The Twilight of Tokenmaxxing: Why the AI Story Just Flipped From Consumption to Control
Enterprises are done burning tokens to look busy, Washington now decides who touches the best models, and the smart money is quietly rewriting what an AI moat even is.
The vibe shift no one wants to say out loud
Six months ago the flex was consumption. You proved you were serious about AI by torching tokens like a startup burns runway. That era is ending, and it is ending fast. The same week the industry was digesting yet another record model release, three separate stories converged on one uncomfortable truth: the constraint in AI is no longer capability. It is cost, access, and judgment. Whoever masters those three wins the next 18 months, and it will not be the people bragging about their token counts.
The capability curve is genuinely bending upward. Newer frontier systems can now grind autonomously for hours on complex software work that would have taken a human team a week or more. That is real, and it matters. But raw capability has quietly become the least interesting variable in the equation. The interesting stuff is happening in the plumbing and the politics.
Tokenmaxxing hit the wall, and the wall was the CFO
The clearest signal is the collapse of what got nicknamed tokenmaxxing, the practice of measuring productivity by how many tokens your people burn. It produced exactly the dysfunction you would expect. Meta ran an internal leaderboard called Claudeonomics that let 85,000 employees compete to be the top AI token consumer, and total consumption hit 60 trillion tokens in a single month. The top user was burning 281 billion tokens per month, earning badges like "Token Legend." The dashboard was pulled within days once the numbers hit the press.
Then the bills landed. Uber acknowledged it had spent its entire 2026 AI budget in the first four months of the year, with its COO saying it was becoming harder to justify internal AI costs. The company's response was a hard cap of roughly $1,500 a month per employee per coding tool, with exceptions granted case by case.
Here is where I would push back on the doom narrative. When SemiAnalysis actually talked to more than fifty enterprises, the picture was less dramatic than the headlines. The Meta and Uber blowups looked more like the product of bad incentives and loose oversight than proof that AI does not pay off. Budgets are now normal, but there is no agreed number: some defense and pharma firms cap staff at a few hundred dollars a month, while others run into the thousands, and data scientists reliably get the biggest allowance because they chew through the most tokens.
And the genuinely counterintuitive part, the bit worth tattooing on a whiteboard: tokens got radically cheaper, and spending went up anyway. The price of a unit of inference has been falling by something close to tenfold a year, yet bills climbed. That is the Jevons paradox in a suit. When something useful gets cheap, you consume dramatically more of it. The lesson is not to spend less. It is to spend deliberately, match each task to the cheapest capable model, and measure what actually shipped. The metric that matters now is cost per accepted outcome, not tokens consumed. It is the same arc cloud spending traveled a decade ago, and the discipline that emerged was called FinOps.
Washington took the wheel on the frontier
While companies were tightening the money side, the government tightened the access side, and this is the story with the longest tail. On June 26, 2026 the US gated two American frontier models on the same day: Anthropic's Mythos 5 was re-authorized for a short-list of trusted US organizations, while OpenAI previewed GPT-5.6 Sol only to partners the government had individually approved.
The mechanics differed in a way that tells you everything. Anthropic's models were forced dark two weeks earlier under export-control authority, because deemed-export rules meant even foreign-national employees could not touch them. OpenAI, watching that happen, pre-negotiated a gated preview rather than risk being switched off. The June executive order explicitly rejects mandatory licensing and asks only for voluntary pre-release access, but the earlier forced shutdown gave that voluntary framework de-facto teeth. One lab was compelled, the other cooperated, and both landed in the same place: the best models are no longer things you can simply go buy.
The business consequence is a gift to open weights. The move quickly pushed global demand toward cheaper Chinese open-source models, and the pitch writes itself: intelligence that cannot be revoked at the stroke of a bureaucrat's pen. For a company building on top of a model, this reframes vendor choice as a sovereignty question. Model routing, data control, and a credible open-weights fallback are no longer nice-to-haves. They are risk management.
The moat moved, and so did the money
Step back and the three threads braid together. If capability is abundant, cheap to copy, and occasionally yanked offline by the government, then the durable advantages sit elsewhere: in owning the customer relationship, in the accumulated context you feed the model, in taste, and in the discipline to spend on outcomes rather than optics.
The capital markets already sense this. Anthropic's run-rate revenue reached $47 billion as of late May 2026, up from $9 billion at the end of 2025, driven primarily by enterprise adoption and Claude Code, and both it and OpenAI have filed confidentially to go public. But even the analysts betting on these IPOs are candid that current growth rates are the fastest these companies will ever post, which is a good reason to list now, as is the concern that some of their largest enterprise customers may start limiting out-of-control token spend.
So here is the so-what. The winners of this next phase are not the biggest spenders or even the makers of the smartest model. They are the operators who treat AI like any other line item with an owner and a return, who architect around a model rather than marrying one, and who understand that when everyone can build the thing, the advantage shifts to whoever is closest to the customer and compounds the most context. The chatbot era measured usage. The era starting now measures results. Adjust accordingly.