Satya Nadella now makes the case for control in enterprise AI. His five principles are useful for a mid-sized firm, but capability and compound can, under provider-specific learning infrastructure, work against choice.
First briefing · Operating Intelligence · Annotated reading of Satya Nadella, “The Reverse Information Paradox,” post on X, 12 July 2026.
On 12 July 2026, Microsoft’s chief executive argued that enterprises need a hard boundary around the institutional knowledge that makes purchased intelligence useful.
Microsoft is an interested enterprise-AI vendor.
Kenneth Arrow described a paradox in the market for information: its value to a buyer is unknown until the buyer has it, at which point the buyer has in effect acquired it without cost. In Nadella’s reading of Arrow, the seller risks giving the thing away in the act of selling it. Arrow’s formulation appears in his 1962 NBER chapter, Economic Welfare and the Allocation of Resources for Invention, p. 615.
Nadella’s claim is that AI can invert this. Now the buyer may take the risk. “You essentially pay for intelligence twice,” he writes, “once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.”
Nadella’s argument is about who controls that knowledge and on what terms. Microsoft states that Foundry models sold by Azure do not use customer prompts, completions, embeddings or training data to train foundation models without permission or instruction. Some stateful features store content, abuse monitoring can expose prompts and completions to authorised reviewers when they have been flagged or form part of a potentially abusive pattern of use, and preview practices may differ.
Prompts, tool calls, traces, evaluations and corrections can encode hard-to-replace institutional judgement. The operating risk is both inappropriate exposure where controls are weak and the firm’s failure to retain its own learning in reusable form.
Nadella proposes five: control of your own evals, memory, traces and decisions; capability to train or tune “within the tenant boundary”, the customer’s own space in Microsoft’s cloud; choice, meaning an orchestration layer decoupled from any single model; cost, which follows from choice; and compound, the continuous learning loop that makes the other four pay off.
For a firm of 50 to 500 people, control is the one to act on first. Evals—the tests that define what “good” looks like for quoting, inspection reports or service triage—are among the more portable assets when the firm owns them in documented, exportable formats. A firm with representative evals can compare models. A firm without them cannot tell whether a change improved the workflow.
Capability, training or tuning inside the tenant boundary, requires data, evaluation, engineering and contractual capacity. Microsoft announced Frontier Tuning on 2 June 2026 as a private preview delivered through Forward Deployed Engineers. Microsoft says it applies reinforcement learning inside a customer compliance boundary and plans later availability in Copilot Studio and Foundry. The announcement names customer engagements and reports a Microsoft-internal HR task-completion increase from 13% to 87%, but provides neither the evaluation method nor independent validation. It does not establish general availability, export rights or portability. Microsoft’s product page for Frontier Tuning says “You control the model. No vendor lock-in.” That is a vendor’s claim about a private preview; a firm that relies on it should have it written into the contract as an export and replay term.
Cost can be cut by routing routine, high-volume work to cheaper models, once quality, latency, volume, integration and monitoring costs have been measured.
Choice is stated as a test: if any one model you use were taken away, could you still operate and still optimise against your evals using another? It is in tension with compound, the learning loop Nadella places inside a trust boundary where “an organization’s data, traces, evals, adapted weights, and memory accumulate and improve together”.
Choice and compound conflict when accumulated state or the learning environment is provider-specific. Evals and traces are portable only when the firm controls their formats and export. Adapted weights are portable only when access, licensing, architecture and runtime permit. A learning loop may therefore improve performance while increasing switching costs.
Nominal portability is the right or the technical means to take assets elsewhere; economic portability is moving them at a cost the firm can bear. Interface access or downloadable weights do not remove runtime, hardware-fit, integration and migration costs.
A firm should build control and choice before it commits to capability and compound.
Microsoft sells products—including Copilot Studio and Foundry, with Frontier Tuning in private preview—that map closely to several remedies.
Nadella’s model test applies to Microsoft as well: if this vendor were taken away, what would we keep?
Nadella’s 23 July post describes Microsoft’s practice of ensuring that “the harness, memory, context, skills are externalized outside of the model” and says Microsoft is making “all this available as part of Foundry and our toolchain”. A Microsoft Foundry blog post says that “the model in the middle is a part you can swap” and that the loop around it is “the asset that compounds the longer it runs”; it also says Foundry lets a firm build that loop “in an open, interoperable, and modular way, so you can swap any piece”. State kept outside the model but inside Foundry passes Nadella’s model test and can still fail the vendor test.
For a Mittelstand firm, the answer should include its evals, documented workflows, reusable corrections and an evidence-based migration estimate.
Ask these questions in the next vendor and architecture review:
Write evals before writing large cheques. Then test the exit path.
The workflow must capture corrections in governed, reusable and exportable records; otherwise the firm may fail to compound its own learning even where the provider does not train on its data.
The reverse information paradox adds a second test alongside efficiency: does the firm retain the learning needed to improve the workflow and to change suppliers on acceptable terms?
The tension between choice and compound rests on learning assets that cannot move at acceptable cost. Two observations together would remove it for a given product: a generally available contract that lets the customer export its adapted weights, learning environments and traces and run them elsewhere, and a named customer that has made such a move at a documented cost. The first alone shows nominal portability; the second shows what the economic kind costs. As of 24 September 2026, Microsoft’s published Azure terms do not name Frontier Tuning, announced on 2 June 2026 as a private preview; on the published record the first condition is not met.
Sources and verification
Attributed claims are labelled as such. Nadella and Microsoft are interested parties; the diagnosis is treated as a procurement hypothesis. Product-specific retention, exportability and executed contract terms remain unresolved.