All posts AI-Native Enterprise

Nadella's Reverse Information Paradox, read by a European CIO: where your company's know-how should live

The short answer

The reverse information paradox is Satya Nadella’s name for a hidden cost of buying AI: “You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.” He set it out in an essay dated 12 July 2026 on sn scratchpad, the blog linked from his X and LinkedIn profiles, as the mirror image of Kenneth Arrow’s information paradox: in Arrow’s market the seller risks giving knowledge away, in AI the buyer does.

The diagnosis is right. The remedy needs one correction. Nadella asks firms to build learning environments “within the tenant boundary”, and in cloud usage a tenant is an account on somebody’s cloud. Follow that literally and the evals, traces, adapted weights and memory that hold a company’s know-how accumulate inside one cloud provider instead of one model lab, so the dependency moves from the lab to the cloud. What holds is an owned, model-agnostic context engine, plus contracts that say what may leave and what must come back.

Key takeaways

  • The diagnosis is right. Every correction, eval and trace your staff produce teaches whoever holds it, “trace by trace, correction by correction, eval by eval”.
  • “We don’t train on your data” answers the wrong question. Microsoft, OpenAI, Anthropic and Google all promise it for paid use, with exceptions. Storage is the open question: OpenAI keeps hosted evals, vector stores and conversations “Until deleted”, and Microsoft keeps Responses API history, Assistants threads and stored completions in your Foundry resource until you delete them.
  • The remedy moves the dependency. A tenant stops the leak, not the captivity: inside one cloud tenant, someone else still owns the learning infrastructure.
  • Six artefacts make up the exhaust. They are expert corrections, eval sets, agent traces and memory, prompt libraries, retrieval indexes, and fine-tune data and adapted weights. Audit where each one sits.
  • The exit is written before signature. It includes a way out if the regulator moves. Since 2 August 2026 the Commission can ask for a model to be withdrawn and fine its provider up to 3% of global annual turnover or EUR 15 million, whichever is higher.

What is the reverse information paradox?

The reverse information paradox is the claim that a company buying AI gives away knowledge in order to use what it bought. In Nadella’s words: “In the AI age, the buyer risks giving away knowledge, just in order to use what they bought.” The better the model has to perform, the more context it needs, and over time “the information asymmetry becomes increasingly skewed”.

The leak is mostly not the data you upload. It is what Nadella calls “exhaust”: “the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong.” Take a generic example, not a client case: when a credit controller rewrites a dunning email because the customer is a strategic account with an open dispute, the edit carries a rule nobody ever wrote down.

For a company with EUR 250M to 2B in revenue, that unwritten knowledge is usually the moat. A product range can be copied. The way a company quotes, plans and handles exceptions cannot, until it has been typed, one correction at a time, into somebody else’s system. Own your intelligence covers who builds the alternative inside a company without an AI bench. This piece is about where the know-how ends up, and what to sign.

What does Satya Nadella’s reverse information paradox essay say?

The primary text is a short post on sn scratchpad, dated 12 July 2026. Much of the coverage, TNW’s included, works from the version Nadella posted on X, so the wording here comes from the blog page. It runs in four steps.

The diagnosis is the double payment, with the knowledge leaking “almost imperceptibly”. The principle follows: “In consuming intelligence, you are creating intelligence. And what you create should belong to you.”

The critique is aimed at the model providers. They train on public data, and he finds it ironic that “the status quo is to then turn around and impose restrictive terms on distillation, and to reserve the right to learn from customer usage and interaction data.” Then comes the sentence the rest of this piece turns on: “If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself.”

The remedy is a trust boundary “where an organization’s data, traces, evals, adapted weights, and memory accumulate and improve together”, across which “nothing crosses, not even the intelligence exhaust, without consent.” He predicts that enterprises “will demand the rights to use model outputs to fine tune and/or train their own models”, then sets out five pillars: Control, Capability, Choice, Cost and Compound. The essay names no cloud provider: neither Microsoft nor Azure appears on the page.

How is it different from Kenneth Arrow’s information paradox?

Kenneth Arrow described the original in Economic Welfare and the Allocation of Resources for Invention, an NBER chapter published in January 1962. Nadella quotes him: “Its value for the purchaser is not known until he has the information, but then he has in effect acquired it without cost.”

Arrow’s information paradox (1962)Nadella’s reverse information paradox (2026)
Who risks giving knowledge awayThe seller, to prove what it is sellingThe buyer, to make what it bought useful
What leaksThe idea itselfThe exhaust of use: corrections, evals, traces, prompts
How it leaksIn one disclosureGradually, with every task
The partial fixPatents: disclose without giving awayNot settled. Nadella proposes a trust boundary

Patents solved part of Arrow’s problem, and Nadella writes that “The Reverse Information Paradox needs its own equivalent.” There is no patent office for a correction. For a buyer the equivalent has two parts that only work together: a contract that says what the provider may do with what you reveal, and an architecture that decides where the artefacts accumulate. A contract alone gives you the right to an export from a system you cannot run. An architecture alone leaves feedback and telemetry outside the boundary.

Does your AI vendor train on your enterprise data, and is that the right question?

For paid enterprise use, Microsoft, OpenAI, Anthropic and Google say no by default, each with an exception worth reading. The better question is where the exhaust is stored and whether it leaves with you, which none of these promises addresses. Read AWS Bedrock, or any other platform you run, the same way.

ProviderThe default promiseThe exception to readWhere the exhaust still sits
Microsoft (Foundry; Microsoft 365 Copilot)On Foundry models sold by Azure, prompts, completions, embeddings and training data “are NOT used to train any generative AI foundation models without your permission or instruction” (Microsoft)For Copilot: “We may use customer feedback, which is optional, to improve Microsoft Copilot”, though not to train the foundation LLMs (Microsoft)With the Responses API, Assistants threads or stored completions, “the service creates a data store to persist message history and other content”
OpenAI API“As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models” (OpenAI)Unless you opt in to share dataEvals, vector stores, fine-tuning jobs and conversations are kept “Until deleted”
Anthropic, commercial terms“Anthropic may not train models on Customer Content from Services.” (Anthropic)A thumbs up or down report lets Anthropic use the related chats to train its models (Anthropic)The whole related conversation, stored for up to 5 years and de-linked from your user and customer IDs
Google (Agent Platform; Gemini API and AI Studio)On Gemini Enterprise Agent Platform, Google won’t “use your data to train or fine-tune any AI/ML models without your prior permission or instruction” (Google Cloud). In the EEA, Switzerland and the UK, the Gemini API’s paid-service data terms apply even to free useWith Grounding with Google Search, Google “will store prompts, contextual information that you may provide, and output for thirty (30) days”, and the stored information can be used for debugging and testing (Google)Prompts and responses are logged “for a limited period of time” for abuse detection and “may be stored transiently or cached in any country in which Google or its agents maintain facilities”; on Agent Platform, zero data retention needs specific actions by the customer

Training is the promise every provider can afford to make. Storage is where the exhaust lives, on the provider’s retention clock and in its schema. “Until deleted” is a retention policy, not an ownership model. And the feedback button is where staff hand over the most valuable exhaust, a correction, in one click. Microsoft’s exit, layer by layer, is in our Microsoft Frontier Company analysis.

What is intelligence exhaust? A six-point audit of where your company’s know-how sits

Intelligence exhaust is the by-product of using AI that teaches whoever holds it. Nadella names prompts, tool use and corrections, and his trust boundary adds traces, evals, adapted weights and memory. Inside a mid-cap that comes down to six artefacts. Run the audit on one workflow first, and answer every column about the system as it runs today.

ArtefactWhat it isWhere it usually sits todayWho holds it, on whose termsWhat to checkPasses when
Expert correctionsThe edits, rejections and overrides an expert makes when the model is wrongFeedback buttons in vendor apps, edited drafts in email, comments in a ticketThe vendor behind the feedback button, on feedback terms that can differ from its training termsWhether feedback buttons are on for staff. Copilot feedback is managed through admin controls; Claude Team and Enterprise owners can turn the button off under Rate chatsEvery correction lands in a store you control with the before, the after and who made it
Eval setsThe cases, expected outputs and graders that define “good” for a taskA vendor eval console, or a notebook on one engineer’s laptopThe provider, on its retention clock (OpenAI keeps evals “Until deleted”)Whether cases, graders and run history export and run against another modelThe suite lives in your repository and runs against at least two providers
Agent traces and memoryThe step-by-step record of what an agent read, called, decided and remembered, and who approved itThe agent platform’s tracing product, conversation stores kept by the API, managed memory and thread storesThe agent platform, on its retention settings and in its schemaRetention, export format and schema. Decision traces at system handoffs are covered in nobody owns the handoffsTraces and memory land in your store in an open format with a schema, and the vendor copy follows a retention you set
Prompt librariesThe system prompts and templates that encode how your company does the taskHosted prompt objects in a vendor console, or pasted into a low-code toolThe vendor console, until the vendor withdraws the featureWhether every production prompt is in version controlPrompts are versioned in git and tested against the eval set on every change
Retrieval indexesThe chunked, embedded copy of your documents that agents searchA managed vector store or search serviceThe search service; the vectors only answer queries embedded with the same modelWhether you hold the source chunks, the chunking settings and the embedding model IDYou can rebuild the index on another engine from chunks you hold
Fine-tune data and adapted weightsInput and output pairs used to tune or distil a model, and the tuned weights they produceUploaded files, fine-tuning jobs and tuned deployments on the provider’s platformThe provider, on terms that decide what you may train with outputsWhether you hold the dataset, whether the contract lets you train your own models on outputs, and whether the tuned weights can be downloaded or only called on the provider’s platformThe dataset is versioned in your store, the contract grants output rights for internal tuning, and you know where the weights can run

Two rows need more than a table cell. Hosted prompt libraries can be withdrawn: OpenAI says it “is deprecating reusable prompt objects in the API”, has set v1/prompts to shut down on November 30, 2026, and now tells developers to “Treat prompts as application code” (OpenAI). Prompts kept in that console need a migration; prompts kept in git do not. And an index does not travel alone, because it is tied to its embedding model: “To ensure accurate results, use the same embedding model for indexing and querying” (Microsoft). A vector export without the source chunks and the model ID is a file of numbers no other system can query.

Does Nadella’s remedy fix it? The five Cs, read by a European CIO

Partly. Control, Choice, Cost and Compound are written to hold on any architecture. Capability is the one pillar that names a place, and once the learning environment sits inside one cloud tenant, the other four tend to follow it there, as the table shows. “Build your own proprietary learning environments within the tenant boundary to train or tune models” draws the boundary around the model provider and says nothing about the cloud the tenant runs on. Built entirely in one hyperscaler’s managed services, the contents of his trust boundary accumulate there.

The cloud holds your traces in its formats and at its prices, and the cost of moving them rises with every turn of the loop. A tenant stops the leak. It does not end the captivity, and by his own logic value still converges toward “the owners of the learning infrastructure”.

The essay names no cloud, but its author runs one. TNW noted the next day that “Microsoft, naturally, sells products that do each of these things”, and still concluded that the core point holds. We agree with both halves, and apply his own Choice test one layer further down.

He asks: “If any one model you are using is taken away, do you still have the ability to operate and optimize for your evals using other models? Does your company ‘veteran’ capability remain with you even if a given ‘generalist’ model is taken away?” Ask it twice more. If the cloud tenant were taken away, could you still optimise for your evals? If your delivery partner left on Friday, could you on Monday? The second question applies to us too.

PillarHis askWhere it lands if built inside one cloud tenantWhere it lands in an owned context engineTest question
Control“Create your private evals” and “retain ownership of your organization’s memory, traces, feedbacks, decisions”The cloud’s evaluation console, tracing product and managed memory store. Yours on paper, readable through its APIsEval suites in your repository; traces, memory and decisions in a store in your accounts, with a published schemaCan we run the full eval suite tonight without logging into a vendor console?
Capability“learning environments within the tenant boundary to train or tune models”The cloud’s managed training service. Whether adapted weights can leave depends on the model’s licence and the serviceTraining data and evals kept as your own artefacts; tuning runs where it is cheapest, on open-weight models when the weights must stay yoursCould we reproduce the tuned model elsewhere from data we hold?
Choice“the orchestration layer is decoupled from any single model”Model choice is open across the catalogue, but orchestration runs on the cloud’s agent serviceOrchestration in code you hold, with model calls behind one interfaceSame test, applied to the tenant and the partner
Cost“bring together context, models, and tasks in the most efficient and cost-effective way”Optimised within one price listEach task routed to the cheapest model that passes its eval, across providersDo we know the cost per task per model, and can we act on it without a migration?
Compound“your own continuous learning loop (i.e. hill climbing machine)”The loop compounds in one cloud’s formats, so the switching cost rises with every turnThe loop compounds in assets you hold, so the gain survives a change of model, cloud or partnerIs the asset that grows every month something we could hand to a different provider?

The fourth column is what we mean by a context engine: the owned, model-agnostic layer that holds a company’s reconciled context and its learning record, from eval suites and trace store to correction log, prompt repository and chunk store. It can run on the cloud you already use. Owning it means the artefacts sit in formats and accounts you control, so the cloud stays a supplier you can replace. Putting the tenant in an EU region settles where the data sits. Who controls it, and whether you can leave, are separate questions, and the Frontier analysis linked above covers residency versus jurisdiction. The same reasoning applies to each of the five layers that accumulate in a serious engagement.

Who owns the data your company creates using AI?

Nadella’s answer is that it should be yours: “retain ownership of your organization’s memory, traces, feedbacks, decisions, and institutional context, and ability to use outputs of models from your own tasks and queries.” The contract usually says less, layer by layer, and most contracts only cover outputs. Anthropic’s commercial terms, for example, assign the customer “its right, title and interest (if any) in and to Outputs” and forbid using the services “to train competing AI models” without Anthropic’s approval (Anthropic). Google’s Gemini API terms say “You may not use the Services to develop models that compete with the Services” (Google). These are the distillation terms Nadella finds ironic.

For a mid-cap tuning a small model on its own quoting workflow, the risk is not competing with a frontier lab. It is that “competing” is undefined and the provider interprets it. Write the definition in. Ownership of outputs is also not ownership of exhaust. Where a correction is named at all, it is usually as feedback the provider may use: Anthropic’s terms say it “may use that feedback at its own risk and without obligation to Customer” (Anthropic). The eval case, the trace and the index around each output are rarely named at all. The EU Data Act’s switching rights, and how far they reach into rule files and memory, are covered in the Microsoft Frontier piece. The records regulated workflows must keep are in audit trails for AI agents.

How do you avoid AI vendor lock-in? An exit-clause checklist

AI vendor lock-in rarely sits in the model, which is the easiest layer to switch. It sits in the exhaust that grew up inside one provider’s formats. You avoid it with the architecture in the audit above and with contract wording. The clauses below are drafting for counsel to adapt to the governing law, and the values in brackets are there to negotiate.

A mid-cap on standard online terms is unlikely to get most of these from a hyperscaler or a lab. Against a cloud provider, part of the ground is already statutory: under the EU Data Act, platform and software providers must “export data in a commonly used and machine-readable format”, and switching charges, including charges for data egress, end from 12 January 2027. Whether evals, traces and memory count as exportable data, rather than the provider’s protected material, is the question to put to counsel. Where you can negotiate, in an enterprise agreement amendment, through a reseller, with a smaller AI vendor or with your delivery partner, use the wording. Where you cannot, read the standard terms against each clause and close the gap in the architecture, by keeping the artefact in your own store.

  1. No learning from use. “Except as strictly necessary to provide, secure and support the Services to the Customer, including abuse monitoring disclosed in the Provider’s published terms, the Provider shall not use Customer Content, feedback, ratings, telemetry or usage data, including under any service-improvement or product-analytics programme, to train, fine-tune, evaluate or improve any model or service other than the Customer’s own, without the Customer’s prior written consent for a named purpose.”
  2. Output rights. “The Customer may use Outputs to fine-tune, train, evaluate or distil models for its own internal use. A model used only within the Customer’s group is not a competing product or service.”
  3. Export of the learning record. “On request and on termination, the Provider shall export all eval cases, graders and results, agent traces, conversation and memory stores, prompts and configuration, in [JSONL or Parquet] with a documented schema, within [30] days, at no additional charge.”
  4. Index export. “Any export of a retrieval index shall include the source chunks, chunk metadata and chunking parameters, and the identifier and version of the embedding model used to build it.”
  5. Deletion certificate. “Within [60] days of termination, the Provider shall delete Customer Content and all artefacts derived from it, including backups and copies held by subprocessors, and certify the deletion in writing, system by system, or, for backups, certify that they are isolated from use and expire within [90] days.”
  6. Termination assistance. “For [six] months after notice of termination, the Provider shall continue the Services on existing terms and assist the transfer to the Customer or a successor provider.”
  7. Notice before terms change. “The Provider shall give [90] days’ written notice of any change to its data-use, retention or training terms, and the Customer may terminate without penalty before the change takes effect.” For comparison, Anthropic’s commercial terms take effect 30 days after an update is posted, and immediately where the change responds to law or regulation.
  8. Model deprecation and substitution. “The Provider shall give at least [six] months’ notice before retiring, replacing or automatically upgrading any model version used in production, and the Customer may substitute another provider’s model without penalty or loss of committed spend.” For publicly released models, Anthropic commits to “at least 60 days’ notice before model retirement” (Anthropic). In Microsoft Foundry, generally available models from Anthropic, DeepSeek, Fireworks and Mistral AI “follow a 12-month lifecycle instead of the standard 18-month lifecycle”. Separately, for Global Standard, Data Zone Standard and Standard deployments, Microsoft upgrades a retiring model version automatically (Microsoft). A new model version can move every number in your eval suite.
  9. Regulatory-event exit. “The Provider shall inform the Customer, as far as the law permits, of any request for information under Article 91 of the AI Act concerning a model used under this Agreement. If the European Commission requests or decides on measures under Article 93 concerning such a model, the Customer may switch models or terminate without penalty.”

The last clause separates the two on purpose: a request earns notice, a measure earns an exit. The power is no longer theoretical. The Commission’s enforcement powers over providers of general-purpose AI models have applied since 2 August 2026, “including with fines” (European Commission). Article 91 lets it request documentation and information from a provider, Article 93 lets it ask a provider to “restrict the making available on the market, withdraw or recall the model”, and Article 101 sets fines of up to 3% of global annual turnover or EUR 15 million, whichever is higher. Providers of models placed on the market before 2 August 2025 have until 2 August 2027 to comply. On 29 August 2026 the Commission, through Executive Vice President Henna Virkkunen, confirmed that the AI Office had sent formal information requests to more than 30 companies that build general-purpose AI models, and it has not confirmed the recipients by name (EU Perspectives). A request is not a finding. The buyer’s point is narrower: a model your workflow depends on can be required to change, or be withdrawn, under threat of fines, on a timetable neither you nor your vendor sets, so the learning loop should not depend on one provider.

Put the same clauses in your delivery partner’s contract, ours included. A partner that builds the learning record in its own platform recreates the paradox one company further away.

When is it fine to let the vendor learn from you?

When the knowledge is common. Rewriting a public job advert, translating a standard safety data sheet and code completion on public frameworks teach a vendor how everyone works, which is what you want the product to know. Paying twice costs little when the second payment is knowledge every competitor already has, and auditing those workflows is overhead without a return. The task can be common when the content is not, so the feedback-button setting from the audit still applies.

Following the remedy inside a Microsoft tenant is defensible for a mid-cap already on Microsoft 365, Entra and Purview, even for a workflow that matters, when getting a first workflow into production quickly matters more than portability. Take that trade knowingly: run the audit, sign what you can, and know which layers leave as files. The per-layer answer is in the Frontier analysis linked above.

Nucleo is not the right partner for a company that wants one cloud’s products end to end and has no workflow it considers proprietary. Running the audit is cheap. Keeping six stores current is not, and own your intelligence sets out who carries that cost. If the audit finds proprietary exhaust in one workflow, own that one and rent the rest.

FAQ

What is the reverse information paradox? Satya Nadella’s name for a hidden cost of buying AI: you pay for intelligence twice, once with money and again with the proprietary knowledge you reveal to make it useful. He set it out in an essay dated 12 July 2026 as the reverse of Kenneth Arrow’s 1962 information paradox, in which the seller, not the buyer, risks giving knowledge away.

Where can I read Nadella’s reverse information paradox essay? On sn scratchpad, the blog linked from Satya Nadella’s X and LinkedIn profiles, at snscratchpad.com/posts/reverse-information-paradox, dated 12 July 2026. Much of the news coverage, TNW’s included, works from the version he posted on X. The essay names no cloud provider: neither Microsoft nor Azure appears in it.

What are Nadella’s five Cs? Nadella lists five pillars, all starting with C: Control (private evals, and ownership of memory, traces, feedback and decisions), Capability (learning environments within the tenant boundary to train or tune models), Choice (orchestration decoupled from any single model), Cost (combining context, models and tasks efficiently) and Compound (the four together as a continuous learning loop, which he calls a hill climbing machine).

What is intelligence exhaust? The by-product of using AI that teaches whoever holds it. Nadella names the prompts people write, the tools agents use and especially the corrections people make when the model is wrong. Inside a company it comes down to six artefacts: expert corrections, eval sets, agent traces and memory, prompt libraries, retrieval indexes, and fine-tune data and adapted weights.

Does your AI vendor train on your enterprise data? For paid enterprise use, Microsoft, OpenAI, Anthropic and Google say no by default, with exceptions such as feedback buttons, opt-in sharing, search grounding and, outside Europe, Google’s unpaid tiers. The better question is where your evals, traces and indexes are stored and whether they leave with you: OpenAI, for example, keeps evals, vector stores and fine-tuning jobs until you delete them.

Does a private cloud tenant solve the reverse information paradox? Partly. A tenant stops the model provider learning from you, which fixes the leak. Built entirely inside one cloud’s managed services, the evals, traces, adapted weights and memory then accumulate in that cloud’s formats and at its prices, so the dependency moves from the lab to the cloud. Apply Nadella’s own Choice test to the tenant: if it were taken away, could you still optimise for your evals?

How do you avoid AI vendor lock-in? Keep the learning record in formats and accounts you control, and write the exit into the contract: no learning from use, output rights for tuning, export of evals, traces and memory with a schema, index export with the embedding model ID, certified deletion, termination assistance, notice before terms change, model substitution, notice of EU AI Office information requests, and an exit if the Commission moves to measures against the model.

Who owns the data a company creates using AI? Nadella argues it should belong to the company that created it. In practice it is whoever the contract says, layer by layer. Commercial terms usually give the customer its outputs but restrict using them to build competing models. Where corrections are named, it is usually as feedback the provider may use, and the evals, traces and indexes produced around those outputs are rarely named at all. Name those artefacts in the contract and define what counts as a competing model.



Nucleo is an AI-native enterprise transformation company for European mid-caps. Our forward-deployed engineers build the context engine inside your stack, model-agnostic and governed to EU rules, so the evals, traces and corrections your people produce stay in formats and accounts you control, whichever model or cloud runs on top. See how an engagement works, or write to us.

Sources: Nadella, The Reverse Information Paradox; Arrow (NBER, 1962); Microsoft Foundry data privacy; Microsoft Copilot privacy; OpenAI, your data; OpenAI, prompting; Anthropic commercial terms; Anthropic, model training; Anthropic, model deprecations; Google Cloud, zero data retention; Gemini API terms; Azure AI Search embeddings; Foundry model retirements; Commission, GPAI providers; AI Act Article 91, Article 93 and Article 101; EU Perspectives; Commission, Data Act explained; TNW.

The future belongs to those who see it before it is obvious, and build it while everyone else is still out there scouting for the best AI tools.