I recorded the sixth installment of Who Owns the Plug with Raju Vegesna yesterday, an hour on why sovereignty is a cost most organizations never price out. This morning I read Himanshi Lohchab's report in The Economic Times on six Nvidia-backed neoclouds now in talks to lease Indian data center capacity, and the two conversations would not stay separate. I met Himanshi at Canva Create earlier this year, and her byline is one I read closely. CoreWeave, Together AI, Nebius, Lambda Labs, Crusoe and Nscale have opened talks with Sify Technologies, Yotta Data Services, CapitaLand Investment, CtrlS Datacenters and Tata Communications to lease infrastructure for AI inference workloads, drawn by a 20-year tax holiday, deep engineering talent and infrastructure costs well below the alternatives (The Economic Times, 2026).

Vegesna spent an hour asking who owns the plug. Lohchab's report hands that question to India's data center layer, and the answer splits the plug in two.

Six companies that rent out GPUs, one supplier behind all of them

A neocloud is a cloud provider built around one product: renting out banks of graphics processing units by the hour, rather than the broader menu a hyperscaler like AWS or Azure sells. CoreWeave, Together AI, Nebius, Lambda Labs, Crusoe and Nscale are the best-known names in that category, and Nvidia sits behind every one of them, as an investor, a financing partner, or a company that has guaranteed to buy back capacity these firms cannot sell elsewhere.

What they want in India is inference capacity, the stage where a trained AI model answers a question or generates an output, rather than training, where a model learns in the first place. Inference is the workload that scales with usage, so it needs to sit close to the customers generating that usage. That is the pull toward India: a market with its own growing base of AI users, plus the cost and talent advantages the report names.

None of the six plan to build a facility. They want to lease space inside data centers that Sify Technologies, Yotta Data Services, CapitaLand Investment, CtrlS Datacenters and Tata Communications already operate, rack the GPUs there, and start serving inference traffic without the years a ground-up build would take.

The tax holiday was written for owners, not tenants

India's Union Budget for 2026-27 created a 20-year tax holiday, running until 2047, for foreign companies that serve customers outside India through a data center notified by the Ministry of Electronics and Information Technology (MeitY, 2026). The exemption clears a specific risk: a server rack in Mumbai creating a taxable presence for a company headquartered in San Francisco or Amsterdam. Lawmakers wrote the language around companies willing to build or co-locate a facility and route global traffic through it.

Six neoclouds want the access without the capital commitment. A signed lease is not the same filing as a facility investment, and inference workloads for offshore customers sit in a category Finance Ministry guidance has not addressed directly. Nobody in New Delhi has confirmed, in public, whether a leased footprint inside a qualifying facility clears the same bar as an owned one.

The domestic operators own the plug. Nvidia owns the terms.

Sify, Yotta, CapitaLand, CtrlS and Tata Communications hold the physical plug: the land, the power contracts, the building. What they would be renting it to is six tenants already wired into one upstream supplier for chips, and in several cases for financing and revenue guarantees behind that supplier. Venture money sits alongside Nvidia in most of these companies. Coatue, Jane Street and Fidelity back CoreWeave. Accel and Orbis back Nebius. Founders Fund, G2 Venture Partners and Valor Equity back Crusoe. That capital diversifies the cap table without touching the dependency that matters here, which is that every tenant on the list runs on Nvidia silicon financed in part by Nvidia itself, a pattern that has drawn its own scrutiny over the past year as neocloud lease obligations have grown large enough for Nvidia to co-sign them directly (Data Center Dynamics, 2026).

A default that starts at the GPU layer now reaches into the property layer. If a leased inference cluster inside a Mumbai facility sits idle because a tenant's revenue guarantee never triggers, the Indian operator holds a lease obligation, not a chip it can resell into a tight market. None of the five domestic operators named in the report has previously underwritten counterparty risk at this concentration.

Unknowns and uncertainties

The Economic Times report describes talks, not signed leases. Deal size, timeline and which specific facilities are under discussion remain unconfirmed. Whether the tax holiday's offshore-customer condition extends to inference workloads served through leased rather than owned capacity has not been tested in public by any of the six neoclouds, and MeitY has not issued guidance for the leasing scenario.

A default that starts at the GPU layer now reaches into the property layer.