You already paid for the AI model, but there’s a second price most teams don’t notice until later: the proprietary knowledge you hand over just to make the model useful. That’s not a Leaseweb argument, but the argument Microsoft’s own CEO made in public, in July 2026.
Microsoft’s own CEO argued in July 2026 that enterprises pay for AI twice: once in cost and once in the proprietary knowledge they hand over to make the model useful. His own fix still runs on cloud infrastructure. That’s the part worth pushing on.
This describes almost exactly the tradeoff we spend most of our conversations with infrastructure teams working through. The diagnosis is right. Where the conversation usually stops one step too early is naming the problem, when it’s just as important to ask who’s positioned to fix it.
The paradox, explained by the person who named it
On July 12, 2026, Satya Nadella published a long-form essay on X entitled “The Reverse Information Paradox,” which drew more than 10 million views within two days [1]. His argument inverts economist Kenneth Arrow’s 1962 Information Paradox, which described a seller’s dilemma: you can’t prove what you know is valuable without revealing it, and once revealed, the buyer has it for free [2]. Nadella’s point is that AI flips this onto the buyer. To get real value from a model, you have to feed it your prompts, your corrections, your institutional know-how, the accumulated expertise that makes your business run differently from your competitors’. The better you want the model to perform, the more of that you have to hand over [1].
Even the person naming the problem has skin in the game
Multiple outlets covering the essay pointed out something worth considering: Nadella runs the company that built much of the infrastructure creating this exact risk, and his own prescribed fix (control your data, build private learning environments, avoid single-model dependency) still assumes you’re running that fix on cloud infrastructure [3]. As one outlet put it, enterprises can swap the model, but the essay doesn’t ask them to swap the cloud underneath it [3].
That’s not a reason to dismiss the argument. The mechanism Nadella describes is real, and worth taking seriously regardless of who said it first. The more useful question isn’t whether the diagnosis is correct (it is), but whether the fix should come from inside the same business model that created the exposure in the first place.
What leaks, in practice
Strip away the framing and the substance is concrete: every prompt reveals what your team is working on. Every correction teaches the model how your industry operates, knowledge a competitor could never simply buy. Every tool call and agent trace is a small transfer of institutional expertise to whichever company owns the model you’re using [1].
Separately, at a policy level, the EU’s proposed Cloud and AI Development Act, still awaiting approval from Parliament and the Council, points in a similar direction: reducing dependency on a small number of non-EU providers and pushing publicly funded software toward open, interoperable standards [4]. Different motivation, same underlying instinct: don’t let one vendor own every layer of your stack.
The fix Nadella describes, and who should build it
Nadella’s own prescription is a hard trust boundary: keep your institutional knowledge inside infrastructure you control, don’t let learning compound silently inside a vendor’s training pipeline [1]. That’s the right instinct. The open question is whose infrastructure that boundary should sit on.
A trust boundary built entirely inside the same company selling you the model isn’t really a boundary, it’s a room inside their house. A trust boundary built on dedicated, private infrastructure, one you control independently of any single AI vendor, is a room inside your own. That distinction matters more than the framework itself: a trust boundary is only as real as the independence of whoever operates it.
What this looks like in practice
Running fine-tuning and inference on public multi-tenant environments means paying twice: once in variable, consumption-based cloud costs, and again in the proprietary operational knowledge that flows to the model owner. Running the same workloads on dedicated infrastructure keeps that operational knowledge inside your own environment instead, and turns AI spend into a cost you can plan around rather than one dictated by usage.
The economics back this up, and they’re sharper than “predictable” implies. Independent research from Uptime Institute puts the breakeven point at 33% utilization: past that threshold, dedicated GPU infrastructure costs less per unit than the equivalent cloud instances, and at 40% utilization it runs about 25% cheaper [5]. Below that line, the math flips hard: at 8% utilization, the same dedicated cluster can cost roughly four times what the equivalent cloud capacity would have cost [5]. Sweating a dedicated cluster isn’t optional if you want the economics to work in your favor, it’s the entire mechanism. The honest caveat: dedicated infrastructure only wins if you actually use it, which is a workload-planning question worth answering before you commit, not after.
Leaseweb has been building infrastructure outside any single vendor’s ecosystem since 1997. Our dedicated bare-metal GPU infrastructure and private networking blocks exist for exactly this purpose: fine-tuning, RAG pipelines, and inference workloads that stay inside a trust boundary you control, rather than one owned by the same company selling you the model. Our participation in the IPCEI-CIS and Cloud Campus initiatives is built around the same principle at an infrastructure level: keeping the pipeline decoupled from any single vendor’s ecosystem.
Infrastructure decisions made on the assumption that a vendor’s convenience today won’t cost you control tomorrow are worth revisiting.
A few questions about your own AI stack
Nadella’s essay was about a pattern, not a prescription for any one company. Applied to your own setup, it comes down to a few questions:
- As your AI usage grows over the next year, will your current API or consumption-based pricing protect your margins or erode them?
- Does your current AI architecture establish a private networking boundary where prompt context and feedback stay inside your own tenant, rather than the vendor’s?
- Is your model orchestration layer decoupled, so you could switch underlying models, including open-weight options like Mistral, without losing the domain knowledge your team has built up?
See where your workload’s utilization crosses the breakeven point.
Book a call to work the numbers with you live.