CoreWeave India Entry: 240 MW AdaniConneX AI Data Center

On October 7, 2026, CoreWeave announced its entry into India through a collaboration with AdaniConneX: 240 megawatts of capacity in Navi Mumbai, to be filled with NVIDIA Vera Rubin systems, with a first phase expected in mid-2028. For the neocloud that has spent 2026 racing to be first on Nvidia's newest rack, the CoreWeave India entry is a different kind of milestone. It is less about hardware generations and more about where AI compute gets built, who builds it, and how long it takes to switch on. This piece sets out what CoreWeave's release says, puts 240 MW in context against India's existing data center base, translates it into GPUs with clearly labelled assumptions, and looks at the risks that sit between announcement and capacity.

240 MWCapacity in Navi Mumbaithree 80 MW buildings

Mid-2028First phase expectedlater capacity in phases

+240 MWOption to doubleat the same campus

4.2 GWCoreWeave contracted poweras of Aug 11, 2026 earnings call

NVIDIA

Microsoft

Google

AWS

The platform and the hyperscalers whose India commitments frame the deal.

What CoreWeave announced

According to CoreWeave's press release, the 240 MW will be delivered as three 80 MW buildings at AdaniConneX's Taloja campus in Navi Mumbai. CoreWeave will be the sole tenant across the three buildings and can double capacity by adding up to 240 MW at the campus. It plans to deploy the NVIDIA Vera Rubin platform for training, inference, reasoning and agentic workloads, to open a local office and to hire in India. The first phase is expected in mid-2028, with more capacity arriving in phases. The facility is described as built for high-density AI, with direct liquid cooling for GPU halls and a non-evaporative chilled water system.

Two things the release does not give are worth noting. The Tribune India's report states that no deal value, investment amount or lease terms are disclosed, and the release does not name the power source for the site. Bloomberg also covered the deal, but its article was not accessible to us, so nothing here relies on it.

AdaniConneX is, per the release, a 50:50 joint venture between Adani Enterprises and EdgeConneX that aims to build a 1 GW data center platform in India. COO Sachin Jain said in the release: "India has the talent, developers, and ambition to become a major center of AI innovation." The release also says the move follows CoreWeave's recent entry into Indonesia and supports national goals including the IndiaAI Mission.

How big is 240 MW in India?

India's data center base is small relative to the ambition. A March 18, 2026 Ministry of Electronics and IT statement, reported by TelecomTalk, put national capacity at over 1,500 MW in 2025, up from 375 MW in 2020. Mumbai and Navi Mumbai held the largest share at 790 MW, roughly half the national total, ahead of Chennai (305 MW), Bengaluru (182 MW) and Hyderabad (152 MW).

On those figures, 240 MW is about 16% of India's 2025 national capacity and about 30% of the Mumbai region's, both our arithmetic. The comparison is imperfect: the government numbers count capacity in service in 2025, while CoreWeave's is a planned build that starts arriving in 2028, and the option to double would take the total to 480 MW. Still, a single-tenant block of this size is large by local standards, and it lands in the market that already hosts the most capacity.

A single-tenant 240 MW block is large against India's roughly 1,500 MW base, but it starts arriving only in mid-2028.
Metir analysis of MeitY and CoreWeave figures

The wider India buildout

Hyperscalers have been committing far larger headline sums. Google said on October 14, 2025 that it would invest $15 billion over five years in an AI data centre in Visakhapatnam, with an initial 1 GW of capacity and Adani Group and Airtel as partners. Microsoft announced $17.5 billion over four years on December 9, 2025, its largest investment in Asia. Amazon said on December 10, 2025 it would invest more than $35 billion by 2030 in India's cloud and AI infrastructure.

Headline hyperscaler commitments to India

Announced figures in US dollars. Scopes and time frames differ (Amazon's runs to 2030 and Microsoft's over four years), so the bars are not like-for-like. CoreWeave did not disclose a deal value.

Amazon$35B+

By 2030, cloud and AI infrastructure

Microsoft$17.5B

2026 to 2029, cloud, AI, skilling, operations

Google$15B

Five years, AI hub in Visakhapatnam

These figures are not directly comparable. They cover different time frames and scopes, and none of the sources split out how much goes to data centers versus skilling, operations or other spending. CoreWeave's deal has no disclosed dollar value at all, so it can only be compared in megawatts. What the pattern does show is that Adani-linked infrastructure appears in both Google's and CoreWeave's India plans, which makes AdaniConneX a recurring local partner.

On the public side, the IndiaAI Mission has an outlay of Rs 10,372 crore and, per a government release dated April 2, 2026, has onboarded more than 38,000 GPUs to a common compute facility offered to startups and academia at an affordable rate. That is a pool of GPUs the state can allocate, not a count of data center power. A private, single-tenant build like CoreWeave's serves a different buyer base: commercial AI labs and enterprises that rent capacity.

What 240 MW means in GPUs

There is no official GPU count for the deal, so any number is an estimate. One public reference point is Supermicro's Vera Rubin design, which lists 1,152 GPUs per 5 MW power envelope, sized at 227 kW per rack and scalable from 5 MW to 1 GW. Our assumption is that CoreWeave's 240 MW is comparable facility power and that its racks resemble this design.

~55,000Rubin GPUs at 240 MW48 x 1,152, our estimate

~110,000If the 240 MW option is exercised96 x 1,152, our estimate

72GPUs per NVL72 rack16 racks per 1,152 GPUs

Dividing 240 MW by 5 MW gives 48 units, and 48 times 1,152 is 55,296 GPUs; doubling gives 110,592. Treat these as order-of-magnitude figures. The actual count depends on how CoreWeave defines the 240 MW (facility power or IT load), cooling efficiency, rack density, and whether Nvidia counts GPU packages or dies, which has varied by product generation. Capacity may also be filled in stages, so the full figure will not exist in 2028.

Aerial view of residential towers beside mangroves and a creek in Navi Mumbai — An aerial view of Navi Mumbai, the region where the campus sits. It is a general city view, not the Taloja campus or any data center. Photo by Anurupa Chowdhury via Wikimedia Commons, CC BY 3.0.

Why India, and why a neocloud now

For CoreWeave, India extends a geographic pattern. The release says it follows the Indonesia entry, and its contracted power stood at about 4.2 GW as of the August 11, 2026 earnings call, which our Q2 2026 earnings analysis covers. Other neoclouds are also widening their footprint: Firmus is building a 170,000-GPU cluster in Batam, Indonesia, and our look at neocloud gigawatt buildouts shows how much of the sector's growth now depends on securing power years ahead.

Data rules are a plausible driver, though CoreWeave does not cite them. India's DPDP Rules, notified in November 2025, phase in over 12 to 18 months and allow cross-border transfers subject to government-prescribed conditions. Significant data fiduciaries must keep categories of personal data, still to be notified, within India. Until those categories are defined, how much AI workload must run domestically is an open question, but local capacity reduces exposure to it.

Power and timeline risk

The binding constraints are not chips. First, timing: mid-2028 is roughly 20 months away, and the release gives no schedule for the later phases. Second, power: the release does not say where the energy comes from, and the government's own figures show a market where one metro holds half of all capacity, concentrating grid demand. Third, the earlier Vera Rubin rollout shows how quickly hardware generations turn: by the time capacity arrives in 2028, today's Rubin may be a prior-generation platform. The doubling option is just that, an option, and the release does not say what triggers it.

For teams building on AI models rather than renting racks, the practical angle is regional availability. More in-country capacity should eventually widen where inference can run, and a model-agnostic workspace such as Metir AI lets you route to whichever provider or region fits your data requirements as that capacity appears.

What to watch

Phase schedule. Whether the first buildings reach customers in mid-2028 and how later phases are staged.
Power sourcing. Grid connection and any renewable contracts for Taloja, which the release does not describe.
Anchor customers. Which Indian or global AI companies commit to the capacity.
Option exercise. Whether CoreWeave adds the second 240 MW.
Data rules. Which data categories the government notifies for localisation.

The takeaway

The confirmed facts are 240 MW across three 80 MW buildings at Taloja, a sole-tenant arrangement, Vera Rubin as the planned platform, a first phase in mid-2028 and an option for another 240 MW. Everything about price, power source and GPU count is undisclosed or estimated. The deal is best read as a long-dated capacity reservation in India's largest data center market, set against much larger but less comparable hyperscaler pledges.

Sources:

Image credits

Header and in-body image: Navi Mumbai skyline by Anurupa Chowdhury via Wikimedia Commons, licensed CC BY 3.0. It is a general aerial view of Navi Mumbai, not the Taloja campus. Reviewed before use.