Karp calls token AI billing a 'wealth tax' — own your stack

Palantir CEO Alex Karp went on CNBC this week and called the way most companies buy AI a "wealth tax." He wasn't talking about the sticker price. He was talking about what you hand over every time you send a prompt to someone else's model — your data, your workflows, the proprietary edge Karp calls your "alpha." It's a loud quote, but the underlying point is one we've been building around for a year.

What actually happened

On July 1, on CNBC's Squawk Box, Karp bashed the token-based AI model as "completely wrong," accusing frontier labs of charging businesses heavily for tokens that deliver little value while harvesting the data those businesses feed in to improve their own models. Forbes reported he framed it as a "wealth tax" on companies — high fees on one side, quiet data collection on the other. On national security he was blunter: "Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane."

The pitch behind the outburst: Palantir and Nvidia announced a sovereign-AI reference architecture — a full-stack, air-gapped deployment that keeps the model and the data inside your own walls. Palantir has a product to sell here, so read the theatrics accordingly. But the mechanism he's describing is real.

Why it matters for your business

Karp runs a company that sells to governments. You run a business in Phoenix or Peoria. The tax is the same shape at both scales. When your quoting logic, your customer records, and your operational playbook all flow through a metered API you don't control, two things are true: the meter runs on someone else's terms, and the thing you're feeding it is the exact data that makes your business yours.

You don't need an air-gapped Nvidia cluster to defend against that. You need architecture. Keep your data in systems you own. Put a model behind an interface you can swap. Log what leaves your building and why. Then a model is a tool you rent by the hour — not a landlord that learns your trade while charging you rent.

The point isn't to fear the frontier labs. We use them daily; they're good. The point is that "convenient" and "captured" look identical right up until pricing or terms change — and by then your data's already been the tuition.

Sources: CNBC, Forbes.