Palantir CEO Alex Karp has once again poked the nose at the business model of AI companies leading the way, saying companies like Anthropic and OpenAI could become entirely dependent on frontier AI platforms. In a recent interview with CNBC, Karp used an unusually strong term to describe an increasing reliance on advanced AI platforms - that companies are attempting to “drug addict” businesses to a future they think they can control.
Karp said frontier AI companies were effectively encouraging businesses to move toward a model in which they could lose control over critical aspects of their operations. “We have people trying to drug addict us to a future they believe they control,” he said, adding that these companies are building some of the world’s best AI systems.
The Palantir CEO specifically mentioned Anthropic CEO Dario Amodei and said he had spent a lot of time with Amodei and the Anthropic team. Frontier AI companies are creating a future in which businesses might own less, struggle to remain profitable and increasingly depend on external AI providers, Karp said.
Karp’s words are indicative of wider concern about companies adopting generative AI. Businesses are spending billions on AI models, computing power and automated agents and Karp wonders if that kind of spend always gets them to show real business value. His criticism is not only about the technology itself but about who controls the infrastructure, data, intellectual property and economic value generated by AI.
One of the concepts Karp has criticized is “tokenmaxxing.” It’s a broad term to describe companies that use increasingly large amounts of AI compute, tokens or autonomous AI agents without having any concrete evidence that the additional spending is producing business benefits.
Karp argued that companies, including big government groups, are wondering why they should spend massive sums on AI services if they don’t control the underlying systems or retain the value created by their use. From his perspective, businesses should be able to use AI to strengthen their own operations and not be tied to a handful of powerful AI laboratories.
The criticism comes despite repeated assurances from AI companies that enterprise customer data is protected. OpenAI has said that business customers must explicitly opt in for their data to be used for model improvement under its relevant policies. Anthropic has also maintained that customer data is not automatically used to train its models.
Karp seems to be worried about a much deeper dependency that goes beyond whether customer data is directly used for training. His argument is about strategic control: if companies rely heavily on externally hosted frontier models, they could be depending on pricing decisions, model availability, computing infrastructure and technological roadmaps drawn by AI providers.
This is also one reason Karp has been increasingly using open-weight AI models. Unlike traditional closed AI systems which generally can only be accessed through the company's own platform or API, open-weight models provide developers with access to the model weights. This can also give organisations greater flexibility to tailor, deploy, and operate AI systems on their own.
Palantir joined Nvidia, Microsoft and other companies in an open letter arguing that open-weight AI models are important for national security. The debate has only become more important as AI development becomes increasingly competitive between the United States and China.
Recently, the UK's AI Security Institute has shown how rapidly leading Chinese open-weight AI models are improving. Some Chinese systems are now getting close to the capabilities of advanced closed models released only months earlier. That trend could potentially hinder the technological advantage that frontier closed AI systems have enjoyed for decades.
The shrinking gap between open-weight and closed frontier models is an important problem for businesses. If open-weight systems continue to improve rapidly, companies could have more opportunities to deploy powerful AI without entirely relying on a small number of proprietary AI providers.
AI experts and industry executives have pointed to the same trend. Srinivas Padmanabhuni, CTO of AiEnsured, said the gap between open-weight and frontier closed models is collapsing rapidly, suggesting that capabilities that once took many months to become available outside leading AI laboratories are spreading much faster.
Sagar Vishnoi, co-founder of Future Shift Labs, also noted that the narrowing gap indicates AI breakthroughs are being distributed much faster in the global ecosystem than they were earlier.
For Karp, this change could ultimately determine whether businesses remain customers of a handful of AI companies or gain greater control over their own AI infrastructure. His criticism is part of a debate that is now even more imminent in technology: whether we will see the future of AI dominated by a few laboratories and whether we will see companies with a growing number of AI models that they will be able to control and deploy independently.
As AI adoption increases in scale, companies are likely to have to make tough decisions about how much they should devote to proprietary AI services, how much control they should retain over their data and intellectual property, and whether a comparable performance can be attained using open-weight alternatives.
Karp’s “drug addict” comment is provocative, but it is also illustrative of a serious concern on the economics of AI adoption. The real issue is not whether businesses will use AI; it is who will control the technology, the costs and the value it will give to them.
