20VC with Harry StebbingsSatya Nadella: Whether to Trust Frontier Models | Why Companies Need to Own Their Own Intelligence
20VC with Harry Stebbings
Model trust and containment 1:01
Nadella says scaling laws still hold, but he no longer thinks alignment will simply emerge from them. He wants an engineered zero trust setup for AI: give each agent an identity, log every tool and model call, keep the records tamperproof, and monitor behavior against policy. He says chain of thought is not reliable, so trusted partners need inspectable, auditable outputs and multiple-model checks.
Open models and firm value 6:01
He treats AI safety and security as separate problems. Safety belongs in training and pre-training, while firms should build their own trust systems around the model. He expects a mix of frontier and cheaper models, open weights and licensed weights, with most token use becoming commodity work. He also says companies should keep the value from their own data, traces, and private evals, so the surplus stays inside the firm.
Competition and scale 15:01
He expects competition across every layer, from silicon to models to apps, with horizontal specialization and multiple winners people can compose. He says demand for inference is much larger than he had realized and that training runs may reach $100 billion. He thinks the scale will come from world GDP rising, since training spend is only a share of R&D. Power, silicon, and fabs are still shortages, but he expects AI to lift GDP through health care, software, science, and maintenance work.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
