Possible: Satya Nadella on making human and token capital compound

Satya Nadella argues that AI is not merely a technology but the future of the firm itself, and that most CEOs have not recognized this shift, remaining satisfied with press releases and a handful of agents as a substitute for strategy. The central challenge he frames is compounding returns from both human capital and token capital, where token capital is the accumulation of knowledge from operations that a firm owns, controls, and has put in place a system to compound, taking the form of context, skill, or model weights.

Nadella warns that enterprise tacit knowledge can be extracted and encoded into model weights through human work trajectories, and that most enterprises carry no balance sheet line item for this asset and therefore do not recognize the risk of losing it. He identifies a specific leakage mechanism: model companies use employees who previously worked at client companies as raters or judges, transferring enterprise-specific knowledge to external parties in a one-way and permanent process. His prescription is that every enterprise must run models inside infrastructure it controls so that human-agent work trajectories remain proprietary and continuously retrain internal models.

Knowing what data to train a model on and how to reward it is where the next level of IP gets created, because everything else is mechanical. Nadella describes the new core capability as a hill-climbing machine that takes an objective, an outcome, and an eval and learns to achieve it using data and reinforcement reward. He states that frontier models should not be used for non-frontier problems, and that a small model like MAI 5B using traces to hill climb in a reinforcement learning regime can outperform even a frontier prompted model on repeatable deterministic workflows such as processing trade promotion claims for retailers. Evals and rubric scoring dimensions must capture high taste that only humans can define in order to achieve token efficiency, which he identifies as a key source of competitive advantage.

Microsoft Build was framed around helping every developer and enterprise build their own hill-climbing machine. Nadella introduced Agent 365 to provide inventory, reasoning trace inspection, auditability, identity management, sandboxing, and policy governance for agents, and extended Entra for agent identity, Defender for agent security, and Perview for automatic data labeling and protection. Microsoft added a feature called Asserts to Foundry to enforce execution boundaries for long-running agents, going beyond simple guardrail classifiers. He also introduced cognitive coverage, a GitHub Copilot skill that creates a quiz for humans to learn from what an agent did, which Nadella views as a critical human skill to develop in an agentic era. Microsoft envisions two million or 20 million agents operating alongside its approximately 200,000 employees, with tacit knowledge emerging from the continuous interplay of humans and agents.

Agent runtime workloads have fundamentally different call patterns compared to training and inference workloads and did not exist at scale three years ago. Microsoft is optimizing its ARM-based Cobalt chip using agentic traces from GitHub coding workloads and co-designing its Maya chip with MAI and OpenAI models. Microsoft will benchmark all chip options and remain open to outside innovation rather than committing exclusively to Maya or Cobalt, while also using older NVIDIA chips to accelerate its Fabric data warehouse. At the infrastructure level, Microsoft is redesigning data center civil engineering, cooling, and power distribution with an ambition to deliver power straight from kilowatt-level input to silicon with minimal distribution losses.

Nadella identifies losing social permission as the risk he is most concerned about, pointing to a college commencement speaker being booed for promoting AI as a signal that public trust has eroded. He argues that abstractly invoking the lump of labor fallacy is insufficient and that the industry must identify specific new jobs, wages, and training pathways. He frames child safety as a first-class AI safety concern alongside cybersecurity, bioweapons, and alignment, warning that chatbot interactions with children risk undermining their agency by persuading rather than empowering them.

On the geopolitical dimension, Nadella draws on a thousand-year comparative history of China and the West to argue that Western prosperity resulted from a virtuous cycle of moral philosophy, political systems, markets, and scientific revolution, not technology alone. He warns that convergence growth for the global south could slow or reverse in the age of AI, but holds out the possibility that if AI reaches all corners of the world simultaneously, every country could express its comparative advantage fully, creating a positive-sum outcome. He predicts that if everything breaks humanity's way, the world could compound at 10 percent GDP growth over the next 15 years. Reid Hoffman separately disclosed that Manus is producing chemistry results that top computational chemists describe as novel and potentially viable against cancers, while cautioning that the results are very early and discoveries take time to develop.