India's AI Data-Center Supercycle: Reliance, Meta and the Neocloud Gold Rush

The capital flowing into Indian AI data centres in 2026 has stopped looking like a build-out and started looking like a supercycle. Within four months, Reliance broke ground on a gigawatt-class campus at Jamnagar, Meta signed the first foreign-hyperscaler AI capacity lease in the country, and two independent "neoclouds" raised over a billion dollars between them to stand up GPU fleets. The headline numbers, $110B from Reliance, $100B over a decade from Adani, run into figures that would have read as implausible eighteen months ago.

Underneath the announcements sits a single constraint that determines who actually delivers: power. The Indian AI data-centre race is, in the end, an energy and execution race wearing a silicon costume.

Reliance and Meta turn Jamnagar into an AI campus

The anchor project is Reliance Intelligence's facility at Jamnagar, Gujarat, the same industrial complex that houses the world's largest single-site oil refinery. The first phase is 120 MW of AI compute, targeted for commissioning by end-2026, built on NVIDIA's GB300 Grace-Blackwell platform. Reliance has framed the campus as scaling from an installed base equivalent to more than 75,000 H100s toward 200,000 H100-equivalents, powered by the group's renewable-energy build-out. It forms part of a roughly $110B AI plan developed with NVIDIA and Google.

The validating event came on 10 June 2026, when Meta leased 168 MW of capacity at Jamnagar, the first time a foreign hyperscaler has contracted dedicated AI data-centre capacity inside India. The facility is seawater-cooled, drawing on Jamnagar's coastal location to reject the enormous heat loads that Blackwell-class racks generate. For Reliance, a hyperscaler tenant de-risks the build by underwriting capacity before it is fully filled with its own workloads. For Meta, it secures Indian compute close to one of its largest user bases while sidestepping the years-long grid-interconnection queues that constrain new capacity in the US and Europe.

What makes Jamnagar defensible is the stack of assets behind it. Captive renewable power, owned land at industrial scale, and Jio's telecom backbone for connectivity are precisely the inputs that are hardest for a pure-play data-centre developer to assemble. That bundle, power plus land plus telco, is the moat.

The neoclouds: Neysa and Yotta build the GPU layer

Below the strategic giants, a layer of independent AI-cloud providers, the "neoclouds", is raising the capital to deploy GPUs at scale and rent them out.

Neysa, the Mumbai-based AI-cloud company, raised $600M of equity on 16 February 2026 within a round of roughly $1.2B, led by Blackstone at an enterprise value near $1.4B. It is the largest Indian AI-cloud raise to date, and the capital is earmarked to deploy more than 20,000 GPUs. Blackstone's involvement signals that global infrastructure capital now treats Indian AI compute as a financeable asset class rather than a venture experiment.

Yotta, operating its Shakti Cloud, is deploying 20,736 liquid-cooled NVIDIA Blackwell Ultra B300 GPUs in an investment exceeding $2B, with the cluster targeted to go live in August 2026. Yotta is also a launch partner for NVIDIA DGX Cloud in India under a multi-year arrangement reported above $1B. Liquid cooling is not a luxury here: Blackwell Ultra parts draw enough power per rack that air cooling becomes thermodynamically and economically impractical, which is why every serious new Indian deployment, Reliance's seawater loop, Yotta's direct-to-chip liquid, is built around heat rejection from the start.

The neocloud model is straightforward in theory: borrow and raise to buy GPUs, sell capacity by the hour, and earn the spread over the cost of capital and power. The risk is equally straightforward. GPUs depreciate fast, NVIDIA's roadmap compresses each generation's useful pricing window, and utilisation must stay high to service the debt. These are infrastructure businesses with semiconductor-grade obsolescence risk bolted on.

Adani's decade-long bet and the export ambition

Adani has committed to $100B over a decade in AI data-centre infrastructure, with AdaniConnex targeting roughly 2 GW of capacity and a reported Google engagement around $15B. The group's logic mirrors Reliance's: it already controls power generation, ports, land banks and transmission, the unglamorous physical inputs that gate AI compute. Adani has also moved one step upstream, into an alliance with Jabil to manufacture AI server hardware in India, signalling that the largest players intend to capture more of the value chain than hosting alone.

The strategic prize both groups are chasing is the same. India consumes a large and fast-growing share of global AI workloads; the question is whether that compute runs on domestic soil or in Singapore and Virginia. The conglomerates are betting it runs at home.

The real constraint: power, grid and execution

For investors, the discipline is to separate megawatts announced from megawatts energised. A 120 MW or 168 MW AI hall is a large, concentrated electrical load that must be reliably fed, and India's grid, while improving, carries real reliability and interconnection risk in the volumes AI requires. This is the central reason captive power is decisive: Reliance and Adani can self-supply, largely from renewables backed by firming capacity, while developers dependent on grid offtake face both availability risk and the carbon-intensity problem of drawing from a coal-heavy network.

Execution risk compounds it. Securing GB300 and B300 allocation from NVIDIA, importing and installing liquid-cooling plant, commissioning at gigawatt scale and hitting utilisation targets are each non-trivial. Announced timelines, end-2026, August 2026, should be read as targets, not guarantees.

The policy tailwind and the investment frame

Two policy currents push the cycle forward. The DPDP data-protection regime creates data-localisation pressure that favours in-country compute, giving Indian capacity a regulatory tailwind that pure cost competition would not. And Budget 2026 introduced a data-centre tax holiday, improving project economics at the margin and signalling that the state regards AI infrastructure as strategic.

The investment frame that follows is specific. The durable winners are those who own the binding constraint, energy, rather than those who simply rent GPUs. Captive-power conglomerates (Reliance, Adani) carry the strongest structural moat; neoclouds (Neysa, Yotta) offer higher-beta exposure to compute demand but bear depreciation and utilisation risk and depend on power they do not control; and hyperscaler leases (Meta) validate demand while transferring build risk to local partners.

The supercycle is real, the capital is committed, and the data-localisation and tax tailwinds are genuine. But the bottleneck has moved from chips to electrons. In Indian AI infrastructure, the firms that control power and land, not those with the longest GPU order, are the ones positioned to convert these announcements into operating assets.