Palantir CEO calls OpenAI, Anthropic token model 'completely wrong,' 'insane'

Alex Karp, CEO of US data analytics firm Palantir, has leveled sharp criticism at closed AI companies including OpenAI and Anthropic.

Karp appeared on CNBC on Tuesday and said, "I'm not trying to put them down, but something is completely wrong."

He said American corporate executives are frustrated at paying for tokens that generate no value, adding that closed AI models have been "fundamentally, completely irresponsibly oversold." He said US enterprise customers believe these AI companies will leave them "wasting time on tokens, doing nothing, and having my intellectual property taken without getting any value."

Karp expressed particular concern about the US Defense Department's reliance on closed AI infrastructure. "Are we going to leave the battlefield of this country to the mainstream opinion of Silicon Valley?" he said. "That is insane." When the interviewer suggested he seemed angry, Karp rebuffed the characterization. "No — this is the voice of American companies coming through me," he said.

Karp's broadside against major AI companies is widely seen as an effort to highlight Palantir's own work building open-source models.

Palantir has recently partnered with Nvidia to build open-source custom models that run in "sovereign" environments. The company said the approach allows customers to maintain control over their data, intellectual property and AI systems while owning their own models tailored to their specific operations.

"What connects me and Nvidia, and what technology customers want, is for them to have control over compute, models, data and alpha — their objectives," Karp said.

In practice, the AI industry is shifting away from so-called "token-maxing" — the tendency to consume tokens without limit — toward a greater focus on return on investment, as the cost of running AI models continues to rise.

Even Amazon, the world's highest-revenue company, is reportedly weighing greater use of its own models such as "Nova" after facing rising costs when its contract with Anthropic shifted from a "compute time" basis to a "token" basis.

A token is the basic unit of computation — and the primary billing unit — for language model-based AI. In English, one word corresponds to roughly 1.3 tokens.