20VC: Satya Nadella on Open vs Closed Models: Who Wins | Whether to Trust Frontier Model Providers | Why Companies Need to Own Their Own Intelligence | The Importance of Verticalisation: From Chips to Models and More

Satya Nadella · Microsoft

Every firm should own its intelligence through private evals and data, or model companies will capture the value.

Nadella draws on Hayek, who saw distributed tacit knowledge as the basis of an economy, and on Arrow's information paradox: a seller who discloses a secret without legal protection leaks its value. With AI, the buyer now needs protection, because its data and usage signals, such as thumbs up and thumbs down, create value for the model company. For that surplus to stay in the firm, it must sit in context, weights or evals. He is 'very gung ho' that every firm will have a private eval benchmark as its core IP, and that every new model will try to climb it. He says this is how a firm works, and that the approach has spread before, as PCs and then cloud did.

“one of the things that I'm very gung ho about is we all become, every firm will have a private eval benchmark, which is their core IP,”at 12:14

Satya Nadella · Microsoft

AI deployment needs a new hybrid of business and technical consulting, distinct from classic systems integration.

Nadella says systems integrators have always existed to drive technology diffusion, but AI deployment has different contours. It is not classic management consulting, systems integration or technology support. It combines business and technical consulting. In the Unilever example, the work started with the business problem, the eval and the agentic architecture. Some FTEs were deployed, but Unilever's IT team now runs the system. He expects a new set of skills to form in the ecosystem, and traditional systems integration to turn into this new kind of FTE work.

“So it's sort of business consulting combined with technical consulting. And I think, yeah, you're right in saying that there will be a new set of skills that'll get built in the ecosystem.”at 15:12

Satya Nadella · Microsoft

Alignment won't emerge from scaling, so agents need an engineered zero-trust containment layer outside the model.

Nadella says his earlier assumption, that alignment would fall out of scaling, no longer holds: 'it doesn't seem to be,' and mechanistic understanding is elusive. He argues that alignment likely has to be embedded in the weights during training or pretraining, and cannot be reached through RL afterward. Since capable models may be less aligned, he proposes treating them inside a zero-trust architecture that can contain any malfunctioning agent before it does harm. Chain-of-thought is unfaithful and can get worse if model companies hide it, so inspectable, auditable reasoning from trusted partners matters. He also says AI safety failures are bad for China, the US and the UK alike, since an unaligned open-weight Chinese model could also harm China.

“So the claim, at least I'm making, is let's create an engineered zero trust environment that can contain any malfunctioning agent before it does any harm.”at 4:14

Satya Nadella · Microsoft

Most future token usage will be commodity, served by cheap open or closed models rather than frontier models.

Nadella expects most models to saturate, after which open-weight or cheaper closed models will do the job. Firms can train on their own traces, so they won't need the latest model to solve a problem that hasn't changed. Frontier models will still have a market for new problems such as science, but predominantly token usage will be commodity usage with heavy price competition. Inside Copilot, Microsoft already uses many cheaper closed models alongside open-weight ones. Demands for transparency and auditability of chain-of-thought may also push more open weights. He doesn't give a split of traffic between open and closed.

“But I think predominantly token usage will be commodity token usage.”at 10:02

Satya Nadella · Microsoft

Full vertical integration from silicon to apps is unlikely; competition is intensifying at every layer of the stack.

Nadella says he once thought one vertically integrated company might control everything from models and apps down to silicon and systems, but he now thinks that is unlikely. Instead, competition is growing at every layer: silicon, systems, the model layer and the app layer. He expects roughly five or six closed and five or six open-weight models to compete. A builder can route each task to the best model through a decision engine. He concedes the eventual outcome depends on which dominant design emerges.

“If anything, it's going the other way, which is every day you wake up and you say, Wow, there's more competition in every layer of the stack, whether it's in silicon, whether it's in systems”at 16:28

Satya Nadella · Microsoft

Power and chip-fab capacity are long-lead constraints, so supply may lag AI demand, with timing unknown.

Nadella says the world has power shortages and silicon shortages. These are long-lead items: you cannot snap your fingers and get power generation or a new fab. Given how much demand everyone is building for, the open question is when supply comes online, and no one knows the demand-supply balance. He argues the constraint isn't a limit on what companies can spend on training, since spend scales with revenue. Revenue depends on AI actually contributing to economic growth, so the binding limit is physical supply.

“They're long lead items. You just can't snap your finger and have power generation, or you can't have a new fab.”at 18:12

Satya Nadella · Microsoft

AI could meaningfully lift world GDP growth above the roughly 2% trend, since it adds to productive output.

Nadella notes that world GDP growth has been stable at about 2% a year for a very long time. He thinks AI could inflect that rate in a short period, and says he hopes it does. His argument is that AI spend should be a share of GDP, as R&D is for a company. Using healthcare as an example, he says a Copilot deployment that improves patient outcomes and cuts cost helps UK GDP by definition. He is hedged, and the view depends on adoption reaching sectors where costs are high.

“Does AI have the ability to to seriously inflect that in a short amount of time? I think so, and I somewhat hope so because that is the point.”at 18:33

Satya Nadella · Microsoft

AI displaces the least and benefits the most in pockets like healthcare and software, where demand is elastic.

Nadella argues that AI adoption should start in the economic pockets with the least displacement and the most benefit. Software development is one example: demand for it is very elastic, and IT backlogs were large, so cheaper development expands work rather than cutting it. Coding has been the easiest domain because verifiable rewards allow RL on top of pretrained models. Knowledge work is harder to crack, but he thinks contact centers, supply chains, order-to-cash and science may have elastic demand too. He points to the NHS Copilot deployment as a case where AI improves outcomes and cuts cost.

“Really, in fact, there's the least amount of displacement, the most amount of benefit.”at 19:39