Satya Nadella has taken a public stance against OpenAI and Anthropic, the very companies in which Microsoft has invested billions. In a blog post, the CEO accuses AI labs of applying asymmetric rules: they freely train on public data invoking fair use, but contractually prohibit distillation of their own models (the practice of using a model’s outputs to train a smaller or cheaper one). Worse: they learn from customer interactions — prompts, corrections, feedback — accumulating proprietary knowledge from companies using their services.

Nadella calls it the «inverse information paradox»: you pay twice, with money for tokens and with the knowledge you reveal each time you use the model. «Every correction gets distilled into institutional know-how», he writes. This is exactly the kind of information a competitor could never buy, yet companies are handing it over.

Why it matters to you. If you use proprietary models for internal tasks, you’re implicitly allowing vendors to learn from your work. Nadella suggests three moves: maintain control of your own data (prompts, feedback, interactions), build proprietary learning environments on the cloud (surprise: Azure is an option), and adopt orchestration layers that let you switch providers without rewriting everything. The implied direction is clear: open-source models installed on-premise, where the learning loop stays within your walls. Idit Levine of Solo.io confirms the trend: «Enterprise customers increasingly ask whether they can run open-source models on-prem. It does 90% of what the big one does, costs less, and you control it».

In detail

The context: what came before

Distillation is an established ML practice: you use a large model’s outputs to train a smaller, faster, or cheaper one that mimics the first’s behavior on a specific task. This is how many competitive open-weight models emerge: they start from frontier model outputs. OpenAI and Anthropic explicitly prohibit this in their terms of service, and in February Anthropic publicly accused Chinese labs of sending millions of prompts to Claude to improve their own models, asking the US government for stricter export controls.

Until now, the labs’ narrative was simple: we train on public data for fair use, you cannot train on our outputs because they’re protected intellectual property. Nadella calls it hypocrisy: if fair use applies to scraping the web, it should also apply to learning from other models.

What Nadella says about customer data

The harshest point in his post concerns the «exhaust»: interaction data that companies generate using models. Prompts, corrections («the model got this wrong, the right answer is»), ratings, choices about which outputs to keep. Nadella argues this is valuable business knowledge, and current terms of service often allow vendors to use it to improve their own models, potentially building competing products from what they learned from your use cases.

The proposed solution is predictable from a cloud provider’s CEO: keep your data and learning loop on your own cloud, with tools that let you switch models without rewriting the integration. Nadella never says «open source», but the subtext is clear.

What’s happening in the field

Testimonies converge: many companies are already migrating to open-weight models on-premise. Idit Levine, CEO of Solo.io (the tech behind the Linux Foundation’s Agentgateway project), says enterprise customers — T-Mobile, ADP, SAP — after the pilot phase with proprietary models ask themselves: «Can I take an open-source model and run it on-prem? It’ll do 90% of what the big one does, cost far less, and I’ll control it».

Traffic data confirms it: Vercel reports that open models represented 29% of total traffic through their gateway last month. OpenRouter sees the same pattern. The push isn’t ideological: it’s cost control, predictable latency, and above all ownership of the learning loop.

Limits of Nadella’s position

Nadella has an obvious conflict of interest: Microsoft has invested billions in OpenAI, sells Azure as an AI platform, and now pushes companies to keep everything on the cloud (coincidentally, his). His criticism of the labs is genuine, but the proposed solution isn’t neutral: moving everything to Azure is still delegating part of your control to Microsoft.

Second: distillation isn’t a right symmetrical to training on public data. Public web data doesn’t have terms of service forbidding it (the debate is whether it should). A model’s outputs are generated by a system that cost billions and which the vendor sells as a paid service. Nadella’s analogy holds morally («if you can, why can’t I?»), but legally the difference exists.

Practical implications

If you’re using ChatGPT or Claude for internal work, read the terms about what the vendor can do with your interactions. OpenAI and Anthropic offer enterprise options where data isn’t used for training, but you must explicitly enable them and often pay more.

If your use case is repetitive and well-defined, evaluate open-weight models you can run yourself. Not for ideology, but because the learning loop stays in-house, costs are fixed rather than per-token, and you’re not implicitly training someone else’s model on your cases.

If you need the frontier (complex reasoning, novel tasks), proprietary models still lead. But use orchestration layers (gateways, routers) that let you switch providers without rewriting everything. The tools exist: Vercel AI SDK, OpenRouter, LangChain with provider abstractions. Don’t lock yourself to a single API.