Every enterprise racing to adopt AI assumes the cost is the monthly bill for tokens. Microsoft CEO Satya Nadella says that's only half the price. In a blog post published this past weekend, Nadella argues that companies using proprietary AI models are quietly paying a second, hidden tax: their own institutional knowledge.
The warning lands at a moment when enterprise AI adoption is accelerating faster than most companies can think through the tradeoffs, and it's coming from one of the most powerful figures in the industry — a man whose own company has invested billions in both OpenAI and Anthropic.
The "Pay Twice" Problem
Nadella's argument centers on what he calls model "exhaust" — the prompts employees write, the tools AI agents use, and, most importantly, the corrections humans make when a model gets something wrong. Each of those corrections, he says, becomes distilled institutional know-how that the model provider can potentially learn from.
As Nadella put it in his post, buyers pay for AI capability once with money, then a second time by handing over the proprietary knowledge needed to make the model actually useful in their business. The better the desired output, the more sensitive data has to flow into the system to get there.
This isn't a purely theoretical concern. It echoes a broader unease that has been building among venture capitalists and enterprise leaders alike — including Palantir CEO Alex Karp — that the giant AI labs could eventually become competitors to the very companies feeding them data.
Turning the Distillation Argument Around
The most pointed part of Nadella's post is his call for reciprocity. AI labs freely scrape the open internet to train their frontier models, arguing fair use. Yet those same labs often restrict distillation — the practice of using a model's outputs to train a smaller, cheaper model that mimics its behavior.
Nadella calls this arrangement hypocritical. If model makers get to learn from the world's data without restriction, he argues, enterprises should be able to study the models they pay for in the same way. The tension is real: Anthropic itself accused Chinese open-source labs earlier this year of mass-querying Claude to train competing models, and pushed for tighter export controls in response. Readers who want the fuller economic picture should look at how DeepSeek's aggressive V4-Pro pricing has already reshaped enterprise AI economics, a trend directly connected to the distillation debate Nadella is now wading into.
Nadella's Proposed Fix
Unsurprisingly for the CEO of a major cloud provider, Nadella's solution steers enterprises toward infrastructure Microsoft is well-positioned to sell. He's urging companies to:
- Retain ownership of their prompts, feedback, and interaction data rather than ceding it to a model provider
- Build proprietary learning environments hosted on their own cloud infrastructure
- Adopt orchestration layers, or AI "gateways," that make it easy to switch between different model providers instead of locking into one
Nadella never says "open source" outright, but the subtext is hard to miss. Enterprises that run open models on their own infrastructure keep full control of both the model and the data that flows through it.
Enterprises Are Already Moving This Way
This shift isn't just a talking point in a CEO blog post — it's showing up in the numbers. Idit Levine, founder and CEO of Solo.io, told TechCrunch she's watching customers evaluate proprietary models, then ask a familiar question: can an open-source model running on-premises deliver 90% of the performance for a fraction of the cost, with full control over the deployment? Her company's technology powers the Linux Foundation's Agentgateway project, and counts T-Mobile, ADP, and SAP among its customers.
The traffic data backs this up. Open-source models accounted for 29% of all traffic routed through Vercel's AI gateway last month, and OpenRouter is reporting a similar surge. This mirrors the demand story playing out across the AI infrastructure stack more broadly — a dynamic also visible in TSMC's 36% quarterly revenue jump driven by AI chip demand, which shows just how much capital is chasing AI compute right now, proprietary or open.
What This Means for Enterprise AI Buyers
Nadella's post effectively hands enterprise decision-makers a checklist: audit what data your AI vendor contracts allow them to retain, evaluate whether an on-prem or open-source deployment fits your risk tolerance, and build in the flexibility to switch providers if terms change. As pricing models continue to shift — see how Anthropic has begun localizing Claude pricing for major markets like India — the commercial terms around data and distillation rights are likely to become just as competitive as pricing itself.
Whether or not Microsoft's motives are entirely altruistic, Nadella's core point is hard to dismiss: in an economy where corrections and feedback loops train the next generation of models, the enterprises generating that value have a legitimate claim to it. As he writes, "in consuming intelligence, you are creating intelligence." The question now is whether AI labs will agree to share it back.
