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

Harry begins by asking Satya about a recent piece in which he argued that people should assume an AI model is compromised. Satya clarifies that he still believes scaling laws are producing increasingly capable models, but one assumption he previously held—that alignment would simply emerge from scaling—no longer seems reliable. Mechanistic understanding remains elusive, and alignment may require work during training or even pre-training rather than being added later. His proposed response is not to stop using sophisticated models, but to place them inside an engineered “zero trust” architecture, analogous to zero-trust approaches in cybersecurity. The surrounding system should contain a malfunctioning or misaligned agent before it can cause harm. Satya says Microsoft is building this through products including Agent 365 and what sounds like “Entra” in the transcript: every agent receives an identity, its activity and tool calls are logged, model calls and chain-of-thought units can be recorded, and its behavior is monitored against policy. He distinguishes this external containment and monitoring system from AI safety itself: safety may eventually be embedded in the model, but companies still need an operational trust system now.

The conversation then turns to whether a model’s chain of thought can be trusted. Satya says it cannot automatically be treated as reliable because chain-of-thought explanations can be unfaithful to the reasoning that actually produced an answer. Hiding the reasoning may make the problem worse. He argues for greater standardization and access to chain-of-thought information—at least for trusted partners—so it can be inspected, audited, and checked with multiple models. He suggests that a lack of transparency from closed-model providers could push users toward open weights. Harry asks whether the predominance of Chinese open-source models creates a safety risk. Satya says AI safety is in everyone’s interest: an unaligned open-weight Chinese model would not necessarily limit its harmful effects to the United States, the United Kingdom, or other countries. He is encouraged by progress on Western open-weight models and mentions Jensen Huang and other industry participants, as well as Microsoft’s forthcoming MAI models. Satya says Microsoft may think of these as “licensed weights,” positioning them as a straightforward software business rather than a network-effects intelligence business. Ultimately, he expects adoption to depend on whether users can control and safely operate the models.

Asked to predict where traffic and spending will go over the next 24 months, Satya says the answer depends on what remains at the frontier. Once a task becomes saturated, an open-weight model—or a cheaper closed model—may be sufficient. If a company can train on its own traces, it may not need the newest model to repeat a task whose underlying problem has not changed. Frontier models will continue to attract spending when they provide genuinely new capabilities, scientific advances, or other improvements, but routine token usage will increasingly become commoditized. Satya expects a mix of open weights and inexpensive closed models rather than a simple victory for one side. He points to Microsoft’s own Copilot usage, which combines multiple closed-source cheaper models with open-weight models. He is also enthusiastic about “HydraFusion” and the broader move toward agent swarms and multiple sub-agents, where many local models can handle parts of a task. A router or decision engine could send each task to the best available model. His conclusion is that there will always be a market for frontier systems, but price competition and model specialization will make commodity usage much cheaper and more distributed.

Harry asks about Satya’s argument that companies should own their own intelligence, especially when they have proprietary data. Satya grounds his answer in Friedrich Hayek’s description of the economy as local intelligence: knowledge is distributed throughout firms and markets, and capitalism and price discovery work because no single entity possesses all of it. In an AI economy, each company therefore needs to accumulate and retain knowledge rather than continually leak it to a model provider. Satya calls this the “reverse information paradox.” Kenneth Arrow described how a seller can lose the value of information by revealing a secret without patent, trade-secret, or copyright protection. With AI, Satya says the buyer—the company using the model—also needs protection because its prompts, usage patterns, thumbs-up and thumbs-down signals, and other exhaust can create value for the model provider. He expects each firm to develop a private evaluation benchmark containing its core intellectual property. Models would then be tested against that benchmark, while the firm captures more of the value created by its own use of intelligence. This intelligence need not be embedded only in model weights; it can also live in context, evaluations, and company-specific workflows.

Harry asks whether private evaluations and AI-native practices will remain limited to unusually forward-thinking companies. Satya says they will spread, comparing the transition to the adoption of PCs, spreadsheets, email, and cloud computing. In the late 1980s and early 1990s, using spreadsheets to forecast or replacing fax machines and interoffice memos felt exceptional; over roughly a decade, those practices became ordinary. He expects today’s AI-native companies to seem unremarkable within about five years, just as virtually every company—including the 500-year-old Lloyd’s—is now cloud-native. The pair then discuss whether this transition will create a large implementation economy. Satya says systems integrators have always helped diffuse technology, but AI implementation will combine business and technical work rather than fit neatly into classic management consulting, systems integration, or technical support. He cites Unilever’s effort to automate order-to-cash and trade-claims processes with an agentic system. The company began with a frontier model, then substituted a lower-cost MAI model for trade claims because it could control the model and train it on its own traces. Microsoft initially supplied support, but Unilever’s own IT team now manages the system. Satya expects new implementation skills and a new class of specialized practitioners to emerge.

Asked whether companies must own the entire stack—from chips through models to applications—Satya rejects a universal rule. He expects perhaps five or six major closed models and a similar number of open-weight models, allowing application builders to use routers and decision engines to select the best model for each task. Although he acknowledges the possibility of a vertically integrated winner, he says competition is currently increasing in silicon, silicon-management systems, models, and applications. His preferred outcome is horizontal specialization, with multiple winners at each layer and the ability to compose products across them. Satya says the major thing he failed to anticipate 12 months earlier was the “infinite demand for inference,” a realization that would have changed some of his investing decisions. Harry then raises claims that training runs costing $100 billion may arrive and asks whether energy, chips, fabs, and power generation can support them. Satya says such spending must ultimately be justified as a percentage of research and development and by growth in world GDP. Supply constraints are real because power plants and fabs have long lead times, but companies are building in response to demand. The timing of supply catching up with demand remains unknowable.

Satya argues that the key economic question is whether AI can increase world GDP growth. He sees major opportunities in sectors such as healthcare, citing Microsoft Copilot’s deployment in the NHS as an example: better patient outcomes and lower costs would directly benefit the United Kingdom’s economy. He also sees software development as highly elastic because large IT backlogs reveal that demand for development was much greater than people realized. Science and engineering may show similar elasticity. Beyond coding, he points to silicon design, faster tape-outs and simulation, and Rolls-Royce’s optimization of maintenance scheduling as evidence that AI can address long-standing operational problems. He says finance, sales, and marketing may improve, though knowledge work is harder than coding because it lacks the same straightforward reinforcement signals. In the quickfire round, Satya says he has changed his mind most on alignment: in 2018–19 he joined the scaling-laws believers, including Dario, Ilya, Sam, Greg, and Elon, and expected alignment to emerge from scaling; he no longer believes that. He predicts that companies having private AI evaluation benchmarks will look normal within five years. For exercise, he walks because his knees no longer allow him to run and tries to get half an hour in each morning, using a gym even while traveling or jet-lagged. The transcript does not include an answer to the advertised cheat-meal question. Asked what excites him most over the next decade, he hopes AI diffuses faster and more broadly than previous general-purpose technologies, lifting entire populations rather than merely creating winners and losers, especially by accelerating convergence in the Global South.