Packy McCormick arrives at a conclusion that cuts through the current AI hype cycle with surgical precision: the industry's obsession with spending more is actually a delusion that masks a fundamental lack of value creation. By reframing the conversation from "how many tokens can we burn?" to "what return are we getting?", McCormick exposes a market failure where Fortune 500 companies are burning cash on digital labor that often produces nothing but expensive dashboards. This is not just a critique of software; it is a warning about how easily capital allocation can be hijacked by vanity metrics, echoing the dot-com era's obsession with "eyeballs" rather than revenue.
The Delusion of Tokenmaxxing
McCormick introduces his co-author, Markie Wagner, a founder who saw through the industry's noise long before it was fashionable to do so. He notes that while everyone else was chasing the latest leaderboard, Wagner heard directly from CEOs who were baffled by their own spending. "We committed to all this token spend and I have no idea what we're getting out of it," McCormick quotes these executives as saying, capturing the growing anxiety in boardrooms. The author argues that the market has created a perverse incentive structure where spending is mistaken for progress.
"Tokenmaxxing - literally maximizing the amount of tokens you or your organization spends... was a mass delusion, something like a commercial form of AI psychosis."
This framing is powerful because it strips away the technical jargon to reveal the human behavior at play: a collective fear of missing out that drives irrational consumption. McCormick describes how companies like KPMG encouraged employees to use AI agents with "no spending limit," effectively treating them as digital staff with unlimited credit cards. The result, he argues, was not innovation but a race to the bottom in terms of efficiency. He points out that when skeptics asked for proof of utility, they were met with the dismissive retort: "Skill Issue." This cultural dynamic mirrors the early days of cryptocurrency, where speculative fervor often drowned out questions about actual utility or underlying value.
"The market incentivized companies to spend tokens, so boards incentivized leaders to spend tokens, so leaders incentivized managers to spend tokens... Nobody had an incentive to say that the tokens aren't doing useful stuff."
Critics might argue that this is simply a growing pain of a new technology and that early adoption always looks wasteful. However, McCormick counters this by highlighting that even the vendors are admitting the problem. He cites Sam Altman, who acknowledged on CNBC that companies are waiting for costs to come down while wondering when revenue will appear, calling it a "huge issue." The evidence suggests this isn't just a learning curve; it's a structural flaw in how these tools are being deployed.
The Return on Tokens Equation
Once the illusion is shattered, McCormick pivots to a practical framework: Return on Tokens (ROT). He argues that tokens must be held to the same standard as any other business investment. "When you invest in a new machine, you expect it to generate a return," he writes, applying basic capitalist logic to the AI era. The formula is simple: Value of Output minus Cost of Tokens, divided by Cost of Tokens.
"The question is always: can the thing generate returns? For tokens, the question is: what is your Return on Tokens (ROT)?"
This section is the article's strongest analytical move because it forces a shift from volume to value. McCormick notes that companies are now scrambling to lower costs by routing tasks to cheaper models, including open-source alternatives from China. While this is a necessary step, he argues it misses the larger point: simply making agents cheaper doesn't solve the problem of them doing the wrong work. He draws a historical parallel to the mid-nineteenth century railroad boom, where companies raced to lay miles of track without regard for profitability, only to find that not all tracks led to gold.
"Agents are great at some things, but they're not the right shape for a lot of others."
The author suggests that the real solution lies in returning to deterministic code for repetitive tasks, a lesson learned half a century ago when NASA replaced human computers with software that never made mistakes. He points out that AI agents, which improvise and lack consistent accuracy, are ill-suited for high-stakes economic work like fraud detection or underwriting. This connects to the broader theme of tacit knowledge; as discussed in related deep dives on how expertise lives in people's heads rather than databases, agents cannot easily access the unwritten rules that make a business run smoothly.
"AI can only evolve what it can touch... The original sin is that there are no goals."
McCormick argues that without clear goals, code decays into "slop" because there is no purifying force to evaluate quality. This is a sharp critique of the current trend of setting agents loose on vague instructions and hoping for the best. He notes that even consulting subsidiaries from major labs are being launched to help companies figure this out, implying that the technology itself is not yet ready for prime time in its current form.
"If the Old Economy can't generate a ROT, well, this is creative destruction baby."
The Structural Flaw of Agents
The final leg of McCormick's argument identifies three structural reasons why agents fail to deliver positive returns: they cannot maintain high-quality standards over long-running tasks, engineers don't understand the work they are automating, and there are no clear goals. He emphasizes that while 80% accuracy might be fine for a prototype, it is "0% usable" for critical business functions.
"Agents improvise. They're spawned fresh onto repetitive tasks like every day is their first day on the job, which hurts consistent accuracy."
This observation highlights a fundamental mismatch between the probabilistic nature of current AI models and the deterministic requirements of enterprise software. McCormick's analysis suggests that until this gap is bridged, companies will continue to burn cash without seeing the promised efficiency gains. The argument is compelling because it moves beyond the usual "AI is amazing" or "AI is a bubble" dichotomy to offer a specific diagnosis: the architecture is wrong for the job.
"The issue is the companies have focused on maximizing tokens, assuming that tokens = value."
Bottom Line
McCormick's most significant contribution is his refusal to accept the current narrative of AI progress at face value, forcing readers to confront the reality that spending more does not equal building better. While his critique of "tokenmaxxing" is persuasive and backed by real-world examples of waste, it assumes a level of organizational discipline that many companies may struggle to implement quickly. The strongest takeaway for any leader is that the era of blind AI adoption is over; the next phase will be defined by rigorous measurement of return on investment.