What Nadella’s manifesto skips
He named the new IP of the firm. The mechanism that turns it from vision to jobs sits underneath.
Last Sunday, Satya Nadella published a long post on X titled “A frontier without an ecosystem is not stable”. The core of the argument fits in a single sentence: “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”.
Nadella is right about the loop. He presents it in the manifesto as a competitive strategy. Every firm should own its loop; otherwise, value flows to the few companies that own the models. The framing is strategic, the audience is the CIO (and CEOs), and the call to action is to build infrastructure, competence, and processes around AI rather than buy it off the shelf.
The manifesto does one thing well and skips one thing entirely. It names a real asset, but it does not explain why that asset is real in the first place. Not the strategic reason, the structural one. Why is owning a loop an actual asset, and not just a managerial aspiration?
On April 27, 2026, I submitted a paper for publication that discusses these questions. The paper was written from the perspective of LLM epistemology, not from the economics of the AI market. It focuses on Software Engineering, and within it on coding agents, because that is where a structural vacuum is most visible (and where most of the media attention is concentrated).
But the proposed reasoning generalizes well beyond software development. It is based on three structural properties of LLMs, four levers, and three key roles.
These three properties are architectural, not application-specific, which is why the framework derived from them applies to any work in which an organization wants to govern an agent rather than supervise it on a case-by-case basis. The tradition the paper rests on is decades old; what the paper adds is a structured account of why the architecture itself requires external knowledge by construction, not because of the current state of the technology.
The mechanism Nadella names but does not explain
Three properties of a large language model, taken together, produce the condition that makes Nadella’s loop necessary. Once they are on the table, the loop stops being a generic declaration and becomes a description of the work that needs to be done.
The first property is probabilistic generation. An LLM does not compute; it generates. Every output is a navigation over a probability distribution conditioned on training and context. The model carries no internal reference for whether its output is adequate. Correctness lives outside the model, as a specification, a test, a domain constraint, or an organizational rule. Any judgment of adequacy requires an external reference. The model cannot supply it. No improvement in scale changes the situation, because what is missing is not training data but a reference against which the output can be checked. That reference is rarely fixed: it varies by context, evolves with the project, and has to be negotiated among the people who hold the relevant knowledge.
The second property is agnosticism. The model has absorbed text and code from the public body of knowledge. It has not absorbed organization-specific knowledge: the risk models of this bank, the legacy constraints of this system, the tacit understanding that has never been articulated, not even by the people who hold it. Part of this can be reduced with training, retrieval, and fine-tuning. Part of it cannot. Organizational knowledge is continuously renegotiated, semantically defined by the practice it describes, and permanently inconsistent in functional ways. Training presupposes a stable ground truth that organizational knowledge lacks. A model that forces this material into a consistent representation assumes a coherence the organization itself does not possess.
The third property is statelessness. Within one session, the model retains a working footprint. Across sessions, it retains nothing. The architectural decision made last Tuesday, the domain constraint introduced by the expert last week, and the trade-off accepted last month: none of them survives a session boundary unless they have been externalized as a persistent artifact. New LLM mechanisms being introduced extend the ability to retain knowledge, but they cannot entirely fill the gap or provide structure and governance.
None of these three properties can be eliminated by more training data or by larger models (scaling). The vacuum is a condition of the conceptual architecture, not a gap that closes with scale. Each property resists reduction for its own reason. Probabilistic generation lacks an internal reference for correctness because correctness lives outside the model by construction, no matter how much training it has received. Agnosticism toward organization-specific knowledge resists training because that knowledge is continuously renegotiated by the practice that holds it, and no stable corpus exists for a training run to absorb. Statelessness across sessions is architectural rather than transitional. The most sophisticated state-management mechanisms are powerless if the first two properties are not addressed first: without an external reference for correctness and without organization-specific knowledge, there is nothing of substance for memory to preserve. And storing “a state” is not the same as representing it correctly and completely. The knowledge must be captured in a structured, navigable form for the agent. Representation is what makes preserved knowledge usable. The conclusion is the same in all three cases: scaling produces a larger generative engine, not one that holds external reference, organization-specific knowledge, or project memory inside itself.
The three properties together produce a structural condition. An LLM is a generative engine with no internal reference for correctness, no capacity to evaluate its own output for adequacy, no persistent memory of the project, and no orientation toward the system’s future. The loop is what fills that condition from the outside. Without the loop, the model produces output that looks plausible and breaks on contact with reality. With the loop, the model has the external scaffolding it needs to be governed.
What goes inside the loop, and who owns it
The loop is not a monolithic entity. It carries four kinds of knowledge, each with its own source and channel of provision.
Methodological knowledge is the discipline of building correctly: specifications, architectural principles, testing frameworks, and process governance. It is fifty years of accumulated practice, codified as a discipline by Software Engineering.
Domain knowledge is what the system must mean and do in its world. It is held by the people who work in the domain: the clinicians, traders, operators, and analysts. It is partly tacit, often inconsistent, and irreducible to any corpus that lives outside the organization.
Design choices are the project-specific decisions made by the engineer as the system grows: architectural structures, component boundaries, trade-offs, and interfaces. Each choice becomes a constraint on what follows. None of them is elicited from anyone; they are produced and owned by the engineer.
Process choices concern how the work is organized for this project: the adopted lifecycle, the criteria for agent autonomy, and the review structure. They derive from methodological knowledge and must be instantiated for each context.
All four kinds of knowledge require reification. They must be materialized as persistent, accessible artifacts. Knowledge that exists only in someone’s head, or in a conversation that ended last week, does not exist for a stateless agent. Reification is not documentation overhead. It is the condition under which any knowledge becomes operational for the system at all.
Three actors structure the provision of all four. The domain expert is the source of semantic knowledge. The AI agent is the executor. The software engineer is the mediator and the custodian: the one who knows what to give the agent, how to structure it, and when to override its output.
Where the manifesto stops short
Nadella collapses these four kinds of knowledge into two (“human capital” and “token capital”) and the three actors into one (“humans”). The collapse obscures the real work.
You cannot compound what you cannot distinguish. If methodology, domain, design, and process are one undifferentiated blob, the loop has nothing to grip. The engineer’s role becomes indistinguishable from the domain expert’s; the domain expert becomes a generic stakeholder; and the agent is asked to do work that requires structured input nobody is responsible for producing. The result is what already shows up in many enterprise AI deployments: a generic agent dropped into a workflow with the assumption that judgment will somehow emerge from use.
The strategy and asset Nadella names are real, but the operational shape of that asset is not in the manifesto. The shape is the four levers, the three actors, and the reification that makes the loop into something an organization can own. Without that shape, the loop is a phrase a CIO (or CEO) repeats in steering meetings while nothing changes underneath.
Strategy and mechanism
Nadella has packaged the architecture that compensates for his company’s generic model into a statement. The statement does something the architecture itself cannot do: it sells. The statement invites a CIO to spend money on infrastructure that doesn’t yet have a shape. The conceptual architecture tells the CIO what shape the infrastructure must have, who must build it, and what knowledge must be reified for it to work. The first sells a vision. The second describes a job to be done.
Nadella’s manifesto is not empty. The loop is real, the warning about value concentration is real, and the principle that owning a learning loop is the new IP of the firm is correct. Correctness without structure is rhetoric. The asset Nadella names is the asset my paper describes. The difference is that the paper explains why building the asset is not optional and what it takes to do so.
This post was written with the assistance of Claude. The ideas, positions, and reasoning are mine.
© 2026 Alfonso Fuggetta & Sonia Montegiove. Salvo diversa indicazione, tutti i contenuti di questa pubblicazione sono protetti da copyright e rilasciati con licenza CC BY-NC-ND 4.0: https://creativecommons.org/licenses/by-nc-nd/4.0/deed.it
Very interesting! Congratulations