Alex Karp is half right: most AI spend creates no value. Here is why.

On Wednesday morning, Palantir CEO Alex Karp went on CNBC to talk about a new partnership with Nvidia and, somewhere around minute three, stopped talking about the partnership. What followed was roughly twenty minutes that the internet quickly labelled a "televised nervous breakdown." Palantir stock jumped more than 9% anyway. The clip went everywhere. And buried inside the digressions was a single sentence that every mid-market operator should sit with, because it is more true than Karp probably intended it to be.

American businesses are "livid" because they are "paying for tokens that create no value."

Strip away the theatrics and that line is the most honest thing a technology CEO has said about the AI market all year. It is also, in the way Karp meant it, wrong. Both of those can be true at once, and the gap between them is exactly where the money is being lost.

What he actually said, under the noise

The performance obscured the argument, so it is worth separating the two. Karp called the AI industry "effing insane." He accused the large model providers, naming OpenAI and Anthropic, of overcharging enterprises, of exploiting the data those enterprises hand over, and of putting national security at risk by becoming the default brain behind government systems. He described the prices being charged as a "wealth tax" on business. On the national-security point he was at his sharpest:

Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.

When a host noted he sounded angry, he answered, "This is the voice of American business that is being channeled through me." After the segment appeared to wrap, he asked, "Are we still on?"

The commercial point underneath all of it was straightforward. Palantir and Nvidia are pairing open models with Palantir's software so that governments and large enterprises keep more control of their data, customise more of the stack, and pay less than closed frontier products demand. It is a real pitch aimed at real buyers. It is also, conveniently, a pitch that requires the audience to believe the incumbents are overpriced data-hoovers. So the rant and the product are not separable. The anger is the marketing.

None of that makes the core observation false. It makes it worth examining without the salesmanship attached.

What social media did with it

The clip did what chaotic clips do. Within hours a post sharing the full segment was racing across X, and the framing stuck:

Here is the entirety of Palantir CEO Alex Karp's televised nervous breakdown this morning on CNBC.

Timelines filled with the meltdown read: the stuttering, the tangents, the moment he asked whether the cameras were still rolling. The dominant mode was mockery, and on the delivery alone it was earned. But watch the replies for more than a minute and a second reaction shows up, and it is the more interesting one. A steady stream of founders, operators, and engineers said some version of the same thing: the man is clearly having a bad morning, and he is also not wrong. Unhinged delivery, correct diagnosis. The "paying for tokens that create no value" line got quoted approvingly by exactly the people who are writing those cheques, often by accounts that have no love for Palantir at all.

That is the tell worth noticing. When a point survives the dunking, when people repeat it even while they laugh at the man who said it, the point is doing work the messenger could not. The discussion split into two conversations: one about a CEO losing the plot on live television, and one quiet, cross-partisan agreement that enterprise AI spend is not delivering, and that everyone has known it for a while and mostly kept it polite. Karp made it impolite. That is why it travelled.

The part he got right

Enterprises really are paying for AI that produces no measurable value, and they really are quietly furious about it. This is not a Palantir talking point. It is one of the best-documented findings in the whole field. MIT's 2025 work on enterprise AI found that roughly 95% of generative AI pilots delivered no measurable impact on the P&L. RAND has put the share of AI projects that never reach production or business value even higher. The pattern is not that the models do not work. The pattern is that the spend does not convert.

So when Karp says businesses are livid about tokens that create no value, he is describing something an operator can feel in their own budget. A team stands up a copilot, a summariser, a support assistant. The invoice arrives every month. The productivity, the retention lift, the cost line that was supposed to move, does not move. Six months later the tool is a browser tab nobody opens and a renewal nobody remembers approving. Multiply that across a company and you get a real number, and real anger, and a CEO channeling it on live television.

Karp deserves credit for saying the quiet part loud. Most people in his seat are still selling the dream. He named the disappointment. That is useful, even if his diagnosis of the cause serves his own book.

The part he got wrong

Here is where the framing falls apart. Karp's explanation is that the value is missing because the vendors are villains. They overcharge. They take your data. The tokens themselves are worthless. Switch to open models on better-aligned software and the value appears.

That is not what the evidence says, and any operator who has actually run one of these projects knows it.

The same MIT and McKinsey research that documents the failure rate also names the fix, and it has nothing to do with which model you bought. McKinsey's analysis of where AI actually moves earnings puts end-to-end workflow redesign as the single biggest driver of impact, ahead of model choice, ahead of budget, ahead of talent. The projects that pay off are the ones where the company changed how the work is done and trained the people who do it. The projects that die are the ones where AI was bolted onto an unchanged process and everyone was left to figure it out on their own.

This is the flaw in the "worthless tokens" story. A token that creates no value in a broken workflow will create no value in a cheaper, more open, more private workflow either. You can move from a closed frontier model to an open one running on your own infrastructure, cut the bill in half, keep every byte of your data, and still end the quarter with a tool nobody uses and a metric that did not budge. Karp is selling a cheaper engine to people whose actual problem is that the car has no wheels.

The data and lock-in concerns are legitimate, and mid-market companies should take them seriously. Owning your data and keeping optionality on models is good hygiene. But hygiene is not the same as impact. You can do all of it perfectly and still get no return, because the return was never going to come from the model. It comes from the process and the people around it.

Why this matters more for the mid-market than for Palantir's customers

There is a second trap in taking Karp's advice at face value, and it is specific to the companies we work with. Palantir sells to governments and the largest enterprises on earth. Those buyers have the scale to run their own model stack, the engineering depth to customise a platform, and the leverage to negotiate. When Karp says the answer is more control and a more open stack, he is describing a solution that fits his buyers.

A mid-market consumer-tech company is not that buyer. If you run an app, a D2C brand, a marketplace, or a subscription business, you are not going to out-engineer OpenAI on model efficiency, and you should not try. Rebuilding your own AI infrastructure to save on tokens is the mid-market version of a nervous breakdown: a lot of motion, a large bill, and a national-security speech nobody asked for. Your constraint is not that you lack control of the model. Your constraint is that you have a handful of workflows that matter, a team that is already busy, and a very short runway to prove that any of this pays for itself.

For you, the lesson from the interview is not "buy open, own the stack." It is the quieter thing Karp stumbled into and then talked past: stop paying for value that never arrives, and find out why it is not arriving before you spend another dollar.

What paying for value actually looks like

If Karp's rant lands anywhere useful for a mid-market operator, it should land here. Not in a model migration. In a discipline. Four principles separate the AI spend that pays off from the spend that becomes next year's livid CNBC clip.

• Measure the value before you scale the spend. The reason 95% of pilots show no impact is partly that nobody defined the impact. Pick the metric the use case is supposed to move, retention, conversion, cost-to-serve, cycle time, before you sign anything. If you cannot name the number, you are buying tokens, not outcomes.
• Redesign the workflow first, then add the AI. This is the single highest-leverage move, and the one everyone skips because it is the hard, unglamorous part. Map the actual process, find where the time and errors and drop-off live, redesign that, and only then decide where a model fits inside it. A model dropped on top of a broken process just makes a faster broken process.
• Treat adoption as the deliverable, not the afterthought. The tool is not done when it ships. It is done when the team's habits have changed and the new way of working survives past the novelty. That is training, change management, and the operating cadence that keeps it alive in month six. No model provider, open or closed, will ever do that work for you.
• Keep optionality, but do not confuse it with impact. Own your data, avoid getting locked into a single vendor's roadmap. Karp is right that these are real risks. Just do not let good hygiene masquerade as a strategy. Optionality protects your downside. It does not create your upside.

The real takeaway from a bad morning on CNBC

Alex Karp had, by most accounts, a rough twenty minutes. But the reason the "tokens that create no value" line travelled is that it is true, and everyone paying an AI invoice knew it was true. The mistake is in what he did with it. He turned a problem about how companies work into a problem about which company they buy from, because the second version has a Palantir-shaped answer.

The honest version is less dramatic and more expensive to ignore. Anyone can spin up AI now. The tokens are cheap and getting cheaper, from every provider, open and closed. What is scarce, and what actually decides whether the spend shows up in the P&L, is the process redesign and the adoption work that turns a launched tool into a durable gain. That is the 80% of the job that Karp did not mention, because it is not for sale in a partnership announcement.

It is the only part that makes AI pay off. And it is the part most teams are still skipping, quarter after quarter, right up until they too become livid about tokens that created no value.

Anyone can spin up AI. We make it pay off. 10dem is an applied AI consultancy for mid-market consumer tech: we do the hard 80%, process redesign, adoption, and enablement, so AI actually moves your P&L. If you are not sure your AI spend is paying off, that is exactly the question our teardown is built to answer.

Palantir's Karp bashes OpenAI, Anthropic token model: 'Something has gone completely wrong'

CNBC's report on Alex Karp's remarks about AI pricing, customer data, national security, and the expanded Nvidia partnership.

CNBC

More on the reaction: Forbes and Futurism.