Palantir CEO’s Critique of AI Giants Prompts Banks to Reassess Deployments

Closed AI systems from OpenAI and Anthropic are typically accessed through APIs and priced by tokens, keeping model weights under vendor control. Palantir Technologies CEO Alex Karp says that arrangement can leave companies paying for usage without measurable returns while exposing proprietary data, workflows and “alpha” — the know-how that creates competitive advantage. His alternative is to deploy open-weight models within an enterprise-controlled stack, allowing customers to retain control of their compute, models and data as AI becomes embedded in core operations.

On July 1, 2026, Karp used an appearance on CNBC’s Squawk Box to accuse OpenAI and Anthropic of encouraging token consumption while customers received little value, calling frontier models irresponsibly oversold. On July 24, Palantir joined Nvidia, Meta, Microsoft and 21 other organizations in an open-weights policy letter. The signatories said distillation is a longstanding research technique that should not automatically be treated as intellectual-property theft, reinforcing support for open models voiced by Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg.

Palantir Technologies supplies data-integration and artificial-intelligence software to governments and businesses, positioning its platforms as a way for customers to deploy AI while retaining control of proprietary data and operational systems. Chief Executive Alex Karp has increasingly challenged frontier AI labs, arguing that large-language-model developers may use partnerships to move into customers’ most valuable businesses and ultimately control their “means of production.” The dispute highlights growing concern over data sovereignty and bargaining power in enterprise AI.

Palantir said on Aug. 3 that second-quarter revenue surged 93% from a year earlier to $1.935 billion, while net income climbed to $1.07 billion, or 41 cents a share, from about $329 million, or 13 cents. U.S. commercial revenue jumped 149% to $764 million in the quarter ended June 30, as demand for AI software accelerated. In his latest shareholder letter, Karp called some frontier labs untrustworthy and described their business ethos as “Marxist,” accusing model developers of seeking to capture partners’ production infrastructure.

Companies deploying large language models typically pay according to the number of input and output tokens used. Palantir argues that the real value lies in an “ontology layer” that connects models to corporate data, permissions and workflows. The architecture gives organizations control over their computing resources, models and data, which is particularly important for governments, defense organizations and regulated industries.

Palantir and NVIDIA unveiled a reference architecture for a sovereign AI operating system on June 29, 2026, integrating Blackwell Ultra, Nemotron and Palantir’s ontology layer. The value of the partnership was not disclosed. In a July 1 CNBC interview, Alex Karp criticized token-based pricing, asking why a model that could create $1 billion in value should not charge a 30% share of the outcome. He also said companies were concerned about data and intellectual property leaking outside their organizations.

Palantir has long focused on bringing AI and data analytics into government and corporate operations, leading CEO Alex Karp to emphasize products that can be deployed immediately to solve existing problems. He argues that most Silicon Valley AI companies are obsessed with sweeping visions while overlooking customers' practical demands on cost, results and deployability.

Karp has again criticized AI companies and their leaders for failing to recognize how negatively they are perceived. Their products are not only expensive but also frequently perform below expectations, he said. He singled out OpenAI's deployment model as a “complete farce,” arguing that the industry should prioritize practical AI. The report did not specify when he made the remarks or provide figures for contract values or product costs.