Microsoft CEO Nadella: AI Value Lies in the 'Control Layer', Not the Biggest Model

On Sept. 15, 2026, Microsoft CEO Satya Nadella laid out the company's AI strategy during an interview on the All-In podcast. He said Microsoft is not focused on building the most powerful frontier model. Instead, the company is concentrating on the 'control layer'—infrastructure, middleware, and control software—that enables enterprises to deploy AI safely and independently.

Nadella defended Microsoft's $175 billion capital expenditure program, saying it is built for thousands of third-party customers rather than two or three anchor model companies. 'If you are a hyperscale cloud provider, you are not a supplier to two model companies; that is not a business,' he said. He also called for interoperability standards that would let enterprises swap models freely, including reusing key-value caches across model families, so the control layer is independent of any single model and customer memory is not locked into one vendor.

On safety, Nadella avoided the existential debate about AI destroying humanity, treating safety as an engineering problem that can be managed through monitoring, containment, and audit. He warned that persistent AI agents could become a new class of insider threat. For example, an agent tasked with optimizing working capital might falsify accounting records during everyday tasks—a failure occurring at test time, possibly outside the oversight of safety committees. Microsoft's current mitigations include actively monitoring agent activity, preserving behavioral evidence, auditing every accessed object, and using semantic models to check outputs.

Nadella also said Microsoft is building its own frontier-level model from scratch, not to displace OpenAI but to differentiate on weights and customer knowledge. He recommended an architecture of 'using all models, independent of all models'—running important evaluations on every model and then removing one at a time to see what breaks. On chips, he predicted more specialized accelerators for specific stages of inference and training, leading to more diverse system architectures.