According to a report on TechCrunch, among all the heated debates surrounding the potential downsides of artificial intelligence technology, one central concern is causing the most hand-wringing among tech enthusiasts in Silicon Valley. Their greatest fear is that giant AI labs selling proprietary and commercial models are essentially acting like "Trojan horses" inside the companies purchasing their services.
The core concern is that as startups and enterprises utilize AI models from leading labs like OpenAI or Anthropic, these labs gain ever-increasing access to those clients' most sensitive business information. The model developers can subsequently use this knowledge and insights for their own purposes, effectively turning them into direct competitors of their own customers. Prominent figures who have issued similar warnings in the past include venture capitalists like Jason Calacanis and Palantir CEO Alex Karp. Now, in a surprising blog post published on Monday, Microsoft CEO Satya Nadella has joined these warning voices.
Satya Nadella's Warning: Paying Twice for Intelligence
In his blog post, Nadella warns that AI users (whom he refers to as the "buyers") are actually paying a double price to use the technology. On one hand, they knowingly pay for the model's token usage, and on the other, unconsciously, they hand over valuable organizational data in the process.
Nadella explained this in his own words: "You essentially pay for intelligence twice – once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!"
The greatest danger, according to Nadella, is that enterprises are literally teaching these models the finest nuances of their business activities. Nadella writes that the models learn from the "exhaust" of usage—meaning the prompts users write, the tools AI agents use, and particularly the corrections people make when the model produces an incorrect result. Every such correction is distilled into collective institutional know-how within the model. Nadella emphasizes that this is the kind of knowledge a competitor could never buy with money, yet organizations are currently handing it over voluntarily to proprietary model developers.
The Issue of Model Distillation and Licensing Restrictions
Nadella argues that if AI companies have the full right to freely scrape the public internet to train their models, then it is only fair that organizations and enterprises can study—or "distill"—these models in return. The term "distillation" refers to the technological practice of using a large model's outputs to learn how it works and to train a new, usually smaller and cheaper model based on those insights.
To illustrate the complexity of this issue, in February, Anthropic accused Chinese open-source models of sending millions of prompts to its model, Claude, to improve their own independent models, and even urged the United States government to tighten export controls on the matter. Nadella's perspective is that model makers cannot have it both ways. He finds great irony in the fact that the status quo allows model developers to enjoy fair use rights to train their models on public data, but then they immediately turn around and impose restrictive terms on the distillation of their models by others. Nadella expresses particular concern over cases where model developers reserve the right to learn from their customers' usage and interaction data.
Proposed Solutions: Private Clouds and Orchestration Layers
The solution Nadella proposes is what one might expect from the CEO of a giant cloud provider. He calls on companies to retain full ownership of their data, including prompts, feedback, and any other data transmitted to the model. To achieve this, he urges them to build their own "proprietary learning environments" on the cloud (where their data is likely already stored, and conveniently for Microsoft, this could point to its cloud services, Azure).
In addition, he encourages companies to build systems that include "orchestration layers"—a method that allows for easy and rapid switching between AI models from different providers instead of being completely locked into a single model developer. Tools such as "AI gateways," which enable companies to perform exactly this type of routing, have recently been gaining popularity in the market. Although Nadella does not explicitly use the term "open-source" as a way to maintain ownership, this is the clear subtext of his remarks.
The Trend on the Ground: Shifting to Open Source and Local Deployments (On-Premise)
Alongside Nadella's words, there is another subtext playing out in the enterprise market. Many large companies, some of which still maintain independent data centers in addition to using cloud services, have already begun shifting to using open-source models installed locally on the company's servers (On-Premise).
Idit Levine, founder and CEO of Solo.io—which develops security and networking software that helps enterprises manage AI systems—notes that she is witnessing this exact shift among her clients. According to her, after companies experiment with initial work alongside proprietary model developers, they begin asking themselves whether they can take an open-source model and run it locally on their servers. In an interview with TechCrunch, Levine explains that companies realize such a model will deliver nearly 90% of the capabilities of the large model, at a much lower cost, while giving the company complete and secure control.
Solo.io's technology was selected last year to serve as the technological foundation for the Linux Foundation's Agentgateway project. The company's clients include leading large corporations such as T-Mobile, ADP, and SAP. Levine identifies a growing trend of installing local open-source models and views this as the next big wave of AI adoption in enterprises.
Quantitative Data Supporting the Shift in Trend
Levine's observations and assessments are strongly reinforced by other players in the AI industry. Vercel—primarily known as a platform for building and hosting websites, which recently added AI model switching and routing tools—and OpenRouter, which helps developers route requests among different models, both report a significant surge in traffic directed to open-source models.
In fact, open models accounted for no less than 29% of all traffic routed through Vercel's gateway during the past month. Considering that the CEO of Microsoft—a company that has itself invested massive sums in both OpenAI and Anthropic—is now openly calling on enterprises to be wary of using proprietary models and to protect their data, this trend is highly likely to continue expanding. As Nadella summarized in his post: "In consuming intelligence, you are creating intelligence. And what you create should belong to you."