On July 1, 2026, Palantir CEO Alex Karp went on CNBC's “Squawk Box” and lit into the way the biggest AI companies charge for their products. He named OpenAI and Anthropic. He called the pay-per-token pricing model a “wealth tax” on businesses. CNBC's own headline said something had “gone completely wrong.”
Karp runs a company that sells to governments and huge enterprises. His fix (own your compute, your models, your data) is not something a solo founder or a five-person shop is going to do next week. So why should you care?
Because underneath the fight is a plain question every small business now faces when it buys AI. Are you renting by usage, and what happens to your data while you do it? That question does not need a billion dollars to answer. It needs about ten minutes and the right things to ask.
What Karp actually said
Speaking on CNBC's “Squawk Box” that morning, Karp said other CEOs are privately “livid” with the leading AI companies. Two things set them off.
First, the bill. In his words, describing the customer's view: “I am paying for tokens that create no value.” A token is roughly a chunk of text the AI reads or writes. Consumption pricing means your cost climbs with how much the tool processes, whether or not the output helped you.
Second, the data. He put the customer complaint bluntly: “These people are stealing the weights and alpha of my business.” Translated out of Wall Street language, “alpha” is your edge. The proprietary way you price, sell, or serve that makes you money. His worry is that feeding that into someone else's model helps improve their product.
He tied it together with the line that got quoted everywhere: the labs are “creating a wealth tax that does not help the poor, it just punishes.”
Karp's answer is ownership. He said enterprises and governments want “control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production.” Two days earlier, on June 29, Palantir announced an engine with NVIDIA for running open models inside “classified, air-gapped, and other sensitive environments.” The commercial version of own-it-yourself. Investors liked what they heard. Palantir stock rose roughly 8 to 9 percent that day.
You do not need to take Karp's side to use his questions. A billionaire yelling on cable television is one thing. The checklist buried in his complaint is another, and it is genuinely useful.
The two questions any AI buyer can ask
Strip away the theater and Karp's grievance becomes a short diligence list. Before you sign up for a paid AI tool or connect it to your customer data, ask the vendor two things.
Are you keeping my data? Look for the answer in the terms, not the marketing page. Does your input get used to train their model? Can you turn that off? Many business and team plans already promise they will not train on your data. Free and consumer plans often say the opposite. The setting is usually there once you go looking.
Are you going to enter my business? Meaning, does using this tool hand over the thing that makes you distinct. Your client list, your pricing logic, your intake process. Sometimes the tradeoff is worth it for the speed. Sometimes it is not. The point is to make that call on purpose rather than by clicking Accept.
Neither question requires a legal team. They require reading two paragraphs and deciding what you are comfortable with. That habit, knowing what to ask a tool before you trust it, is exactly the kind of judgment The 7 Levels of AI Proficiency is built to grow. Early on, a working tool feels like enough. Further up the climb, you start interrogating that tool the way you would a new vendor or a new hire.
The uncomfortable ROI number (and it is not Karp's)
Here is a fact that has nothing to do with Palantir and holds up on its own.
MIT researchers, through a project called NANDA, studied how generative AI is working out inside companies. Their finding: about 95 percent of enterprise generative-AI pilots delivered no measurable return to the bottom line. Only about 5 percent captured significant value. The study drew on 150 leader interviews, a survey of 350 people, and 300 public AI deployments.
of enterprise generative-AI pilots delivered no measurable return to the bottom line. Only about 5 percent captured significant value.
Source: MIT Project NANDA, 2025Read that again. Nineteen out of twenty AI projects at real companies showed no money back.
Nineteen out of twenty AI projects at real companies showed no money back.
That is the anxiety Karp is channeling, whether he cites the number or not. Businesses feel the spend and cannot find the payoff. But the study points somewhere more useful than blaming the AI companies. When almost every pilot stalls, the common thread is rarely the model. It is the approach. The tool got dropped on top of a messy process and told to perform.
This is where a small business actually has an advantage over the enterprise. You can see your whole operation. You know where the real hours go. You can pick one task, redesign how it flows, then point the AI at it. The 5 percent that works tends to start there, with a specific job and a clear before-and-after, not with a company-wide rollout and a hope.
What to do with all this if you run a small business
You are not going to build your own AI models. Ignore that part of the story. Four steps this week put the rest of it to work.
- Read the data terms on the AI tools you already pay for. Find the training toggle. Decide on purpose whether to leave it on.
- Separate your spend from your results. Look at what you paid an AI tool last month, then name one concrete thing it saved you. Hours, a won customer, a task you stopped dreading. If you cannot name it, you found your next question.
- Pick one workflow, not ten tools. Redesign a single repeatable task first. Quoting, follow-up, scheduling, whatever eats your week. Then automate that. The MIT number says the scattered approach is where the money disappears.
- Track the tradeoff honestly. Renting AI by the month is fine for most small businesses. Convenience and speed are real value. Just know what you are trading for it, and revisit the call as you grow.
Where to start
Pick the AI tool you use most. Open its terms, find whether it trains on your data, and set that toggle the way you want it. Then write down one task you would hand to AI if you trusted it fully. That single task, redesigned and measured, is worth more than any argument happening on cable news.
If you want a read on where you stand and what to build next, The 7 Levels of AI Proficiency assessment takes about ten minutes and gives you a starting point instead of a guess.
Related reading: Level 3: The Lieutenant (Critical Thinker).
Sources
- Palantir's Karp bashes OpenAI, Anthropic token model: 'Something has gone completely wrong'
- Palantir billionaire Alex Karp calls AI industry 'effing insane' in heated interview
- Palantir Launches Engine for Deploying NVIDIA Nemotron Open Models in Sovereign Environments
- The GenAI Divide: State of AI in Business 2025 (MIT Project NANDA)
Frequently Asked Questions
Is token-based or usage-based AI pricing bad for a small business?
Not on its own. For most small operations, a flat monthly plan or modest usage-based billing is the practical choice, and the value can be real. The concern Karp raised applies more to companies spending heavily on tokens that produce nothing. The lesson for you is to watch whether your spend is tied to actual results, not to defend a particular pricing model.
Should I worry about my data training someone else's AI?
Worth a look, not a panic. Check whether the specific plan you use trains on your inputs and whether you can turn that off. Business and team tiers often already promise they will not. The habit of checking before you connect sensitive data is the win.
Does the MIT finding mean AI does not work for small companies?
No. It means most projects that fail do so because of how they were rolled out, not because the technology is broken. The small share that worked tended to start with one well-chosen task and a clear measure of value.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.