Microsoft CEO Satya Nadella says Copilot already mixes cheaper closed-source and open-source models, while predicting a fierce price battle for everyday AI tasks. His comments put model selection and enterprise-owned evaluation at the centre of how businesses may get value from AI.
Microsoft AI Watch analysis
What happened
In a conversation with 20VC host Harry Stebbings on 26 October, Nadella argued that as capabilities for many routine business tasks level off, cheaper models will take more of the work. BigGo Finance reports that he described Microsoft Copilot as already mixing and matching lower-cost models at scale.
Nadella also argued that companies should build private evaluation benchmarks to compare models against their own business needs, rather than assuming the most capable or expensive option is best. He pointed to Unilever as an example of using frontier models to develop a process before moving claims processing to Microsoft’s lower-cost MAI model. Those are Nadella’s account and argument, as reported by BigGo Finance.
Why it matters
If Nadella’s view holds, the contest may be less about one model winning every task and more about routing work to whichever model offers the best fit for the price. That could intensify pressure on model providers while giving businesses more reason to compare options using their own data and criteria.
A private benchmark is not a magic moat: building one takes work, and the result depends on whether it reflects the tasks a company actually cares about. But it gives buyers a more practical question than “which model is smartest?”: which one performs well enough on our work, at a cost we can defend?
Our read
This is a useful glimpse of how Microsoft says it is using models behind Copilot, alongside a prediction about where the market is headed. The sensible takeaway for businesses is to test models against real tasks and costs, not to treat a provider’s model rankings as a shopping list. The benchmark, not the slogan, should get the final say.
What to watch
- Whether Microsoft shares more detail about how Copilot routes work between models.
- Whether enterprise customers publish results from their own model evaluations.
- Whether lower-cost models can maintain the quality businesses need as they take on more routine tasks.
Discussion spark: Should businesses build and control their own private AI evaluation benchmarks, or are independent shared benchmarks a better way to compare models?
Sources and evidence
not affiliated with or endorsed by Microsoft