Palantir Technologies CEO Alex Karp said Wednesday that AI labs are jeopardizing corporate intellectual property by prioritizing token growth over security [1].
This tension highlights a growing rift between the developers of frontier AI models and the companies that use them. If enterprises believe their proprietary data is being used to train models that eventually compete with them, the adoption of generative AI in the corporate sector could slow.
Speaking in a CNBC interview on July 1, 2026 [1], Karp said that many CEOs are "livid" regarding the current trajectory of the industry [2]. He said that AI labs are chasing "tokens"—the basic units of text processed by a model—rather than protecting the business value of their clients [1].
Karp said that "something has gone completely wrong" with the way these models are being deployed [3]. According to Karp, the current business models of AI labs incentivize the extraction of data in a way that threatens the competitive advantage of the enterprises providing that data [4].
Palantir specializes in big data analytics and software, often working with government and corporate entities to secure and organize sensitive information. Karp's comments suggest that the industry is moving toward a conflict over who owns the value generated by AI when that value is derived from private corporate intelligence [4].
He said that the focus on token-driven metrics creates a misalignment between the goals of the labs and the needs of the business world [1]. This misalignment, he said, puts corporate IP at risk as models absorb the very logic, and data, that give a company its edge in the market [4].
“"Enterprises are livid"”
Karp's warning reflects a fundamental conflict in the AI economy: the tension between 'data hunger' and 'data sovereignty.' While AI labs require massive amounts of high-quality data to improve model performance, corporations view that same data as their primary competitive moat. If the industry does not shift toward more secure, isolated environments where corporate data is not used for general model training, enterprises may retreat from using frontier models in favor of smaller, private, and more controlled AI systems.
