On 6 October 2026, CoreWeave announced plans for 240 MW of data-centre capacity at AdaniConneX’s Taloja campus in Navi Mumbai, across three 80 MW buildings. The company expects the first phase to be available in mid-2028. That makes this a notable AI-infrastructure commitment—but not 240 MW of computing capacity available in India today.
Executive summary
The CoreWeave–AdaniConneX announcement is one of the clearest recent signs that a specialist AI-cloud provider is planning a dedicated India footprint. CoreWeave says it will be the sole tenant across three planned 80 MW buildings at AdaniConneX’s Taloja campus in Navi Mumbai. The company intends to deploy NVIDIA’s Vera Rubin platform on its AI cloud, and says the GPU data halls are being designed for direct liquid cooling and a non-evaporative chilled-water system.
The public release gives a phased schedule: the first phase is expected to be available in mid-2028, with additional capacity coming online in phases. CoreWeave also says it has the opportunity to expand by up to another 240 MW at the campus. That additional block is an option—not a confirmed build-out. The announcement does not disclose a project capex figure, lease terms, MW measurement basis, GPU count, named customers or a commissioning certificate.
The most useful way to read the deal is as three linked commitments: AdaniConneX’s campus and data-centre infrastructure; CoreWeave’s role as the planned anchor tenant and AI-cloud provider; and a future-generation compute platform that is intended to be installed when capacity becomes available. Each part has a different delivery risk. The 240 MW figure should be quoted as the company’s stated data-centre capacity, not silently relabelled as IT load, installed GPU power or current operating supply.
For Indian customers, the project could eventually add another route to specialised AI compute, subject to delivery, product availability, commercial terms and service design. For suppliers and competitors, it points to the importance of high-density power, liquid-cooling capability, network integration and phased campus planning. For analysts, it is a pipeline item until an independently verifiable facility and service are live.
Contents
- The announced project at a glance
- What CoreWeave and AdaniConneX each bring
- What 240 MW does—and does not—tell us
- Why AI cloud infrastructure is different
- Cooling, power and the water question
- The delivery milestones to watch through 2028
- What the announcement could mean for Indian customers and suppliers
- Frequently asked questions
- Related Gridvik research
- Sources and disclosure
The announced project at a glance
| Question | What CoreWeave’s 6 October release says | Status to use in analysis |
|---|---|---|
| Where? | AdaniConneX’s Taloja campus in Navi Mumbai | Announced site; do not infer all site approvals or utility milestones from the release. |
| How much? | 240 MW of data-centre capacity | Planned capacity; the public release does not define the MW basis as IT load, total facility load or grid connection. |
| How is it divided? | Three buildings of 80 MW each | Planned configuration. |
| Who is the tenant? | CoreWeave says it will be the sole tenant across the three buildings. | Company statement; lease or commercial terms are not public in the release. |
| What workloads? | AI training, inference, reasoning and agentic workloads | Stated intended use for the CoreWeave AI cloud. |
| Which platform? | CoreWeave plans to deploy NVIDIA Vera Rubin | Future deployment intention; no GPU count or installation schedule by building is provided. |
| When? | First phase expected to be available in mid-2028; more capacity in phases | Forward-looking schedule, not an operating date. |
| Could it grow? | Opportunity to add up to another 240 MW at the campus | Expansion option, not a confirmed second 240 MW commitment. |
| What is undisclosed? | Project capex, lease price, detailed utility schedule, customer list and GPU count | Leave these fields as “not disclosed” unless a later primary source confirms them. |
The facts in this table come from CoreWeave’s announcement. The release is the primary source for the planned capacity and technical design, but it is also a corporate announcement. It describes intent and expected timing; it is not independent proof of construction progress or service availability.
What CoreWeave and AdaniConneX each bring
A data-centre campus and an AI cloud are related but different parts of the value chain. AdaniConneX is the infrastructure partner associated with the Taloja campus. CoreWeave is the company planning to use the three buildings as a dedicated base for its AI cloud platform. The arrangement means one announcement can describe a property/infrastructure development, a customer tenancy and a compute-service expansion at the same time.
That structure matters when counting capacity. The 240 MW should not be counted once as AdaniConneX’s new capacity, a second time as CoreWeave’s capacity, and a third time as an India AI-compute addition. Those may be different perspectives on the same campus. Likewise, the possible additional 240 MW should not be included in the initial plan as though it were already contracted or being built.
CoreWeave says it intends to establish an Indian office and hire locally, creating a presence to work with customers and partners. The release frames the project as an expansion of its AI cloud into India. It does not state that CoreWeave will own the Taloja land or the buildings, and it does not disclose the full commercial agreement with AdaniConneX. Avoid describing the arrangement as a confirmed property acquisition or publishing a project value that the primary announcement does not provide.
The sole-tenant description is significant for campus planning. A single large tenant can coordinate a particular accelerator stack, service architecture and deployment plan across buildings. It also concentrates initial demand in one customer relationship. But the release does not reveal the lease duration, take-or-pay terms, power pass-through, phased acceptance conditions or remedies for delay. Those are all commercially material unknowns.
For a broader explanation of how owners, tenants, cloud providers and equipment suppliers earn value at different layers, see Gridvik’s seven data-centre business models and India investment map.
What 240 MW does—and does not—tell us
A megawatt figure is not self-explanatory. Data-centre announcements can use MW for IT load, utility power, critical load or a broader facility figure. The CoreWeave release calls the number “data center capacity” but does not specify the measurement boundary. Until that is clarified, the clean formulation is “240 MW of planned data-centre capacity, as stated by CoreWeave.”
That caution prevents false comparisons. A state policy may define capacity as cumulative IT load. A market report may track operational IT stock. A campus plan may describe a utility connection or building design limit. Those measures should not be placed in one league table as if they were identical.
The announcement also does not disclose how much of the 240 MW will be available in each phase, the IT load per building, how the power is sourced, or how much capacity is reserved for the GPU systems versus supporting equipment. The 3 × 80 MW configuration is useful information about the plan, but it does not answer those questions.
A disciplined project ledger would record the announcement this way:
| Ledger field | Entry as of 10 October 2026 |
|---|---|
| Announced capacity | 240 MW, company-described data-centre capacity |
| Status | Planned / announced; the release is not commissioning evidence |
| Location | Taloja campus, Navi Mumbai |
| Tenant | CoreWeave, described by the company as sole tenant across three buildings |
| Delivery target | First phase expected in mid-2028; remaining capacity in phases |
| Optional expansion | Up to 240 MW additional, not a committed base-case amount |
| Project cost | Not disclosed in the primary announcement |
| IT-load definition | Not specified in the public announcement |
| GPU quantity | Not disclosed |
If a later filing, utility document or project update specifies the IT-load basis or commissioning status, update the ledger with the new date and source. Do not silently overwrite the original plan date; preserving the chronology makes it easier to distinguish announcements from delivered capacity.
Why AI cloud infrastructure is different
A conventional colocation customer may rent powered space and install a mix of servers chosen for enterprise, storage or cloud workloads. A specialist AI cloud provider sells access to accelerated-computing capacity, software, networking and operational support as a platform. That can create different requirements for rack density, cooling, internal fabric, data movement and software orchestration.
CoreWeave says the Taloja facilities are being designed for high-density AI infrastructure, including direct liquid cooling for GPU data halls. It says the platform is intended for AI training, inference, reasoning and agentic workloads. Those categories can have different usage patterns: model training can involve tightly coordinated large clusters; inference can serve repeated requests from users or applications; and agentic workloads may combine multiple model calls with tool use. The release does not provide a customer mix, utilisation forecast or product tariff for India.
The plan to deploy NVIDIA’s Vera Rubin platform is a forward-looking platform choice. It should not be described as GPUs already delivered, installed or available in the country. Hardware generations, software compatibility, supply, cooling requirements and customer demand can evolve between announcement and service launch. The infrastructure has to be ready for the design that is ultimately commissioned, not merely for the product name in a press release.
A practical buyer will eventually want to know more than the accelerator model. Important details include:
- which instance types and cluster sizes will be offered in India;
- whether customers can reserve capacity or use on-demand instances;
- network bandwidth and topology within a cluster and between regions;
- data-residency, privacy, security and audit arrangements;
- software environment, storage options and model-serving support;
- service-level objectives, incident response and maintenance windows;
- pricing, minimum commitments, taxes and currency terms;
- how capacity is allocated during a demand spike; and
- whether a workload can fail over to another Indian or overseas region.
None of these commercial answers can be inferred from a campus capacity announcement. They will become relevant as CoreWeave publishes an India product catalogue or begins contracting with local customers.
Cooling, power and the water question
The company says the Taloja GPU data halls are being designed for direct liquid cooling and a non-evaporative chilled-water system. That is a useful design signal: higher-density AI computing can generate concentrated heat that is difficult to manage with room-level air cooling alone. Liquid systems can move heat closer to the chips, while a chilled-water loop transfers heat to the facility’s heat-rejection equipment.
“Non-evaporative” should not be rewritten as “zero water use.” The announcement does not disclose the water source, make-up-water requirement, cooling-tower design, annual WUE, discharge plan or water-reuse percentage. A closed or non-evaporative system can reduce one type of water loss without establishing the total site water footprint. The full design, climate conditions, backup arrangements and operational rules still matter.
Power is equally central. Large AI clusters require a reliable electrical path from the grid or other supply arrangement through substations, switchgear, UPS systems and rack-level power distribution. A site’s nominal MW figure does not tell a reader whether every phase has an energisation date, how redundant feeds are designed, which upgrades are required, or whether renewable energy is contracted. CoreWeave’s release does not provide these details.
This is why the campus is a useful case study for Gridvik’s India data-centre power guide and connectivity/site-selection guide. The physical building is only one layer: power readiness, water, fibre, equipment delivery and customer operations have to arrive on compatible schedules.
A project update worth watching would disclose, or independently verify, the following:
- site or building permits and construction milestones;
- the planned energisation capacity and phase schedule;
- cooling-system design criteria and water-source plan;
- fibre and interconnection arrangements;
- initial GPU platform and service availability; and
- a commissioning or customer-acceptance milestone.
Until then, the appropriate status remains “announced plan with a stated first-phase availability target.”
The delivery milestones to watch through 2028
CoreWeave expects the first phase to be available in mid-2028, with additional capacity coming online in phases. That is a schedule expectation made in October 2026, roughly twenty months before the stated target. It gives customers and suppliers a planning horizon, but it leaves many interim milestones unspecified.
The project’s execution can be tracked as a sequence rather than a binary yes/no:
| Milestone | Evidence that would improve confidence | What it still would not prove |
|---|---|---|
| Site readiness | Confirmed parcel, permits, site works and responsible project entities | That power, water or fibre are fully delivered |
| Building construction | Dated construction progress, named phase or building and third-party confirmation | That the building is commissioned or customer-ready |
| Utility readiness | Written connection/energisation schedule and substations or upgrades completed | That GPU systems are installed and working |
| Cooling and network acceptance | Commissioning or performance tests for relevant plant and routes | That CoreWeave services are generally available to customers |
| Hardware deployment | Confirmed installation and cluster testing | That commercial capacity is available at scale |
| Service launch | India product listing, customer terms, region documentation or availability notice | That all 240 MW or optional capacity is online |
A supplier should not treat the expected mid-2028 date as a confirmed tender deadline. Package-specific timelines depend on the project’s design, qualification process, procurement route and construction sequence. A developer or customer should ask for a dated milestone plan and update it as evidence changes.
There may be capacity-accounting overlap with broader AdaniConneX campus announcements or national investment estimates. Keep the Taloja plan at the announced 240 MW level and record any later phase separately. Do not add the optional 240 MW until it becomes a firm, documented commitment.
What the announcement could mean for Indian customers and suppliers
For Indian enterprises and AI teams
The project may add a future domestic option for specialised accelerator compute. Local capacity can be relevant to latency, data handling, network design and customer support, but it does not automatically guarantee lower prices, better availability or a particular regulatory outcome. Buyers should compare the future India service with alternatives on workload fit, software, performance, security, geographic resilience and total cost—not only on campus size.
A team with a 2027 or early-2028 requirement should not count this campus as available capacity. A team planning beyond the stated first-phase window should keep it as a prospective option and seek written updates on product availability, cluster size and customer contracting.
For colocation and infrastructure operators
The announcement highlights a customer segment with requirements that may differ from ordinary enterprise colocation: concentrated load, high-density racks, liquid-cooling readiness and large-scale network connectivity. Operators can use it as a demand signal, not as proof that every AI customer wants the same architecture. Design flexibility, direct-to-chip liquid-cooling expertise, serviceability, power density, water planning and commissioning capability may be differentiators.
For equipment and engineering suppliers
Potential work could exist across power distribution, cooling loops, heat rejection, controls, monitoring, water treatment, fibre, construction, testing and commissioning. But a news announcement is not an open tender. Suppliers should verify the procurement owner, qualified-vendor process, package scope, design freeze and schedule directly before allocating bid resources.
For analysts and investors
Treat the plan as evidence of an announced tenant/campus relationship and a future supply pathway. It does not establish the project’s capex, lease economics, revenue, profitability or risk-adjusted return. The “option to add up to 240 MW” is potential upside only after evidence of exercise or a new binding commitment.
The cleanest way to avoid hype is to monitor a small set of measurable updates: power-delivery milestones, construction progress, utility connection, equipment installation, first customer service, and whether the optional expansion moves from option to commitment. Those observations are more informative than repeating the headline capacity.
Frequently asked questions
Is CoreWeave’s 240 MW India capacity operational now?
No operating status is established by the October 2026 announcement. CoreWeave says the first phase is expected to be available in mid-2028, with additional capacity in phases. Treat the full 240 MW as planned, not currently usable capacity.
Where will the CoreWeave–AdaniConneX campus be?
At AdaniConneX’s Taloja campus in Navi Mumbai, according to CoreWeave’s announcement. The release describes three planned 80 MW buildings.
What does “sole tenant” mean here?
CoreWeave says it will be the sole tenant across the three buildings. The announcement does not disclose lease economics, duration, power pass-through, acceptance conditions or other contract terms.
Is another 240 MW already committed?
No. CoreWeave says it has the opportunity to expand by up to another 240 MW at the campus. That is an expansion option, not part of the initial 240 MW plan unless later exercised and documented.
Will the site use NVIDIA Vera Rubin GPUs?
CoreWeave says it plans to deploy NVIDIA’s Vera Rubin platform on its AI cloud. The public announcement does not give a GPU count, delivery schedule by phase or evidence that the hardware has been installed.
Did CoreWeave disclose the project cost?
The primary announcement does not provide a capex figure, lease price or total investment amount. Do not insert an estimate as a reported fact without a separate, clearly attributed source.
Does non-evaporative cooling mean no water is used?
Not necessarily. The release names a non-evaporative chilled-water design but does not disclose total water consumption, water source, annual WUE or the complete cooling and heat-rejection design.
Related Gridvik research
- Karnataka’s Sustainable Data Centre Policy 2026–2031
- India data-centre connectivity: fibre and site selection
- India’s data-centre power bottleneck: grid readiness and policy
- India’s data-centre expansion: capacity, cities and AI
- India data-centre investment map
Sources and disclosure
Research checked 10 October 2026. This article relies primarily on CoreWeave’s own announcement, which is the source for the planned capacity, site, tenant description, cooling design, hardware intention and first-phase target. Corporate plans are forward-looking and can change.
Gridvik is an independent research firm. This article is informational and is not investment, legal, tax, procurement, engineering, energy-market or environmental advice. The illustration is a conceptual rendering and does not depict the Taloja campus.