Nadella’s sharpest line cuts against the instinct to chase model selection: “the real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound. You can offload a task, or even a job, but you can never offload your learning.” A task or even an entire job can be handed to an AI, but the learning generated by doing the work is the asset that stays with the firm — “the future of the firm is the ability to compound that learning across people and AI.”
This reframes where to invest: not in the model layer (fungible) but in the retained loop. It is why The system of work is the moat, not the model — the model is fungible underneath and why Persistent agent memory preserves institutional knowledge that walks out the door with employees that would otherwise walk out the door with employees. The compounding nature of retained learning is the same mechanism behind Compound engineering makes each unit of work improve all future work and explains why The context flywheel is a Day 90 moat — Day 0 comparisons are misleading — the value accrues to whoever keeps the loop, not whoever rents the smartest model this quarter.
Nadella names the loop itself — private evals, internal-trace RL, queryable institutional memory — as the firm’s new IP: “This loop becomes the new IP of the firm. I think of it as a hill climbing machine. And unlike most assets, it compounds. Every improved workflow generates better training signal, which accelerates the accumulation of tacit knowledge unique to the firm.” The payoff is timing-sensitive: “The companies that build this early will have an advantage that is hard to replicate, regardless of any new individual model capability.” A better model from a lab does not erase the lead, because the lead lives in accumulated tacit knowledge, not the model. That makes the loop the compounding mechanism behind Proprietary feedback loops create moats that widen with every interaction, and operationally it is the The trace→eval→harness flywheel compounds agent quality — every production interaction generates its own training data elevated to the level of firm strategy.