Sovereign AI isn’t just a buzzword, it’s also severely misunderstood, being conveniently flattened into a binary test: is the data sitting inside the country or not?
That test isn’t wrong, it’s just insufficient. I put this to Andrew Winney, General Manager and Global Head of Cloud Product Management at Tata Communications, because his company sits at an unusual vantage point. Tata Communications has spent decades building and maintaining undersea cables that connect the global internet – over 500,000 km through which 30-35% of internet traffic flows every second.
“We identified in this journey that we have a lot of distinctive competencies that allows us to do a lot of these services, especially our play around infrastructure, such as cloud, cybersecurity, a lot better and stronger,” Winney told me, before making the connection that explains the whole strategy: “You look at any large hyperscalers, they want strong network.”
That is a telling sentence. Because the hyperscalers are all more or less Tata Communications’ customers. Which makes the company’s own cloud play – the Vayu Cloud Platform, its AI Cloud and AI Studio layers – simultaneously complementary and competitive. For Tata Communications, it allows them to do something very unique that goes beyond just data residency requirements in the age of sovereign AI.
Data residency compliance requirements predate COVID, to be honest, so why is it being revisited now in the need for sovereign AI? Winney’s response goes beyond just marketing slogans.
Also read: Before AI takes over, fix the wiring: Tata Communications’ infrastructure warning
“Sovereign infrastructure and sovereign AI are not the same. There are a lot of differences,” he said. “When we spoke about sovereign infrastructure, we would have looked at where your data is stored, and where your infrastructure is really present.”
That was the old test, in terms of localising your data in a database hosted on a server within national borders. But according to Winney, AI stacks four more layers on top of data alone – like compute, storage, GPU, etc – and each one can leak sovereignty independently, which is the real risk.
The model layer is where it starts getting uncomfortable. “If your model is still communicating some of the data that you are using outside the country, your purpose is defeated. The sovereignty should also be looked at from a model perspective,” Winney emphasizes.
“When I talk about a platform, it isn’t just a data plane that has to be sovereign. The control plane of how is this platform controlled from? Are these also controlled from India?,” Winney asks as the million dollar question. “It is not really about where the data is, but who has keys to the data. And where are folks located who have access to the data, who can access the data? All of these are today important, a lot more important for AI to be truly sovereign.”
None of this is abstract for Indian buyers, and according to Winney the demand is already commercial rather than merely regulatory: “Many of them are now looking at sovereignty. This could be driven by regulations. This could be driven even by cost of ownership, or this could be driven by other concerns,” he points out.
Going back to Winney’s five-layer problem of sovereign AI, an organisation that hasn’t worked out its data layer isn’t going to be sovereign AI compliant just because it fixed the other layers. And that is what’s currently killing most of the Indian enterprise AI projects from demo to deployment..
Winney suggests Tata Communications built what he describes as, “the first GPU as a service in India around 2023. When there were a very limited number of competitors at that point.” Three years of watching customers attempt the same transition has given Winney deep insights into what they do wrong.
Also read: Microsoft’s Sovereign AI cloud push and its India significance explained
“As of today we have crossed the stage of planning for AI. We have crossed the stage of working on pilots. Most of the enterprises are moving into full-on production from pilots.” But moving into production and surviving it are two completely different things, as Winney identifies two failure modes.
The first is infrastructure planning. “When you do a pilot with a certain amount of data, in a simulated environment, you are not able to reflect or understand what would be the impact when you take it production grade,” he explains. The pilot succeeds precisely because it was small, and its success tells you almost nothing about the economics at scale.
The second is more damning, because it’s self-inflicted. “In many cases enterprises overestimate the cleanliness of the data that they have, and the availability of data itself, which becomes a hurdle for them as they move into this production phase,” according to Winney.
The wastage that follows isn’t trivial, and Winney’s second example of it is genuinely counterintuitive. GPUs are the scarcest resource in this market, and yet, “When you do pilots, it is not that your GPUs utilize 24×7. So these are typically used for a few hours a day, and after that your GPUs would be idle,” Winney highlights. This means customers keep paying for the compute regardless. The country’s most contested resource is being rented by the month but only used by the hour, suggests Tata Communications.
“The bottleneck is across multiple layers even today, we are all GPU constrained. In India, the demand for GPU outnumbers the available GPU by a lot, that’s the reality of the market today,” Winney says. Sure, this isn’t a surprise anymore, given NVIDIA’s stock price over the past year or so. Except the interesting part is who’s demanding more GPU access.
“If you look at it in the last few months, we actually see many customers from outside the country actually wanting to come in and consume GPUs based out of India, for training requirements where typically your latency or delivery of service to customers is not really a problem,” Winney says.
Think about it for a second. India, the country that has spent three years worrying about whether it has enough compute for its own ambitions, is now fielding demand from foreign customers who want to train AI workloads inside GPU clusters inside India. This is against the traditional narrative of India being a consumer of somebody else’s GPU training cycles.
Where does an infrastructure provider fit into a five-layer sovereign AI problem? All of them, is Winney’s answer, or you’re just selling capacity.
“If you have thousands of GPUs, you are considered an AI factory because you generate intelligence. That’s how NVIDIA really defined it,” recounts Winney, before suggesting what comes next. “From your AI factory where you manufacture intelligence, AI grids are essentially going to be where you deliver intelligence.”
Also read: Yotta to Adani: India building sovereign, frontier AI with Global South relevance
In an AI factory where training happens centrally at scale, going forward inference has to happen close to the user, suggests Winney. For a company like Tata Communications, which already owns globally distributed points of presence, it allows them to make every edge node a GPU-powered inference layer.
And that brings me to my favourite piece of jargon from the entire conversation. “I’ll throw you one more jargon, now customers are talking about what they call a token path, which essentially is what it takes to generate a token and getting it consumed.”
The token path is a genuinely useful frame, because it exposes where the accountability gaps are. That’s the bet Tata Communications is making. “With the capabilities that we have with the cloud and platform on one side, the edge capabilities and the network that sort of connects it all, I think we got the complete token path covered, which means we’ll be able to provide better experiences,” Winney emphasizes.
Whether owning the network turns out to be a durable moat or merely a nice-to-have is the open question – but there is one data point Winney offers that’s hard to ignore. “We work a lot with even our government. A significant percentage of our capacity today is leveraged by our government for many sensitive use cases. That also shows the trust our infrastructure has and what we have proven with them.” India’s most sovereignty-sensitive customer has already voted.
The broader lesson from this conversation isn’t about any one provider. That data sovereignty isn’t equal to sovereign AI. The harder work of defining control planes, enforcing model boundaries, and who holds the keys to every Indian enterprise’s kingdom and where they live – is the part that needs to be relooked and re-examined going forward.
Also read: Infrastructure control is key for sovereign AI stack, says IBM