Microsoft CEO Satya Nadella once again publicly addressed the risks of enterprises using AI. Foreign media reported that his remarks this time were more direct than those at the beginning of the month: if companies rely long-term on a single AI model provider, they may gradually lose autonomy and ultimately undermine their own competitiveness.
Advocate for data retention and control invocation
In an interview with CNN, he said businesses need to carefully handle the content they provide to model providers—not just training data, but also prompts, context, and metadata generated during usage. According to him, this information should be kept under the enterprise’s own control as much as possible.
Nadella believes that enterprises don't necessarily need to build large models themselves right now, but they shouldn't hand over complete control of critical layers to external platforms. By retaining call logs and related metadata, companies can position themselves to train their own models in the future—or at least preserve the ability to switch vendors.
The model and memory layers should be separated.
He proposed that the model itself should be best separated from the invocation layer, context layer, and memory layer. This way, enterprises can connect to multiple models simultaneously and allocate them according to different tasks; even if a model provider exits, business operations will not be forced to halt.
Nadella also specifically mentioned the coding tools provided by the model providers themselves—that is, the entire development layer wrapped around the model. The article cites Anthropic’s Claude Code and OpenAI’s ChatGPT Codex as examples of such tools.
Worried that a supplier might become a competitor
Foreign media believe this warning is not without Microsoft’s own interests in mind. Microsoft is both a major cloud service provider and a seller of enterprise-grade AI infrastructure. Nadella’s advocacy for multi-model management and isolation of model invocation aligns precisely with service directions that cloud platforms can offer.
However, the article also notes that this assessment is not unfounded. As companies adopt more AI applications, issues related to cost, performance, and vendor lock-in are becoming more pronounced. An increasing number of companies are beginning to focus on open-weight models that can be fine-tuned independently and deployed on their own hardware.
Nadella’s primary concern goes beyond budgeting. He believes that once companies outsource their “thinking process” to model providers, those providers could later leverage these usage scenarios to launch competing services directly. This risk increases further as AI agents begin integrating into enterprise internal processes.
Additional information: The article notes that Nadella’s stance toward general consumers differs from his position in enterprise scenarios. He stated that in consumer internet, users exchanging data for free services is a common model, so this warning is primarily aimed at enterprise customers.