Nadella — The Reverse Information Paradox

Verdict: READ (~10 min). Founder-shared 2026-07-13. 7.7M impressions / 18k bookmarks in ~20h — this framing is entering the enterprise-AI vocabulary NOW.

The argument

Arrow's Information Paradox (1962): the seller of information risks giving it away in order to sell it. Patents were the fix — disclose without surrendering.
The reverse (Nadella's coinage): in AI, the buyer gives away knowledge just to use what they bought. You pay for intelligence twice — money, plus the proprietary context you must reveal to make the model useful. The better you want it to perform, the more you feed it.
Models learn from "exhaust" — prompts, agent tool use, and especially corrections. "Every correction is distilled into institutional know-how… the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval."
Named irony: providers claim fair use to train on public data, then impose restrictive distillation terms and reserve the right to learn from customer usage. "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure."
Karp quote (Palantir): technical customers "want to know they own the means of production."

His prescription — the "trust boundary" + five C's

Control — create your private evals ("evals define what 'good' looks like inside the organization"); own your memory, traces, feedback, decisions; retain rights to use model outputs on your own tasks.
Capability — proprietary learning environments inside the tenant boundary; models learn against real workflows without exposing company knowledge.
Choice — orchestration layer decoupled from any single model. "If any one model you are using is taken away, do you still have the ability to operate?"
Cost — decoupled orchestration = route context/models/tasks efficiently.
Compound — the four together = a continuous learning loop ("hill climbing machine").

Why this matters to RDCO/phData (the real reason to read)

This is the brigade house described from the demand side, by the largest software company's CEO:

Nadella	House
Private evals define "good"	Eval-proven skills; "evals prove it works at all, MISE proves it works HERE" (founder's two-halves-of-trust moat, 7/9)
Own your traces/memory/decisions	Cellar + tickets-as-build-records + close-out signatures
In-tenant learning environments	House runs in YOUR env; RDCO-owned private center; client cellars on client infra
Orchestration decoupled from any model	Walk port + station roster; model-agnostic adapters; the founder lived the "model taken away" test when Fable was pulled in June
Compound learning loop	fill → exemplar → rubric eval → tasting → freshness-watch

Sales-narrative ammo: CAF/DIE's pitch ("governed knowledge graph at the center, client owns the learning") now has a named market concept behind it, with 7M impressions of demand-side validation. Usable in Kwik Trip discovery framing this week: your corrections and evals are the asset — who owns them in your current stack?

Bias flag

Not neutral analysis — this is also Microsoft positioning. "Build in your tenant, decouple from any one model" is an Azure/Foundry wedge against OpenAI/Anthropic API terms (and quietly against his own OpenAI partner). The diagnosis is real; the prescription conveniently lands on his cloud. Cite the concept, not the vendor conclusion.

Sources cited in the article

Arrow 1962 NBER chapter (nber.org/system/files/chapters/c2144/c2144.pdf)
Palantir/Karp post · Nadella's earlier post

[[2026-07-09-anthropic-plugin-ecosystem-vs-rdco-brigade-plugins]] — the moat framing this validates
01-projects/phdata/fabric-spec/FABRIC-SPEC-v0.md — the Fabric = the trust boundary, specified
plugins/ab-skill-factory/AGENT-BRIGADE-STANDARD.md (brigade-house) — the machinery