- Microsoft CEO Satya Nadella warned that businesses using proprietary AI models may be paying twice: paying for the tokens consumed, while inadvertently feeding their valuable internal knowledge to AI model developers.
- According to Nadella, for AI to work effectively, enterprises must share significant information about their processes, data, and operations. These user prompts, feedback, and corrections gradually form organizational knowledge that competitors can hardly buy.
- Nadella believes that every time a user fixes an error or guides the AI, they are transferring experience and corporate knowledge, creating an extremely valuable data source.
- He also pointed out the contradiction where AI companies argue that scraping public data on the Internet to train models is fair use, but strongly object to other parties using distillation techniques to learn from the outputs of those very models.
- Nadella is particularly concerned about clauses that allow AI providers to learn from the usage data and interactions of corporate clients.
- Microsoft proposes that enterprises retain ownership of their data by building private AI environments on cloud platforms, while deploying an orchestration layer to remain flexible in switching between various AI models instead of being locked into a single vendor.
- The article also notes that many businesses are shifting toward on-premises open-source AI models to reduce costs and better control internal data. According to Vercel, open-source models accounted for 29% of traffic passing through its AI gateway platform last month.
- Nadella concluded that when businesses use AI, they are also generating new knowledge. In his view, this knowledge must belong to the enterprise itself rather than becoming the property of AI model providers.
- 📌 Conclusion: The AI debate is shifting from consumption costs to knowledge ownership. Satya Nadella warns that businesses are not only purchasing AI capabilities but are also inadvertently sharing valuable operational experience and data with model developers. He recommends building private AI environments, utilizing an orchestration layer to avoid vendor lock-in, and considering open-source AI models to maintain long-term control over data and internal knowledge.
