Alex Karp, Palantir’s CEO, has dropped an estimate that sounds like a declaration of war: artificial intelligence could make him “20 times wealthier” — projecting his fortune toward $300 billion, up from roughly $15 billion today — while middle-class workers might simply see their salaries double over the next decade. He called it “a complete decoupling of unimaginable wealth and normal wealth.” Beyond the stark numbers lies a signal that should put every enterprise investing in AI on high alert: value will not distribute evenly, but will concentrate in the hands of those who own and control the infrastructure.

This is not an isolated prediction. Palantir builds software that analyzes data for governments and large companies, often in scenarios where sovereignty and security are non-negotiable requirements. Karp himself has consistently pushed for operating models that keep data under the client’s control. His remark, therefore, is not just a billionaire’s provocation: it’s the logical conclusion of an ecosystem where AI is not merely a tool, but a mechanism for extracting rent. Those who train the models, who manage inference workloads at scale, who sit on raw data — these players will capture the lion’s share of the surplus generated.

For a company evaluating how to integrate LLMs into its processes, the question ceases to be “cloud or on-premise?”. It becomes: “Who appropriates the value I produce?”. If AI use happens via third-party APIs, every inference, every fine-tuning feeds a system that strengthens the provider. The wealth concentration predicted by Karp is not an abstract phenomenon: it translates into growing margins for those selling access to models and compute, and structural dependency for those buying. It is no coincidence that the shift toward on-premise and self-hosting is involving sectors well beyond the usual banks or government agencies: manufacturing, healthcare, professional services.

The point is not that every organization should buy a GPU cluster. It’s that the decision of where to run the models — on one’s own infrastructure or on a vendor’s — redefines the perimeter of economic sovereignty. An on-premise deployment does not eliminate costs, but shifts the TCO curve radically: upfront CapEx may be high, but recurring OpEx does not scale linearly with usage, and, crucially, data stays internal. This creates an asymmetry: those who choose self-hosting retain control over what feeds the models and what the models produce, avoiding becoming a cog in someone else’s enrichment engine.

Analytical frameworks — such as those developed on /llm-onpremise — help navigate the trade-offs: latency, VRAM management, quantization, hardware maintenance. But the stakes, in light of Karp’s remarks, are no longer purely technical. It’s a choice between two models of digital capitalism: one where AI value condenses in a few hands, and one where enterprises — even mid-sized ones — retain their own share of appropriation. Palantir CEO’s prediction should not be read as a leak about the future, but as an admission of the present: AI’s real payoff is not the doubled salary, but the positional rent of those who control the stack. And that position, for those who don’t build it in-house, is already reserved.