Palantir's CEO Called AI a 'Wealth Tax' on Your Business. He's Half Right, and the Half He's Right About Tells You What to Sell.

On July 1, Palantir CEO Alex Karp went on CNBC's Squawk Box and torched the frontier AI industry. Enterprises, he said, are "paying for tokens that create no value" while the labs quietly harvest their proprietary data and "alpha" to improve their own models: a "wealth tax" on business. He called the token-based model "completely wrong," said the models have been "irresponsibly oversold," and threw in a line about how outsourcing the country's defense to "the consensus view in Silicon Valley" is "effing insane." Palantir stock rose more than 9% that day.

Now, the easy move is to note that a man whose company sells AI outcomes has an obvious interest in trashing companies that sell AI tokens, and leave it there. That's fair. But it's also lazy, because underneath the performance there's a distinction that actually matters if you're a one-person business, and the guy talking his own book happened to draw it clearly.

The half he's right about

Here's the part that's true whether or not you like the messenger. If your business is reselling raw model output, you are a pass-through, and pass-throughs get squeezed from both ends.

Think about what you actually own in that setup. The model isn't yours. The infrastructure isn't yours. The pricing isn't yours. It moves when the lab decides it moves. Your customer is paying you a markup on tokens they could increasingly buy themselves, and every quarter the "buy them yourself" option gets easier. That's Karp's "tokens that create no value," aimed at you: not that the model does nothing, but that the layer where you're sitting captures almost none of the value and carries almost all of the substitution risk.

And the data point is real too. When you pipe a customer's proprietary workflow through a hosted model, the interesting question of "who benefits from what the model learns about this problem" is not settled in your favor by default. You're doing the integration work (figuring out the prompts, the guardrails, the edge cases) and a lot of that hard-won knowledge leaks upward to the platform rather than compounding into an asset you own.

I've felt the small version of this. Early on I built a thing that was, if I'm honest, a nice prompt wrapped in a nice UI. It worked. It also had no floor, because the second the underlying model got better at the task natively, my entire reason to exist evaporated. That's the wealth tax Karp is describing, and for a solo operator it's not abstract. It's the difference between a product with a moat and a feature waiting to be absorbed.

The half he's wrong about

Where Karp is wrong, or at least where his conclusion doesn't transfer to you, is the implication that you need to be Palantir to escape this. You don't. The thing that makes the outcome layer valuable is not a fifteen-billion-dollar balance sheet. It's proximity to a specific customer's specific problem, and that's precisely the ground a solo operator can hold and a horizontal platform can't.

Tokens got cheap and capable at the same time. That's the setup, not the trap. It means the expensive part of building the outcome layer (the raw intelligence) is now a commodity you rent for cents, and the scarce part is the thing you were always going to have to supply anyway: understanding one narrow domain well enough that your product delivers a result, not a text completion. Karp sells that to the Fortune 500. You can sell the version of it that the Fortune 500 will never bother to build for a 200-person plumbing-supply distributor.

So the "wealth tax" is real, but it's a tax on the pass-through layer, and you are not obligated to live there.

What I'd actually do

Get off the token meter as your value proposition. Not literally (you'll still pay for tokens) but stop letting "we use GPT-5" or "powered by Claude" be the pitch. Nobody's moat is the model they rent.

Charge for the result. Price on the outcome the customer cares about (the invoice reconciled, the ticket resolved, the brief drafted and cleared), not on usage that mirrors your own cost structure, because usage-based pricing on a commodity input is just volunteering to be arbitraged. Own the two things the platform can't take: the evaluation harness that proves your output is good enough for this specific job, and the accumulated domain data and corrections that make your version better than a naked model call. That's the asset. That's what compounds while the token price falls.

And keep the thin abstraction so that when the next model is cheaper or better, you swap the engine and keep the car. If your product is genuinely the outcome, the model underneath is a supplier you can replace, which, conveniently, also means the "wealth tax" gets cheaper for you every year instead of more expensive.

Karp went on TV to protect Palantir's positioning. Fine. But he also, for free, described the exact gap a solo operator is best placed to fill: sell the outcome, own the data and the evals, and let the labs fight over the commodity underneath. The people getting fabulously wealthy, he said, are the ones selling the tools, not using them. The fix isn't to feel taxed. It's to move one layer up, where the value actually is.

Author

Lukas

@lukcombinator

Sources

• Palantir's Karp bashes token-based AI model as 'completely wrong' (CNBC)
• Palantir Billionaire Alex Karp Calls AI Industry 'Effing Insane' In Heated Interview (Forbes)
• Palantir CEO Alex Karp: Enterprises Are "Livid" Over AI Models That "Steal" Their Business Value (24/7 Wall St.)
• Karp unleashes on frontier AI labs (Axios)