Palantir CEO Alex Karp publicly criticized the recent Silicon Valley trend of "tokenmaxxing," arguing that continuously increasing AI usage does not equate to generating real business value. In an interview during Palantir AIP Con 10, he stated that the market has shifted from debating whether AI is real to recognizing that while AI is effective, many use cases are not functioning as expected.
The dispute points to high-energy consumption usage.
A token is the basic unit of text processed by large language models, and AI service providers typically charge based on token consumption. Over the past few weeks, some professionals in Silicon Valley have begun reflecting on the "tokenmaxxing" culture—the practice of expanding AI usage with little to no restraint in an effort to keep pace with the development of AI agents.
Karp’s view is that more tokens often mean more low-quality outputs rather than higher-value results. Last month, Palantir’s Chief Technology Officer, Shyam Sankar, expressed a similar perspective on the earnings call, stating that the company internally emphasizes a “no slop zone” and opposes treating cheap model calls as value in themselves.
Palantir emphasizes systems over model stacking.
At the time, Sankar stated that cheaper AI alone does not automatically lead to higher returns; companies still need systems like Palantir AIP to connect model capabilities with real-world business environments and prevent economic losses caused by incorrect outputs.
Karp further stated in the latest interview that the real challenge is not getting the model to generate generic content, but embedding AI into continuously operating business processes. For example, a large model can fairly well produce a report on China’s GDP growth; however, when it comes to complex tasks such as oil and gas extraction, supply chain adjustments, defense manufacturing, or automobile production, AI itself cannot replace the specific workflows.
Complex operations still require ongoing execution.
He believes that such issues often involve cost, compliance, ethics, and operational details simultaneously, requiring precise and ongoing processes. Large models can enhance these processes but cannot directly replace them.
Karp also noted that the industry is increasingly recognizing that while AI’s capabilities have been validated, the key for businesses to truly turn it into commercial outcomes lies not in endlessly scaling model usage, but in clearly understanding what business problems they aim to solve and how to integrate the model into actionable systems.