Express Computer
Home  »  Guest Blogs  »  Why networking will decide the future of sovereign AI

Why networking will decide the future of sovereign AI

0 0

By Avaneesh Vats, Vice President – Information Technology, Techno Digital

For much of the sovereign AI conversation, the infrastructure debate has centred on what countries need to build or control, including domestic compute, indigenous models, data and the data centres to host them.

In India, that ambition is increasingly taking physical shape. The IndiaAI Mission has expanded shared compute capacity to more than 45,000 GPUs as of June 2026. Installed data-centre capacity reached 1.57 GW in August, with nearly $70 billion in investment already underway and another $90 billion in announced projects. The government, meanwhile, is working with hyperscalers, data-centre operators and infrastructure providers to address the power, land and regulatory requirements of further expansion.

As that infrastructure takes shape, how its different components work together becomes increasingly important. India has seen the consequences of that interdependence in another critical digital system.

Around the time of UPI disruptions in 2025, India’s real-time payments system was processing an average of 591 million transactions a day. The reported causes varied across incidents, from telecom and network issues to hardware problems, bank-side processing loads and an SD-WAN issue. What the episodes demonstrated was not the failure of any one layer, but a characteristic of infrastructure at scale, the performance of the whole system depends on how reliably its different components communicate and work together.

For AI, that relationship becomes even more consequential. Compute, memory, storage, data and applications are in constant communication, and increasingly these resources will be spread across hyperscale data centres, enterprise infrastructure, private clouds, neoclouds and edge environments.
India’s sovereign AI challenge is therefore not only about building more compute. It is also about ensuring that all of this infrastructure can function as a system. That is where networking moves from being the infrastructure around AI to becoming part of the AI infrastructure itself.

Networking is Becoming Part of the AI Infrastructure
As AI clusters grow, enormous volumes of information must move between accelerators, memory and storage. Adding more GPUs therefore does not automatically create proportionately more useful compute. If the network cannot keep pace, accelerators spend time waiting rather than computing. This is already shaping infrastructure choices. Ethernet is gaining ground in AI back-end networks as advances in bandwidth, traffic management and congestion control make it better suited to connecting large GPU clusters. Meta, which expects to invest $115-135 billion in capital expenditure in 2026 largely to support its AI and core infrastructure, is among companies adopting Ethernet, alongside with Amazon, Microsoft, Meta and Oracle, even as InfiniBand continues to grow.

The challenge becomes broader as AI extends beyond a single facility. Training may favour concentrated compute, while fine-tuning and inference can require proximity to enterprise data, applications or users. Compute, data and applications can therefore sit across different infrastructure environments, making networking responsible not only for performance within the cluster, but for how workloads remain connected across locations.

For sovereign AI, that is where the question begins to change. Once workloads and data move across infrastructure environments, how they are connected also determines what they depend upon and how much control an organisation retains.

Sovereignty is About Control, Not Isolation
Sovereign AI does not mean that every layer of the technology stack must be manufactured domestically or disconnected from global technology ecosystems. Modern AI infrastructure inevitably relies on global supply chains, platforms and technology providers.

Sovereignty is therefore less about eliminating dependencies than retaining control over them. Organisations need to understand what their AI infrastructure depends upon and have the ability to manage, diversify or reroute around those dependencies when circumstances change.

This extends the sovereignty question beyond data residency. Knowing where data sits remain important, but so does knowing where it is processed, how it moves, which infrastructure and networks it traverses, and what happens when a provider, route or location becomes unavailable.
For sovereign AI, control over how infrastructure operates therefore becomes as important as control over where it resides.

Designing for Sovereignty: Compute, Connectivity and Control
This changes how AI infrastructure needs to be designed. Compute, storage and networking can no longer be planned as isolated layers; they need to be engineered as a system.

That brings three requirements together. Compute provides the processing capability. Connectivity ensures that data and workloads can move efficiently across the infrastructure without constraining that compute. Control provides the visibility, security and architectural flexibility to manage how those workloads operate and respond when conditions change.

Together, compute, connectivity , and control determine whether distributed AI infrastructure can perform efficiently, remain resilient and operate with the level of independence that sovereign AI requires.

India’s Sovereign AI Architecture Cannot Depend on Compute Hubs Alone
Much of India’s data-centre capacity remains concentrated around established hubs such as Mumbai, Chennai, Hyderabad, Bengaluru and Delhi NCR. AI demand, however, will increasingly come from BFSI, manufacturing, healthcare, government, GCCs and digital public infrastructure, with different requirements around latency, data sensitivity and resilience.

This does not mean decentralising every AI workload. It means having the architectural freedom to place workloads where they make operational sense without losing control over the infrastructure connecting them.

India already has more than 42 lakh route kilometres of optical fibre. The next challenge is to build on that foundation with high-capacity fibre backhaul, low-latency edge infrastructure and stronger links between data centres, cloud and enterprise environments.

For technology leaders, this changes the infrastructure question from simply which cloud or data centre to use, to which workload should run where and how it can remain connected and controlled.
India has an opportunity to design compute, connectivity and geography together, rather than build AI capacity first and solve connectivity later.

Sovereignty will be Tested in Operation
The real test of sovereign AI will not be whether every component of the stack is locally owned, but whether the system can continue to operate under sufficient control as workloads, technologies and infrastructure conditions change.

Can dependencies be understood and managed? Can workloads move without losing governance? Can infrastructure reroute around failure? These are the questions that will test sovereignty in practice. They also point to a different approach to infrastructure: networking cannot be solved after compute has been deployed. Compute, connectivity and control need to be designed together from the outset.

Sovereignty, ultimately, will be determined not simply by where AI infrastructure sits, but by the control retained over how it operates.

Leave A Reply

Your email address will not be published.