CEO Satya Nadella just went against all norms as he publicly admitted that it costs more to use artificial intelligence than what it says in the bill. In a lengthy blog on X published on July 12, the chairman and CEO of Microsoft said that all enterprises utilizing AI effectively are paying for it not once but twice. Once through cash, and once through something he deems far more valuable: the company’s very own expertise.
Also read: Meta keeps launching AI features that assume consent: Why it keeps backfiring
According to him, to obtain meaningful responses from an AI model, one needs to supply the system with the context of the business, its processes, errors, and corrections. This is called “intelligence exhaust,” according to Nadella, referring to the byproducts of prompts, corrections, and evaluations that workers make when employing an AI system. He says that such intelligence exhaust is not garbage information. It’s actually institutional knowledge that cannot be easily purchased from a competitor and is leaking out gradually with each correction in the model.
To drive home the point, he turned to economics. Economics Nobel laureate Kenneth Arrow once talked about an “Information Paradox”: it is impossible for the seller of information to establish its value without disclosing it, while once disclosed, there is no incentive for the buyer to pay. But Nadella says the reverse is true for AI; it is actually the buyer, the enterprise, who ends up disclosing its knowledge. The more efficient you need your model to be, the more institutional knowledge you have to put in.
Also read: Can AI make you a better cricketer? Str8bat co-founder Gagan Daga thinks so
It is disturbing in how it is lopsided. While the AI firm gains knowledge from all the enterprises at once and builds on that knowledge for its customer base, each single enterprise gains only minimal knowledge for itself while losing control of what it has contributed to the machine-learning process.
So what is his solution? He has come up with a solution that he calls the “five Cs.” The first one is Control, which suggests that organizations should have control of their evaluations and institutional knowledge rather than leave this in the hands of a third party vendor infrastructure. The second is Capability; which suggests the need to create private, tenant-aware environments in which models are trained using actual organizational data, keeping them from leaking out. Third is Choice, which entails not getting locked down into a particular model or vendor. Fourth is Cost, which involves making use of different models and workflows in an efficient way rather than just choosing the most expensive option for everything.
It’s worth noting the timing. Nadella has spent much of 2026 talking publicly about AI’s economics turning uncomfortable, from telling Microsoft staff to stop defaulting to the most expensive models, to warning in an earlier essay that AI could hollow out entire industries the way outsourcing once did. This latest post arrived the same week Microsoft was hit with a shareholder lawsuit over its AI spending disclosures. Whether the Reverse Information Paradox is a genuine strategic warning or a bit of framing to explain why enterprises should build more on Microsoft’s own tools remains an open question. Either way, it’s a rare admission from a hyperscaler CEO that AI’s price tag doesn’t end at the subscription fee.
Also read: Robot surpasses human? Neo’s robotic hand dexterity feels scary good right now