He stops mid-sentence on a conference stage, grins, and calls a company feature the “demasturbatory, get off masturbation” thing. The hosts shift in their chairs; a room full of enterprise buyers laughs and winces at once. For a few minutes, the conversation is no longer about models—it’s about appetite.
I watched the clip and I want you to notice what he did next: he named the behavior and turned it into a sales case. That move tells you everything about how Silicon Valley talks about AI when money and ego are in the room.
On stage at Palantir’s AIPCon, Karp likened token play to porn addiction
He didn’t whisper the line. He leaned into it, comparing tokenmaxxing to “a porn addiction” while the hosts tried to steer him back to dashboards. I’m telling you this because the metaphor wasn’t just colorful— it was a signal.
There’s a blunt logic behind the joke: when companies treat model usage as a scoreboard, humans start chasing the number. What looked like experimentation turned into spectacle—internal leaderboards at Meta and Amazon became a public flex until the OnlyFans-sized bills arrived. Now those same firms are pulling back.
At the mic, Karp sold Palantir by framing the problem he criticizes
He repeatedly positioned Palantir as the fixer after the party. I want you to see why that matters before we unpack the pitch.
Karp’s critique doubles as marketing. He praises LLMs for producing “almost right” code—“magical,” he called it—but stresses that domain expertise and secure on-prem patches are different work. That’s the hole Palantir says it fills: not the shiny new model, but the scaffolding that keeps your crown jewels local and functional.
What is tokkenmaxxing?
It’s the informal name for mass token consumption—teams burning model calls to test, toy, or score points. The habit started as curiosity and PR: show investors your usage. Then cost reality hit. When companies traded resource discipline for scoreboard bragging, they got shock bills measured in real cash—not just vanity metrics.
In the trenches, token consumption looks productive but often isn’t
At internal demos, people clap for a prompt that spits out useful-looking code; behind that applause there’s often an engineer refreshing a prompt like someone scrolling for dopamine. I’ve seen that behavior in product rooms and it’s revealing.
Karp’s point: models are excellent at producing plausible outputs that need human domain glue to become safe, repeatable processes. Palantir’s sell is that you need on-prem control, patching, and knowledge retention—services that you pay for with hard budgets. If you’re spending hundreds of thousands of dollars a year—say $200,000 (€186,000)—on cloud model calls and dashboards, you’ve got to ask whether those calls are moving the needle or just lighting up a scoreboard.
Palantir CEO Alex Karp says tokkenmaxxing is like a “porn addiction”:
“We have a product, eternally called something, but internally we call it the ‘demasturbatory get off masturbation’ thing.”
“People are just sitting there all day like a porn addiction.”
“Enterprises are… pic.twitter.com/HQ7N25wY6N
— Jawwwn (@jawwwn_) June 4, 2026
How does Palantir handle token consumption?
They sell governance and on-prem guardrails. Karp named an internal tool with a jokey label and then moved on—because the joke does the work of explaining the product. Palantir’s argument: you can swap any model under the hood—Anthropic, OpenAI—but spotting vulnerabilities is one thing; patching them inside your environment and preserving specialized knowledge is another.
At companies that chased token metrics, the scoreboard became the story
In war rooms where every call registered as a trophy, leaders mistook motion for progress. I’ve sat in those meetings; it feels important when the numbers go up.
That’s where the metaphor matters: token leaderboards can be a carnival mirror, making activity look larger than it is. When the dust settles, many teams find a leaky roof of productivity—lots of noisy fixes, few durable systems. Palantir’s pitch is to sell you the roofers.
You should be skeptical of anyone who critiques a fad and simultaneously stands to profit from the solution. Karp does both—he calls out the masturbatory spectacle and sells the dry plumbing that follows. Sometimes the critique is sincere. Sometimes the critique is a sales demo with charm.
So where does that leave you? You can chase every model, scoreboard, and juice-draining experiment, or you can insist on measurable outcomes, data hygiene, and patchable systems. I’ll wager the latter makes a smarter bet for an enterprise budget than a leaderboard that only impresses investors.
Will you keep feeding the scoreboard or start asking for something that actually holds up under scrutiny?