Palantir CEO Alex Karp has a blunt take on the AI industry’s hottest buzzword. Speaking on the TBPN podcast on the sidelines of Palantir’s AIPCon 10 conference, Karp compared “tokenmaxxing”—the compulsive overuse of AI tokens—to a porn addiction. His point was simple: a lot of companies are burning through tokens that look like productivity but deliver nothing real.Karp told the hosts that Palantir built an internal tool to wean enterprises off the habit. The name isn’t fit for air, but he described it as a “demasturbatory, get off masturbation” thing. “People are just sitting there all day like a porn addiction,” he said. He kept circling back to the comparison, even as the hosts tried to steer him toward the AI side of the analogy.Tokens are the building blocks of large language models, breaking words into numerical units—a single token is roughly three-fourths of a word. AI providers usually charge based on how many tokens you consume and which model you pick. Over the past few weeks, parts of Silicon Valley have turned hard against tokenmaxxing, the culture that cheered nearly unlimited AI use.
Tokenmaxxing backlash builds as the AI productivity math stops adding up
Companies like Meta and Amazon reportedly built internal scoreboards to track token consumption, treating raw usage as a proxy for productivity. Most have since cut back. Uber COO Andrew Macdonald said the rideshare firm was struggling to connect rising AI bills to any real return. Karp said that until about two weeks ago, executives felt they couldn’t question AI publicly without looking foolish—even as many privately admitted it wasn’t quite working.His CTO Shyam Sankar made a sharper version of the argument on a recent earnings call, calling Palantir a “no slop zone.” More tokens, Sankar said, means more slop.
Why Karp thinks LLMs hit a wall on hard, domain-specific problems
Karp didn’t dismiss AI outright. He called it real and said it’s useful for narrow tasks—writing a report on China’s GDP growth, for instance. The trouble starts with harder, domain-specific problems, like figuring out a cheaper, legal, and ethical way to drill for oil and gas. Those need precise, ongoing processes that LLMs can enhance but not replace, he argued.Of course, much of this doubles as a sales pitch. Karp told prospects to spend a couple of days with frontier companies like OpenAI or Anthropic first—and call him when they’re done, because clients usually come back “clamoring.”