Token-maxxing and the quiet transfer of the alpha
The loud objection to enterprise AI spend is that the value does not justify the tokens. The quieter one is what the spend hands over: a view of your business that took years to build, and, more slowly, the problems that kept your people engaged. Trust is a scarce resource right now, and token-maxxing spends it like it is abundant.
There is a phrase going around for the way a lot of companies now buy AI: token-maxxing. The idea is to push token usage up as an end in itself, to reward employees for running more of everything through the model, on the theory that the more you spend, the more the organization changes for the better. It is the familiar boat-store move. You asked for a fishing boat and left having nearly bought a yacht, because at every step the larger option was framed as the obviously smarter one. When it is a boat, the sales motion is easy to see through; it is not exactly subtle. But when it is token spend, the buyers, surprisingly, do not see it at all.
Two separate complaints get bundled together here, and they are worth pulling apart, because one is arguable and the other is not. The loud complaint is that the value does not justify the spend: that much of the usage produces disposable, one-off code that no one ever really reviewed, because everything moves too fast and the developer’s natural incentive, under the market’s constant pressure, is to move on to the next task. What remains is the pleasant feeling of progress, in place of software the developer actually understands, that solves a problem they believe in, built the best way they know how. The quieter complaint is that the spend buys the provider something: a view of how your business actually works, assembled from the prompts you send. The first is a claim about ROI, and reasonable people land in different places on it. The second is a claim about who ends up holding the knowledge the company earned the hard way. And behind both of them sits a third cost that somehow never makes it into the spreadsheet: what this way of buying does to the people who work for you, to their sense of meaning at work, and to the kind of problems they still get to solve themselves. That one took me the longest to see, and I now suspect it is the largest.
The ROI complaint is the weaker one
It is genuinely hard to know whether a given quarter of AI spend paid for itself. But it is worth being precise: the fact that the value is hard to measure is not proof that it is not there. Value from these tools tends to be diffuse. It shows up as small time savings spread across many people, as tasks nobody would have attempted at all without these tools, and, honestly, as some money simply wasted on attempts that led nowhere. Netting that out honestly is real work, and most organizations have not done it. So the strong version of the complaint, that the value is obviously not there, mostly reads as an assertion. It might be right. It is not settled, and I would not state it as if it were.
What is easier to defend is the narrower point buried inside it: usage for its own sake is a bad target. When the metric is tokens consumed, people optimize the metric. You get more calls, more retries, more of the addictive sense that something is happening, and none of that is the same as the work being better. This is an old failure mode wearing new clothes. If you reward the proxy, you get the proxy. The fix is not necessarily to use less AI out of stubborn principle; it is to measure the outcome you actually wanted from the organization and let usage fall out of that.
The real concern is about the data
Here is the part that survives scrutiny. Every prompt you send is a small, honest description of a problem your business is trying to solve, hopefully to make the world a little more pleasant for everyone. In isolation, one prompt is nothing. In aggregate, across a whole organization, across many organizations in the same industry, the prompts describe the shape of the work: what people are stuck on, what they are building, which parts of the market are heating up. A provider sitting at that vantage point can see patterns no single customer can see, because no single customer has the cross-sectional view.
It helps to be precise about the word people reach for here, which is alpha. Stripped of the trading-desk glamour, it just means your edge: the specific, defensible things that let you compete, the accumulated knowledge of how to win in your particular corner. The mechanism is real: heavy AI usage can, in principle, transfer some of that edge to whoever runs the model, because usage data is genuinely valuable and it flows in one direction. You send the problems. They see the pattern of problems. The question is not whether the channel exists. It is whether the party on the other end has earned that view of you, and I think the honest answer right now is: not yet, and not at the pace the spending assumes.
None of this is a private worry, and little of it is original. Palantir’s Alex Karp spent a recent CNBC interview (July 2026) making a much louder version of the same argument: that enterprises are livid about paying for tokens that create no value, and that providers are harvesting their customers’ alpha along the way. He runs a company that competes with those providers, so apply whatever discount you think that deserves. But it is worth noticing when the loudest version of the quiet complaint comes from inside the industry.
Why the pricing question is a fair one
There is a rhetorical challenge that comes up in this discussion and I think it lands: if the model were as transformative as the marketing says, why sell tokens at all? If a provider could reliably take your revenue and multiply it, the obvious move would be to skip metered pricing entirely and take a cut of the upside, the way anyone confident in the value would. Charging by the token is, among other things, a revealed preference. It is what you do when the value per call is uncertain and variable, not when you are sitting on a machine that prints money for whoever plugs it in.
I would not push this too far. Metered pricing is also just the normal way to sell a utility with real marginal costs, and plenty of useful, non-magical products are priced per unit. The point is narrower: the pricing model is evidence against the strongest hype, not proof of a conspiracy. It is consistent with tools that are useful and improving and also not the guaranteed revenue multiplier the loudest pitches imply. Both things can be true, and usually are.
Trust is earned slowly
It is worth granting the strongest version of the worry before answering it. The incentive to exploit a vantage point is structural, and the examples are old and abundant: cloud providers that watched what customers ran on their infrastructure and launched competing managed services, retailers that watched what sold on their marketplace and shipped a private label. You do not need to assume bad intent to see the shape of it; whoever watches the demand is, at minimum, well positioned to build toward it, and structural incentives do not go away because everyone involved is well-meaning. If you can observe where the demand is, the pull to build toward it is constant. Pretending otherwise would be naive.
The standard reassurance is that the serious providers police themselves: alignment and safety teams whose actual job is to make sure the systems built on your data cannot be turned against you, terms that promise your prompts stay out of training, and the reputational fact that a lab caught mining customer edge would pay dearly for it. I believe most of that. The people doing that work strike me as sincere, and the promises are probably being kept today. But sincerity is not a track record, and a promise is not an institution. The kind of trust being asked for here, hand us a live feed of your hardest problems, is normally built over decades of kept commitments, and this industry is a few years old in its current shape. Terms get rewritten, leadership changes, companies reorganize around new incentives, and the customer usually finds out afterwards. None of that requires bad faith. It just means trust is a scarce resource right now, and the token-maxxing posture spends it as if it were abundant. That is the same mistake as rewarding token consumption, applied to something much harder to buy back.
The cost nobody prices
There is a quieter cost, and it has nothing to do with the provider. When the mandate is to push everything through the model, the interesting problems leave the building first, because the interesting problems are exactly the ones the model is most impressive on. What stays behind is review, supervision, and glue. Andrew Wegner has written about what this does to junior engineers specifically: mentorship turns into validating machine output, and the question shifts from why did you choose this approach to did the AI pick the right pattern. A lot of engineers are living some version of this already: the work still ships, the dashboards look fine, and the part of the job that made it worth doing, sitting with a hard problem until it gives, is quietly gone. I think this is a real crisis of meaning, not a mood. People took these jobs to solve problems. Solving problems is the human part of this profession, and a usage mandate optimizes it away because tokens are easy to measure and meaning is not.
There is also a version of this posture that assumes, without ever saying it out loud, that your employees are no longer smart enough to be trusted with the hard parts. That assumption does a lot of silent work in the token-maxxing pitch, and I think it is wrong on the merits. The model has read everything, but your people hold the context: the threat model, the org chart, the history of why the last attempt failed. Those are the inputs the decisions actually turn on, and they do not live in anyone’s training data. This connects to the data concern from earlier: I argued in another essay that this context is the part of the work that stayed expensive, and indiscriminate usage is precisely the channel that streams it out. Let your employees solve real problems. They like it, they are better at the deciding part than the pitch assumes, and every decision made in house is one your organization still knows how to make next year.
Freedom is the thing worth maximizing
If there is a bet underneath all of this, it is about freedom. The pace makes everything feel mandatory: adopt now, spend now, keep up, and people who like the human touch in their work are finding that the pace makes it hard to keep. But the companies that come out of this period strongest will, I suspect, be the ones that maximized their freedom of movement: free to switch providers because no single one holds their whole picture, free to say not yet because their people can still build, free to keep a hard problem in house because the muscle never atrophied. And freedom compounds downward. A company that preserves its own room to move can afford to give its people room to think, and people do their best work with room to think. That is a guess about the future, so hold it loosely. But the opposite bet, that the winning move is maximum dependence at maximum speed, seems to me the stranger one.
What actually follows from this, in my view
Not abstinence. The tools are useful, the usefulness is compounding, and swearing them off mostly means watching competitors who did not. The conclusion is narrower and less comfortable than either buying more or hoarding everything. Treat provider trust as a budget and be deliberate when spending it. Send the mechanical work out freely. Keep the hardest problems, and above all the decisions, in house, not only because of what leaks out but because of what stays alive inside.
The load-bearing question was never whether these tools deliver value. For a lot of work they clearly do. It is what you are trading for that value: a view of your business that took years to accumulate, handed to a counterparty whose track record is measured in quarters, and a kind of work that kept your best people engaged. Token-maxxing treats the spend as the goal, the data as exhaust, and the people as reviewers. I do not have a formula for where the line sits. But trust is scarce right now, meaning is scarcer than it was, and a purchasing strategy that burns both to make a usage graph go up is optimizing for the wrong thing. None of this is a new mistake. We have traded the durable for the measurable before, and the record on how that ends is not flattering. The only new part is the timing: this time the mistake is visible while it is still cheap to stop making it.