Data Centres
A capacity business priced per megawatt, built on ten-year leases and decided by whether an anchor client signs early. Four business models sit under one label, and they earn very different returns on very different risk.
Capacity is contracted, not sold. The question is which model the company is actually running — colocation yield, cloud build, IaaS/PaaS, or managed services — because the capital intensity and return profile differ by an order of magnitude.
Capacity measured in megawatts.
1,263 MW
Installed India capacity — the base from which the build-out runs.
3,000 MW
Roughly 2.4× the 2025 base in three years. Execution risk sits here.
4,500 MW
The longer-dated target; directional rather than contracted.
Who takes the space
- Hyperscalers~50% of end-user demand — AWS, Azure, Oracle, Google Cloud. Large blocks, long tenures, hard negotiation.
- BFSI, IT, healthcare~22%. Regulated workloads with data-residency needs.
- Enterprises & GCCsOwn hardware and software installed into leased space; increasingly repatriating from public cloud.
How space is sold
- Shelf
- Rack
- Cage
- Hall
- Floor
- Whole building
Rents run roughly $70–90/kW for bulk small space and $90–110/kW for large, customisable by size. The underlying driver is data sovereignty — processing and hosting data domestically rather than paying an offshore provider.
Four models, four return profiles.
| Model | What is sold | Economics | Principal risk |
|---|---|---|---|
| Colocation | Rack space, power, cooling, bare shell. Fixed rental per kW per month. | RoE 10–15% EBITDA 60–70% ₹80–100mn revenue/MW |
High capex, leasing risk, rental pressure from oversupply |
| Cloud build | Bare-metal compute, storage and networking layers. | RoE 10–20% Capex ₹800mn–1bn/MW |
Technology obsolescence, underutilisation, security requirements |
| IaaS / PaaS | Virtualised compute and storage with orchestration; prebuilt DB, analytics, container and DevOps environments. Elastic billing or subscription. | Competes directly with public cloud | Margin pressure from hyperscalers, scale disadvantage, integration complexity |
| Managed services | End-to-end IT management — monitoring, backup, OS, security, compliance. Subscription, AMC or project. | RoE 25–30% — highest of the four | Hyper-competition, people-intensive, execution-dependent |
The build cost frame. One MW costs roughly $6–7mn — some 30–40% cheaper than the US and around 18% below the Asia-Pacific average. Land, mechanical and electrical account for 55–60% of that, statutory costs 5–7%, and operational costs 15–20%. In rupee terms a 1 MW installation runs about ₹60–70 crore.
The payback frame. An indicative unit generates around ₹1 crore per month against roughly ₹15 lakh of operational expense, with minimum ten-year leases signed. Securing an anchor client compresses payback to two to three years at a 12–18% ROI. Without one, the same asset is a long-dated bet on absorption.
Anchor, power, and mix.
The anchor client
Pre-leased capacity is the difference between a two-to-three year payback and a speculative build. Contracted MW ahead of commissioning is the number that de-risks the whole model.
Power & efficiency
Power is the largest running cost and the binding constraint on expansion. PUE, access to cheap and preferably renewable power, and cooling design decide the operating margin.
Moving up the stack
Colocation yields 10–15% RoE; managed services 25–30%. The operators that earn a premium multiple layer higher-value services onto the same capacity base.
What to answer before underwriting.
- →Which model, really? Colocation, cloud build, IaaS/PaaS or managed services — and what is the revenue split? The blended return depends entirely on this.
- →Contracted vs commissioned MW. How much of live and under-construction capacity is already leased, and to whom?
- →Anchor tenant. Is there one, on what tenure, and what share of capacity does it take? What does payback look like without it?
- →Customer mix. Hyperscalers versus enterprises versus cloud providers. Hyperscalers bring volume and negotiating power in equal measure.
- →Realised rent per kW. Against the $70–110 range, and how it has trended as new supply lands.
- →Capex per MW. Against ₹60–70 crore, and whether land is owned or leased.
- →Power. Cost per unit, sourcing arrangement, renewable share, and whether grid capacity constrains the expansion plan.
- →Lease structure. Tenure, escalation clauses, and whether electricity is a pass-through or a margin risk.
- →AI workload readiness. GPU-as-a-service is gaining traction; does the facility have the density and cooling to support it?
- →Funding of the build. How is the capex financed, and what happens to returns if absorption lags the schedule by a year?
What to monitor, quarter by quarter.
| KPI | Calculation / source | Benchmark or read-through |
|---|---|---|
| Live vs contracted MW | Company disclosure | The core capacity metric; contracted-ahead is the de-risking signal |
| Utilisation % | Leased MW ÷ commissioned MW | Empty MW carries full fixed cost — the fastest way to destroy returns |
| Revenue per MW | Revenue ÷ operational MW | ₹80–100mn per MW annually for colocation |
| Realised rent per kW/month | Rental revenue ÷ contracted kW | $70–90 bulk small, $90–110 large; falling rents signal oversupply |
| EBITDA margin | Excluding electricity pass-through | 60–70% for colocation; materially lower if power is not passed through |
| PUE | Total facility power ÷ IT power | 1.67–1.8 typical enterprise, ~1.1 best hyperscale. Directly drives opex |
| Capex per MW | Capex ÷ MW added | ₹60–70 crore ($6–7mn); overruns compress the whole return |
| Payback period | Years to recover capex per facility | 2–3 years with an anchor; much longer without |
| Anchor client concentration | Revenue from largest tenant | De-risks payback but concentrates renewal risk |
| Weighted average lease tenure | Contract schedule | 10-year minimums are standard; a clustered expiry is a re-leasing event |
| Power cost per unit | Electricity cost ÷ units consumed | The largest operating cost and the main margin variable |
| Revenue mix by model | Segment disclosure | Managed-services share rising = blended RoE improving |
How the thesis breaks.
- !Speculative capacity. Building MW without contracted demand is the sector's principal way of destroying capital. Rental correction is already flagged as a live risk.
- !Oversupply and rental pressure. A 1,263 MW base heading to 3,000 MW is a lot of supply arriving at once. Rents are the first thing to give.
- !Power constraint and cost. Grid access can cap expansion regardless of demand, and power inflation compresses margin where it is not a pass-through.
- !Hyperscaler bargaining power. Half the demand pool can dictate terms, and can also choose to build its own facilities.
- !Technology obsolescence. Cloud-build and IaaS models carry real risk of stranded hardware as compute architectures shift.
- !Capex overrun and funding. At ₹60–70 crore per MW, a modest overrun or a delayed lease-up materially changes the return.
- !Proxy exposure is not the same thing. Equipment and construction suppliers benefit from the build cycle but carry order-book, not annuity, economics — do not value them as infrastructure.
The figures, and where they stand.
| Metric | Value | Note | Basis |
|---|---|---|---|
| India capacity | 1,263 MW | Installed base | 2025 |
| Capacity target | 3,000 MW / 4,500 MW | 2028 and 2030 | Est. |
| Cost per MW | $6–7mn (₹60–70 cr) | 30–40% below US, ~18% below APAC | Research note |
| Cost split | 55–60% / 5–7% / 15–20% | Land + M&E / statutory / operational | Research note |
| Rent per kW/month | $70–90 / $90–110 | Bulk small space / large space | Research note |
| Revenue per MW (colocation) | ₹80–100mn p.a. | Maximum annual | Research note |
| Colocation EBITDA margin | 60–70% | Excluding electricity pass-through | Research note |
| RoE by model | 10–15% / 10–20% / 25–30% | Colocation / cloud build / managed services | Research note |
| Cloud-build capex per MW | ₹800mn–1bn | Very high capital intensity | Research note |
| Payback with anchor client | 2–3 years | At 12–18% ROI; minimum 10-year leases | Est. |
| End-user mix | ~50% / ~22% | Hyperscalers / BFSI, IT, healthcare | Research note |
| PUE benchmark | 1.67–1.8 vs ~1.1 | Enterprise vs best hyperscale (1.0 = perfect) | Research note |
| Hyperscale threshold | 5,000 servers / 10,000 sq ft | Definitional | Definition |