By Sudhir Kunder, CBO of DE-CIX India
As AI adoption surges, Indian enterprises face a squeeze: limited data center space, rising power demands, and dependence on hyperscalers. Ultra Ethernet is changing the game – enabling distributed private AI infrastructure, reducing vendor lock-in, and powering the rise of AI Internet Exchanges. How India can turn these challenges into a competitive edge. And why neutral interconnection platforms like DE-CIX help enterprises to scale securely while keeping control of their data.
Driven by growing concerns around data sovereignty, compliance, and control, companies across the globe are rethinking how they build and manage AI infrastructure. McKinsey projects that by 2030, 35–40% of AI workloads in Europe and the U.S. will be deployed privately. A survey by data center specialist Vespertec found that a quarter of organizations already use their own hardware, while another 70% are planning to do so.
For Indian enterprises, the motivation is clear: retaining ownership of critical infrastructure while reducing dependence on hyperscalers. Whether run on-premises or within colocation facilities, private AI allows companies to safeguard sensitive data, avoid vendor lock-in, and retain full operational control. Yet, as AI adoption accelerates, the biggest hurdles for enterprises are not only cost but also the availability of suitable data center space and power – both of which are scarce in fast-growing markets like India. This scarcity directly limits how quickly infrastructure can be scaled.
The capacity crunch in India’s data centers
One of the most pressing challenges is power. Running GPUs for AI requires far more energy than traditional workloads. Legacy enterprise data centers, often capped at 5–10 kW per rack, fall short of the 50–100 kW per rack demanded by modern AI deployments. Meanwhile, India’s booming data center industry faces delays due to grid constraints and high land acquisition costs. Where new facilities do get built, hyperscalers and cloud giants often secure entire sites through pre-leasing, leaving little capacity for enterprises.
Even global operators are struggling: some are resorting to investing in their own power generation just to keep up with AI demand. Vacancy rates in primary markets have dropped to record lows – CBRE reports single-digit availability in Europe and the U.S. – and India’s Tier I hubs such as Mumbai, Chennai, and Hyderabad are facing similar pressures.
Breaking away from centralized AI training
Traditionally, AI training required highly centralized, power-intensive data centers. This has been driven largely by InfiniBand, the interconnect protocol linking GPUs in AI clusters. InfiniBand provides extremely high bandwidth (3.2 Tbps per processing unit) and ultra-low latencies measured in microseconds. But it comes with restrictions: GPUs must be located just a few meters apart, forcing companies into tightly packed, centralized facilities.
InfiniBand also comes at a high cost, with limited vendor options and little flexibility for distributing workloads geographically. While Ethernet has long been cheaper and more widespread, it lacked the precision and reliability for AI workloads – until now.
The launch of Ultra Ethernet (UE) Version 1.0 in June 2025 marked a turning point. Optimized to minimize packet loss and congestion, UE enables Ethernet to compete directly with InfiniBand for AI training inside data centers. More importantly, the next UE versions, expected in 2026, will go further – supporting distributed AI training across facilities in a metro region. This promises to break the rigidity of centralized AI and unlock new models of infrastructure design.
Towards distributed private AI in India
Future releases of Ultra Ethernet will allow distributed AI clusters to operate reliably even with latency up to one millisecond. This means enterprises will be able to spread training workloads across multiple facilities in a city while maintaining performance. For Indian enterprises, where rackspace and power are unevenly available across markets, this flexibility could prove crucial.
The benefits extend beyond scale. Because UE builds on widely adopted Ethernet standards, enterprises can leverage existing Internet Exchange (IX) hardware, where software upgrades can be rapidly deployed once vendors roll them out. This lowers costs, simplifies operations, and accelerates adoption. For companies pursuing private AI strategies, it also reduces dependency on a single colocation provider, allowing them to aggregate capacity from multiple data centers and strengthen their negotiating position.
The role of AI Exchanges in the Indian market
The natural next step is the AI-Internet Exchange (AI-IX): a neutral interconnection hub designed to bring together Ultra Ethernet, Internet service providers, and AI routing technologies across a distributed data center landscape. An AI-IX enables direct, low-latency connectivity between data center sites, ensuring optimal data pathways for both training and inference.
For India, where enterprises are increasingly focused on regulatory compliance and protection of intellectual property, the AI-IX model offers a clear path forward. It provides the performance needed for distributed AI training while meeting governance requirements. McKinsey notes that C-level executives worldwide are driving AI adoption decisions – spanning data protection, systems integration, and vendor strategy. In India, this trend is set to accelerate as enterprises look to keep sensitive data within national borders, comply with local regulations, and gain a competitive edge.
India’s private AI moment is approaching
Private AI, powered by Ultra Ethernet and enabled by AI Exchanges, positions Indian enterprises to meet these demands head-on. By combining scalability with sovereignty, distributed AI infrastructure can help India’s businesses not only adapt to the AI era but lead it.
