Palantir CEO Alex Karp has intensified the debate over corporate AI adoption by arguing that heavy AI use does not automatically create business value. Speaking in a live interview on TBPN during Palantir’s AIP Con 10 event, he compared excessive AI use to porn addiction and pushed back on the idea that more prompts or tokens should be treated as proof of success.
The warning lands at a time when many companies are chasing higher AI usage as if it were a performance metric. But rising bills and unclear productivity gains have made some executives question whether “tokenmaxxing” is helping businesses at all.
Why companies are rethinking AI usage metrics
The term “tokenmaxxing” has emerged to describe the pressure on employees to use AI tools as much as possible and burn through more tokens. During the AI boom, that kind of activity was often seen as a sign that an organization was moving fast.
That view is now facing resistance as usage-based pricing becomes more common. When AI spend rises without a clear improvement in output, the question shifts from how much AI is being used to what that usage is actually producing.
Uber COO Andrew Macdonald has publicly said the company has struggled to find a clear connection between growing AI spending and meaningful results. Amazon has also reportedly removed an internal AI leaderboard after employees were suspected of inflating their usage numbers.
Palantir’s comments fit into that broader reassessment. The company is drawing a clear line between experimentation and economic value, arguing that the two are not the same thing.
Palantir’s message is simple, no “slop” from excess tokens
Palantir CTO Shyam Sankar set out a similar view during the company’s earnings call last month. He described Palantir as a “no slop zone” and rejected the assumption that cheaper AI or higher token consumption will automatically create better outcomes.
According to Sankar, using too many tokens can increase the risk of “slop,” meaning output that does not deliver real usefulness. He also said that excess “commodity cognition” requires systems that can prevent economic losses, so companies capture value rather than just add operating costs.
For Palantir, the threshold for useful AI is not volume. It is whether the technology is tied to a real business process and can improve it in a measurable way.
AI can help, but it does not replace complex operations
Karp acknowledged that AI models are effective in many narrow tasks. He noted that a model can easily generate a report on topics such as China’s GDP growth.
But he said the limits become obvious when the work involves more complex business operations. Areas such as supply chains, industrial operations, military logistics, manufacturing workflows, and oil and gas operations require continuous decision-making and precise execution.
In those environments, large language models can improve human work, but they do not replace the underlying processes that make decisions and carry them out. That distinction matters because many companies have treated chatbots and generative tools as if they could serve as universal answers.
Taste, not token volume, may decide who benefits most
Karp also argued that many AI capabilities will eventually become widely available. As that happens, the competitive edge will come less from frequent use and more from choosing the right problems to solve first.
He described that judgment as “taste,” meaning the ability to identify where AI can create real results. In his view, that human judgment will matter more than a company’s raw consumption of prompts or tokens.
The message is blunt, but it reflects a growing industry shift. As AI costs climb, executives are being pushed to ask a tougher question: not how much the technology is used, but whether it actually improves operations and business outcomes.
