Alex Karp cites Palantir-NVIDIA push for “secure AI,” says frontier labs fall short on intellectual-property safeguards

Palantir CEO Alex Karp said his company is working with NVIDIA to deliver what he characterized as “secure AI” to organizations that cannot afford to leak sensitive information, including their intellectual property. Speaking with CNBC, Karp framed the effort as a practical response to the way companies adopt frontier AI models, emphasizing control, protection, and contractual safeguards for customer assets.

Karp’s comments also turned into a broader critique of OpenAI and Anthropic, two of the best-known providers of frontier AI models. He argued that those model developers, as characterized in the interview, do not have adequate measures to protect clients’ intellectual property. The thrust of his criticism was less about raw model capability and more about governance and risk allocation when enterprises rely on third-party AI systems.

A key part of Karp’s argument was that “security” for AI is not a single feature that can be added at the end. Instead, he suggested that the full deployment stack matters, including how data flows during use, how models are integrated into customer environments, and what protections exist if a customer’s work or information is used elsewhere. While the interview positioned these points as a differentiator, the underlying technical specifics were not detailed in the reported account.

Karp’s remarks come at a time when AI buyers increasingly demand more than performance. In many industries, decision makers are concerned about confidentiality, compliance, and the downstream use of their content. For enterprises, intellectual property protection is often a central issue because AI systems can ingest or transform proprietary materials, including designs, internal analyses, and plan documents.

NVIDIA’s role in Karp’s framing is tied to its position in the AI compute ecosystem. The graphics processing unit (GPU) and broader accelerated-computing platform NVIDIA sells are commonly used to train and deploy large AI models. In the context of secure enterprise deployments, buyers often look to the full supply chain, including the hardware layer that supports model execution and the software stack that integrates AI into operational workflows.

Even so, the reported account did not provide details about what, exactly, the Palantir-NVIDIA partnership includes, such as whether it involves specific products, software toolkits, deployment models, or security certifications. It also did not specify any named customers, government agencies, or contract values tied to the partnership in the reported coverage.

Karp’s critique of OpenAI and Anthropic likewise was presented at a high level in the account. It did not cite specific policy documents, technical controls, or legal provisions that would show what was missing, nor did it lay out a point-by-point comparison. In other words, the strongest support in the coverage is the CEO’s stated view of the gap in IP protection rather than an itemized audit of the frontier labs’ practices.

Going forward, market watchers are likely to look for clearer disclosure on what “secure AI” means in measurable terms. That could include documentation of security mechanisms, how customer data and intellectual property are handled in deployment, and what protections are contractually or technically guaranteed. For NVIDIA, any additional detail around partner deployments could further clarify how enterprise security requirements are being translated into practical system architectures. For Palantir and its customers, the next test will be whether the claimed IP safeguards translate into broader adoption where trust and compliance are decisive factors.