AI Data Centres: The Power, Water and Capex Economics Behind the Boom
Current position
Indian data-centre projects operate within electricity, land, building, fire, environmental, telecom, data and state-policy frameworks. Approvals and incentives vary by location. Claims about “capacity” should specify whether they refer to sanctioned power, IT load, installed servers, operational capacity or contracted utilisation.
How it works
A project needs firm and redundant power, substations, backup systems, fibre routes and cooling. Renewable certificates or contracts do not automatically mean the facility runs on carbon-free power every hour.
AI accelerators can cost more and produce denser heat loads than traditional servers. Depreciation life, replacement cycles and vendor concentration should be stress-tested.
Water use depends on cooling design, climate and operating mode. Report direct site water separately from water embedded in electricity generation where possible.
| Issue | Current position | Why it matters |
|---|---|---|
| Core denominator | IT load versus total facility load | Do not mix MW measures |
| Efficiency metric | Power Usage Effectiveness | Lower can be better, but climate and load matter |
| Economic risk | High upfront capex and rapid hardware obsolescence | Utilisation drives return |
Practical example
A data-centre company announces 100 MW of planned capacity. Only 25 MW is operational, 20 MW is contracted and average IT utilisation is 60%. Revenue cannot be estimated by multiplying 100 MW by a market tariff. Analysts must distinguish planned, powered, commissioned, occupied and billed capacity.
Action checklist
- Define capacity consistently: planned, sanctioned, commissioned and utilised.
- Model power price, PUE, cooling water and backup fuel.
- Stress-test hardware refresh, customer concentration and lease tenor.
- Match debt repayment to contracted cash flow rather than planned capacity.
- Disclose renewable sourcing and emissions methodology precisely.
Evidence and document checklist
- Land and power approvals.
- Grid connection, redundancy and outage records.
- Customer contracts and utilisation reports.
- Equipment capex, depreciation and maintenance schedule.
- PUE, water and emissions measurement methodology.
Common mistakes
- Calling sanctioned power operating revenue capacity.
- Using GPU count without utilisation and pricing.
- Ignoring replacement capex.
- Describing annual renewable matching as 24/7 clean power.
Red flags
- Single customer supports most debt service.
- Power connection date is uncertain.
- Cooling design lacks water-risk assessment.
- Capex per MW excludes network or substation costs.
Escalation and complaint route
Project approvals belong with state, electricity, environmental, fire and local authorities. Data-security incidents may trigger CERT-In and contractual duties. Investors and lenders should obtain engineering, legal and commercial diligence.