Who Owns the Plug? - Data Sovereignty, in Conversation

By Raju Vegesna, Chief Evangelist, Zoho

Welcome to Who Owns the Plug?, a multi-format conversation series exploring data sovereignty's place in a world where AI reigns.

In the first episode, I sat down with Jon Reed of Diginomica for an unscripted conversation on one of the defining questions of the AI era: who really controls intelligence, and what happens if that control is taken away? From market instability and the centralization vs. decentralization debate to why "data residency" isn't enough in a world where models absorb your knowledge, this conversation covers the AI sovereignty questions every enterprise and nation should be asking right now.

My chat with Jon can be found here, and I've broken down its key elements below:

The Pendulum Is Swinging Back Toward Decentralization

Underneath the day-to-day noise sits a longer, more predictable pattern. Computing has always swung between centralization and decentralization: mainframes gave way to the PC, and then SaaS and cloud computing pulled everything back toward the center again. Intelligence is now retracing that same arc. We centralized fast around a small number of frontier model providers, and the pull back toward decentralization has already begun.

Data Residency Was the Wrong Question All Along

For years, the sovereignty conversation with governments and enterprises alike started and ended with one question: where does the data live? I've started pushing back on that framing directly. Picture your organization's data as a book. You can keep that book locked in a room inside your own country, and technically you've satisfied the residency requirement. But if someone walks into that room, reads the book cover to cover, and walks back out, the knowledge inside it has left with them. The book never moved. The intelligence did.

That distinction stopped being theoretical this summer. A leading U.S. AI lab had its most advanced models suspended worldwide by a Commerce Department export order, then restored several weeks later after negotiations with the government (CNN, June 30, 2026). Whatever your view of that specific case, it's a clear signal: when your organization's intelligence lives inside someone else's infrastructure, someone else's regulatory environment now governs your access to it. Data residency didn't protect anyone from that disruption. Intelligence residency is the harder, more important question, and it's the one more boards should be asking.

This is exactly why a growing number of enterprises are moving toward on-premise and self-hosted models. It isn't a compliance checkbox. It's an attempt to own the intelligence itself, not just the data it was trained on.

Every Layer of the Stack Wants to Commoditize Every Other Layer

Pull the AI stack apart and you'll find hardware, cloud infrastructure, the model layer, and the applications and services built on top of it. At every layer, the players are trying to push value toward themselves by commoditizing whatever sits next to them. Model companies would like you to believe the model is where all the value lives, so they're building their own data centers and their own chips. Cloud and hardware providers are pushing back by embracing open-source models to keep intelligence from becoming a walled garden. Application and service providers are doing the same thing from the other direction, customizing open models rather than renting a frontier one.

This is more than posturing. Google's own research on agentic systems found that the model itself accounts for only a fraction of real-world output quality; the surrounding context and harness do most of the work (Google's "New SDLC with Vibe Coding" playbook, May 2026). Enterprise infrastructure players are reaching the same conclusion from the ground up. Cisco has said outright that building its own models and routing intelligently across them is the more efficient path than defaulting to the largest available model for every task (Network World, 2026). When the industry's own infrastructure leaders are building around that assumption, it's worth enterprises doing the same.

Right-Size the Model, Then Push It to the Edge

The practical implication for enterprise buyers: you probably don't need the trillion-parameter model for most of what you're trying to do. A smaller model with the right context and the right harness will frequently outperform a larger, more expensive one on the task that actually matters to your business. That also means intelligence itself is portable in a way it wasn't eighteen months ago, capable of running closer to the edge rather than requiring a constant round trip to someone else's data center.
Where this matters most is knowing which domains are ready for AI today and which aren't. Large language models perform best in domains with fast, structured verification, like coding, mathematics, and other rules-based problems where you can check the output almost immediately. They're far less reliable in domains where the feedback loop is slow, like whether a piece of marketing copy converted or a strategy actually worked, because verifiability there takes months or years, not seconds.

Matching the tool to the domain, rather than applying it everywhere because it's the new hammer in the shop, is still the differentiator most organizations haven't figured out.

Human Value Goes Up, Not Down

None of this should read as a case for fewer people. If anything, the opposite is true. As agents absorb more of the repetitive, administrative work, the time that opens up should go toward more human-to-human interaction: customers, partners, and colleagues, not less. The agent becomes the interface between a person and the software, handling the chores so the person can spend more time on judgment, relationships, and creative problem-solving. Organizations that frame this as a straightforward productivity or headcount play are missing the more interesting opportunity in front of them.

Apprenticeship Matters More, Not Less

The next generation entering the workforce is fearless with these tools in a way that's genuinely valuable, and often asks the right questions precisely because they aren't anchored to "how we've always done it." But they don't yet have the domain expertise to catch it when a model confidently reverts to its own internal assumptions instead of the context they gave it. That's where mentorship earns its keep. Pairing new talent with experienced practitioners, so the questions get asked and the mistakes get caught, is a far better long-term strategy than quietly hiring fewer junior people because AI can produce a first draft.

Stay tuned for more Who Owns the Plug? in future weeks!