On 6 October 2026 CoreWeave entered India as the sole tenant of three 80 MW buildings at AdaniConneX’s Taloja campus near Navi Mumbai: 240 MW, an option to double to 480 MW, NVIDIA Vera Rubin on the CoreWeave platform, direct liquid cooling in the GPU halls, a non-evaporative chilled-water system, and a first phase expected in mid-2028. It is the most concrete AI commitment in this registry — a named tenant, a named host, a named silicon platform and a date. It is also the record most likely to be double-counted, because 240 MW here is capacity already inside AdaniConneX’s 1 GW platform, newly occupied, not newly built.
1At a glance
2The tenant, not the campus
This registry holds AdaniConneX’s 1 GW India platform, its 1 GW Maharashtra MoU and the Prestige Taloja campus as separate records. This is a fourth, different kind of record: a customer.
The distinction matters for two reasons that a supplier will feel directly.
First, it changes who specifies. When a campus is being built on spec, the developer chooses the cooling architecture and the vendor list. When a single tenant takes three whole buildings and runs a GPU cloud, the tenant’s reference design wins — and CoreWeave’s reference design is direct liquid cooling with high-density racks, not the conventional air-cooled halls that make up most Indian capacity.
Second, it changes the procurement language. The three buildings are being designed for high-density AI from the start. That means coolant distribution units, manifolds, rear-door or cold-plate systems, high-amperage busway, and substantially more power per square metre than the campus buildings that were commissioned in 2023 and 2024. A vendor with Indian references for air-cooled colo is not automatically an acceptable bidder here.
It is a named, creditworthy tenant committing to a specific megawatt figure at a specific campus with a specific silicon platform and a stated target year — the rarest thing in this registry.
It is not new capacity. Nothing is built. And the money behind it has not been published by any party.
3Three buildings of exactly 80 MW
Three buildings at 80 MW each is a clean, round, satisfying structure, and that is the first reason to look at it closely. Indian data-centre buildings are commonly sized at 20, 30, 40 or occasionally 60 MW per building. Eighty megawatts in a single building is at the upper end of what has been built anywhere in the country, and doing it three times at once, on one campus, for one tenant, is a materially harder engineering problem than three separate developers would face.
| Parameter | At 80 MW a building | Why it matters |
|---|---|---|
| Typical AI rack density | 100 to 160 kW per rack | 7 to 15 times a conventional Indian colo rack at 5 to 10 kW |
| Rack count at that density | about 500 to 800 per building | smaller halls, heavier floor loading, much denser cabling |
| Cooling | liquid, not air | air cannot move 130 kW out of a cabinet reliably at Indian ambient conditions |
| Power delivery | MV distribution inside the hall | busway and PDUs at a scale that changes the switchgear specification |
| Water | non-evaporative design | no evaporative loss, so the heat rejection is dry or closed-loop — larger plant, more electrical load |
| Connection | 80 to 96 MVA a building | 240 to 288 MVA for the campus: a transmission-level connection, not a distribution one |
The two engineering details in the release are the most informative sentences in the whole announcement, and they are easy to read past: direct liquid cooling and a non-evaporative chilled-water system. That combination says the developer has accepted a higher electrical overhead for heat rejection in exchange for zero evaporative water loss — a choice that is expensive and near-certainly driven by the water situation in the Taloja corridor.
It also tells you what will not be bought. No large evaporative cooling towers. No air-cooled CRAC fleet at scale. The bill of materials for these three buildings is not the bill of materials of the last ten Indian campuses.
4What 240 MW is in racks
Neither company has published a rack count, a GPU count or a per-rack power figure. Megawatts alone do not let a vendor size a package, so the arithmetic below is offered as a range with its assumption stated — not as a specification.
| Rack density | Racks at 240 MW | Comment |
|---|---|---|
| 100 kW per rack | 2,400 racks | conservative for a Vera Rubin generation hall |
| 130 kW per rack | about 1,846 racks | the middle of the range used above |
| 160 kW per rack | 1,500 racks | high density, liquid cooled throughout |
Even the bottom of that range is a hall count that would place this deployment among the densest AI facilities in Asia. It is 4 to 10 times the rack count of a conventional 240 MW colocation campus, which is the entire point of an AI cloud and the entire reason the cooling specification had to change.
Gridvik does not publish a GPU count. The platform has been named, the workloads have been described, and the number of accelerators that fit in 240 MW has not been stated by anyone. Any figure circulating with a GPU count attached has been invented by whoever circulated it.
5The number that is missing
Every other large record in this registry has a rupee figure and no megawatt, and we say so. This one is the reverse: 240 MW and no capital figure. Neither CoreWeave nor AdaniConneX has published one. Bloomberg’s "several billion dollars over the project’s lifetime" is the closest thing to a number and it is journalism, not a commitment.
| Basis | Implied capital for 240 MW | Position |
|---|---|---|
| $8 million per MW — shell, power and basic fit-out | about $1.9 billion | the floor for a building of this class |
| $12 million per MW — AI-ready with liquid cooling and dense power | about $2.9 billion | more consistent with the specification described |
| Including the compute inside it | not estimable from public data | Vera Rubin pricing and CoreWeave’s purchase terms are not disclosed |
The floor is useful and the ceiling is not, because the difference between a building and a building with GPUs inside it is a multiple, not a percentage. Note also who is paying for what. AdaniConneX builds and owns the campus; CoreWeave occupies it and pays rent over a long lease. The capital figures above are the developer’s problem, and the tenant’s commitment is a lease obligation whose value has not been published either.
240 MW at Taloja, added to AdaniConneX’s 1 GW platform, plus the 1 GW Maharashtra MoU, plus Prestige’s Taloja campus, quickly becomes a number larger than Maharashtra can physically energise.
Gridvik records this as a tenant reservation and adds nothing. The megawatts are already inside a counted platform.
6CoreWeave’s position at signature
Whether a 2028 target is credible depends mostly on the tenant’s balance sheet, and that part of the record is public because CoreWeave is listed.
| Measure | Value | Reading |
|---|---|---|
| Active power at Q2 2026 | about 1.5 GW | the base the business actually runs on |
| Contracted power at 11 August 2026 | about 4.2 GW | signed capacity not yet active |
| 240 MW as a share of contracted power | 5.7% | India is a material but not dominant commitment |
| 480 MW, with the option exercised | 11.4% | the option is the part that would make India core |
| First India phase | mid-2028 | roughly 20 months after signing |
| Platform named | NVIDIA Vera Rubin | a generation beyond what is installed in most Indian capacity today |
Two readings follow, and both are true.
The bullish one: a company with 4.2 GW of contracted power, a listed currency for raising capital and a single-tenant model that fills whole buildings is exactly the customer a host developer wants, and India gets 5.7% of its pipeline as a first entry. That is a serious commitment from a serious counterparty.
The cautious one: mid-2028 is twenty months away, the first phase capacity is unstated, and the buildings do not exist. Between signing and energising, a company that has grown this fast has to do a great deal of financing and construction in several countries at once. The option for a further 240 MW is not a phase; it is a right that expires or converts on terms nobody has published.
Also worth stating plainly, because the release does: CoreWeave will open an India office and hire locally, and the collaboration references the government’s IndiaAI Mission with no subsidy figure attached. Any incentive this project eventually receives has not been disclosed.
7Timeline
- CoreWeave releases its India entry: 240 MW at AdaniConneX’s Taloja campus, three 80 MW buildings, sole tenant, NVIDIA Vera Rubin, direct liquid cooling, first phase mid-2028, option for a further 240 MW. No capital figure
- Business Wire syndication and international coverage, including Bloomberg and DCD, repeat the deployment and describe the investment as several billion dollars over the project life
- Trade coverage adds the engineering detail: direct liquid cooling in the GPU halls and a non-evaporative chilled-water system. Inc42 and others describe a potential 480 MW footprint
- No construction start, building schedule, capital figure or first-phase capacity has been published. No land, power or environmental approval for the three buildings is in the public domain
8Procurement
9Assumptions and method
| Figure | Basis | Confidence |
|---|---|---|
| 240 MW | CoreWeave release, 6 October 2026. Sole tenant across three 80 MW buildings. IT versus facility basis not stated | High as a claim |
| 480 MW | The stated option to expand by up to a further 240 MW. Not a signed phase | High as an option, zero as a commitment |
| Mid-2028 | Company target for the first phase. First-phase capacity is not stated | Statement only |
| 5.7% and 11.4% | Gridvik arithmetic against CoreWeave’s roughly 4.2 GW of contracted power at 11 August 2026 | Arithmetic |
| 1,500 to 2,400 racks | Gridvik arithmetic at 100, 130 and 160 kW per rack | Arithmetic with stated assumptions |
| 80 to 96 MVA a building | Gridvik arithmetic at 1.0 to 1.2 times IT load | Estimate |
| $1.9 to $2.9 billion | Gridvik arithmetic at $8 to $12 million per MW, excluding compute | Estimate, not recorded |
| Additivity: no | The 240 MW sits inside AdaniConneX capacity already in the registry | Recorded as a rule |
10Limits and open questions
11Sources
Sources: CoreWeave investor release, CoreWeave newsroom, Business Wire, Bloomberg, Data Center Dynamics, ETDatacenters, Inc42, Data Centre Magazine, Converge Digest, The Next Web, TradingView, FourWeekMBA, FinanceFeeds.