The AI Story Is Breaking. Your Best People Were the Hedge All Along.

Over the last couple of months I’ve been watching something happen inside enterprise AI programs that I haven’t seen since this whole wave kicked off. I can’t name the companies, NDAs being what they are, but I can tell you the shape of it. Claude Enterprise rollouts getting paused mid-flight, or at least throttled. Token budgets getting capped. Integration work with development teams deliberately slowed down, not because anyone stopped believing in the tools, but because somebody with budget authority finally asked the question that should have been asked at the start: what value is this actually bringing to the teams, and can we see it before we keep paying for it?

That question sounds boring. It’s not. For two years the answer to “should we spend more on AI” was yes by default, because the story everyone had bought said the spend was a down payment on replacing half the org chart. You don’t measure a down payment on the future. You just keep paying it.

The measuring has started. And that tells you the story it was propping up is coming apart.

Palantir Just Gave the Disease a Name

This week Palantir posted a 9-point manifesto on X about what it calls AI sovereignty, and buried in point three is the most useful new word I’ve seen in this space in a while: tokenmaxxing.

Here’s the line, straight from the manifesto: “Tokenmaxxing hijacks your value orientation and decreases your institutional fortitude and intelligence.” The pursuit of high token usage, Palantir argues, “incentivizes disposable scripts over robust software — with the addictive feeling of false progress.” And then the kicker, the sentence I’d frame and hang in every CFO’s office: “There is a reason why those selling tokens refuse to charge based on value.”

Alex Karp has been circling this for a while. Back in June he told CNBC that the frontier AI labs “don’t understand how unlikeable they are” and that their products “don’t actually work the way” customers expect. You can roll your eyes at the delivery, and plenty of people do. And sure, Palantir is talking its own book here. A company that sells sovereign, keep-your-data-in-house AI infrastructure has every incentive to tell you that renting intelligence by the token is a trap. Fair. But an argument isn’t wrong because the person making it benefits from it being right. It’s wrong or right on the receipts.

So let’s look at the receipts, because I’ve been stacking them for a year.

I’ve Been Charting These Symptoms All Along

Point three of the manifesto, the tokenmaxxing one, is the argument I made in When Your AI Bill Becomes Your Biggest Line Item and again in Tokens Aren’t Going to Zero. Token spend grows faster than the value anyone can point to, the prices aren’t heading toward zero no matter how many keynote slides say otherwise, and the meter is designed so that the seller wins whether or not you do. When Nvidia’s own VP of applied deep learning says the cost of compute for his team runs “far beyond the costs of the employees,” that’s not me theorizing anymore. That’s the receipt.

Point four says your weights, and more broadly your accumulated institutional knowledge, are “the distilled form of hard-won, accumulated institutional knowledge,” and that letting someone else control them means letting them “migrate the alpha of your business to theirs.” I wrote a rougher version of the same idea in The Tokens-for-Engineers Trade Has Three Holes: every agent session starts from zero, the model never accumulates your context, and the twenty years of failed paths living in your senior engineer’s head are not in the weights and never will be. You can buy more tokens next week. You can’t buy back 15 years of somebody knowing why the company does it this way instead of that way. Palantir is saying the corporate version of that: your data and your tribal knowledge are the treasure, and shipping them out by the API call is how the treasure walks.

And point seven might as well be the thesis statement of The CEOs Walked It Back. The Layoffs Didn’t. — “Listen to those closest to the problems, not those speaking most compellingly about them.” For two years, the people speaking most compellingly were lab CEOs forecasting white-collar bloodbaths and boards running replacement math on spreadsheets. The people closest to the problems were the engineers and data leaders saying, over and over, that the tools are genuinely good and also genuinely not a person. One of those groups has since said, in public, that they were “pretty wrong.” It wasn’t the second one.

I’m not laying all this out to take a victory lap, or at least not mostly. The reason to connect the dots is point nine of the manifesto, which says the only signal worth trusting is a track record of being right. That’s a standard you should hold me to as much as anyone. Go read the thread and judge it on the record.

The Wallet Is Doing What the Warnings Couldn’t

Here’s the thing I find almost funny, in a grim way. None of the warnings moved anybody.

A year of essays, mine and plenty of others’, about sample efficiency and capacity constraints and burn rates, and the replacement narrative just kept rolling. What finally slowed it down wasn’t an argument. It was the invoice. Uber’s COO admitting the company burned its entire 2026 AI budget by April and couldn’t draw a line from the spend to better products. Microsoft cancelling most of its Claude coding licenses and pointing engineers at cheaper tooling, with Nvidia making the same call for the same reason. The 99% of executives who told a survey they still expected AI-driven layoffs, right before the CEOs who seeded that expectation started walking it back. And now, in my own orbit, enterprise rollouts pausing so somebody can finally measure value per token before renewing the line item.

That’s what “the AI story is breaking” means, and I want to be precise about it because I’ve been careful on this thread not to call tops or predict crashes, and I’m not starting today.

The technology isn’t breaking. The models are real, the gains this year are real, and I run these tools every day on my own projects. Anyone who reads this as “AI was fake” is reading a different post. The market may or may not correct, and the timing of that is not something I or anyone else can call, which is exactly what I said when I wrote about the labs burning cash.

What’s breaking is the story. The specific story that said tokens trend to zero, agents replace headcount one for one, superintelligence arrives on a schedule, and any spend today is justified by the org chart you won’t need tomorrow. That story is losing its grip on the people who fund it, not because they read the counterarguments, but because they’re now two or three budget cycles in and the promised trade never cleared. Leadership and investors are starting to see AI for what it actually is right now: a genuinely powerful value-add tool for the people you already have, not a wholesale replacement for them. It took the wallet to make that visible. Fine. Whatever works.

Your Best People Were the Hedge

Which brings me to the part I care most about, because it’s the part leaders can still act on.

If the story is breaking, the question is who’s exposed. And the answer, mostly, is the organizations that traded away their hedge. The companies that cut senior people on the replacement thesis are now holding tools they don’t fully understand, integration projects nobody can evaluate, and a compute bill that runs past what the departed people cost. When the pause-and-measure moment arrives at those companies, there’s nobody left who can do the measuring. That’s the trap in full: you can’t assess what AI is worth to your teams if you already let go of the people who knew what good looked like.

The organizations that kept their best people are in a completely different position, and it’s worth spelling out why. They can pause without panic, because the work still gets done while they measure. They can measure honestly, because the people closest to the problems are still in the building to say what’s real and what’s the addictive feeling of false progress. They can negotiate with vendors from leverage, because walking away doesn’t strand them. And they’ve still got their treasure, the data and the tribal knowledge and the judgment, compounding in-house instead of migrating out through somebody else’s API.

That’s why the headline of this post says what it says. Through the entire bubble narrative, the boring move, the one that never got a keynote, was keeping the right people in the right spots and handing them the tools. It looked conservative. It turned out to be the hedge. The people were the thing that let you survive being wrong in either direction, whether AI under-delivered or over-delivered, and they’re the thing that lets you calibrate now that the invoices are forcing everyone else to.

So the Monday-morning version, same as I’ve been telling the execs who ask, just with a fresh manifesto to back it up. Keep the people who have the knowledge, the skill, and the technical ability your company needs to stay competitive. Protect your data and your institutional knowledge like the treasure they are, because whoever controls that controls your fate, and it doesn’t have to be you. Measure value per token before you renew the spend, the way the smart programs in my orbit have already started to. And treat anyone still selling you the unlimited-tokens-to-utopia story with the same skepticism you’d give any vendor who refuses to charge based on value.

The story is breaking. The wallet broke it. And if you spent the last two years holding onto your best people while everyone else ran the replacement math, you didn’t miss the AI wave. You hedged it, and the hedge is about to pay.