposed in the Defense enclave. From a fundamental science point of view, we have little understanding of the internals of large-scale AI models, and the more sophisticated the model, the less control we have of it.

Token consumption is no longer merely a Silicon Valley vanity metric. It has arrived in the halls of the Department of Defense.

During Operation Epic Fury — the 38-day air campaign against Iran — the Pentagon's Maven Smart System, built by Palantir, was used to help plan and coordinate some 13,000 airstrikes across the entire theater of operations. By any standard, this represents a remarkable operational deployment of AI at a historic scale. Pentagon Chief Digital and AI Officer Cameron Stanley described it at a recent conference, noting that AI tools "allow us to take all of this data, synthesize the data, and make better decisions, faster, on the battlefield."[v] Classified usage of Maven surged 89 percent month-over-month. Unclassified usage rose 38 percent. And at peak, DoD users burned through approximately 20 billion tokens per day — a 44x increase measured in those terms.

Stanley described the military's appetite for AI as "insatiable."  What we do not know at this time is whether these tokens are actually making the air campaign better. This is an assignment for future historians.

Insatiable is a difficult word because it is not, as a rule, a virtue. They are symptoms of excitement, of novelty, of the intoxicating feeling of capability without yet having developed the discipline to apply it wisely. As has been considered here[vi], the Pentagon is right to embrace AI for intelligence synthesis, target analysis, and decision support. The technology's potential in those domains is real and significant. But "insatiable" token consumption is not a measure of decision superiority. It may, in fact, be a measure of inefficiency. Inefficiency of systems and operators still learning what questions are worth asking.

To his credit, Stanley himself seems to understand this. He was careful to describe the goal not as maximum AI throughput, but as "the best human-machine team possible." Human analysts, he noted, must "apply operational art, context, and legal intent" — dimensions no model can supply. The tokens are the scaffolding. Human judgment is the building.

The tragedy is that the metric being reported — and celebrated — is the scaffolding.

The tokenmaxxing problem is not new. It is a specific instance of one of the oldest failure modes in organizational management: choosing a proxy for the thing you actually want, then optimizing the proxy until it bears no relationship to the thing.

What to Measure Instead

Tokens measure what vendors charge you for. AWUs measure what agents do. Neither measures what you actually want: decision quality.

Decision quality is harder to measure. It requires defining what a good outcome looks like before you optimize for it. It requires resisting the temptation to substitute a legible proxy for an ambiguous truth. It requires leadership willing to have the harder conversation — the one about results, not inputs.

Sensible leaders measure results. They ask: did the decision we made, with the data and tools available, produce an outcome that justifies the cost — in dollars, in compute, in risk? That question cannot be answered by a leaderboard. It can only be answered by people willing to look honestly at what happened, and why.

The Pentagon CDAO says the right things about human-machine teaming and the irreducible role of human judgment. The question is whether the metrics being tracked — and celebrated — reflect that wisdom or undermine it.

The answer, of course, is 42.

Nick Ulmer is a professional OR analyst and President of the Military Operations Research Society. A retired Navy Supply Corps Officer, his professional interests are at the nexus of supply chains, optimization, and policy.

Harrison Schramm is a professional statistician and CTO of Attritable Machines. A retired Navy Pilot, he focuses on building the smallest, most useful AI. He is a Fellow of MORS.

The authors declare no conflicts of interest. They additionally declare no tokens were used in the development of this article. NEED HELP Quitting? We recommend chewing gum or just getting outside.

[i] https://fortune.com/2026/04/09/meta-killed-employee-ai-token-dashboard/

[ii] https://blog.pragmaticengineer.com/the-pulse-tokenmaxxing-as-a-weird-new-trend/

[iii] https://www.realcleardefense.com/articles/2026/05/15/ai_at_speed_humans_at_risk_1182838.html

[iv] https://open.substack.com/pub/decision/p/hold-my-beer-im-tokenmaxxing?r=3twa9a&utm_campaign=post-expanded-share&utm_medium=web

[v] https://breakingdefense.com/2026/05/insatiable-appetite-for-ai-maven-usage-surged-for-strikes-on-iran-pentagon-ai-chief-says/

[vi] https://cimsec.org/the-battle-to-preserve-the-global-order-part-1/