India committed ₹70,000 crore to a single one-gigawatt AI campus in Hyderabad on 5 September 2026. Yet roughly five per cent of the country’s installed India AI data center capacity can actually run AI workloads today. Both things are true, and the distance between them is where the money will be made or lost.
Table of Contents
- The India AI Data Center Numbers Nobody Agrees On
- Where the Money Actually Is
- Announced Is Not Energised
- The Map Is Being Redrawn
- Three Constraints Capital Cannot Buy Past
- The Missing Layer
- Who This Is For
- Where Hybr® Fits
- Frequently Asked Questions
- References
The India AI Data Center Numbers Nobody Agrees On
India’s installed data centre capacity is somewhere between 1.28 and 2.2 gigawatts. That is not a rounding error — it is the actual spread across six credible sources published within twelve months of each other, and a single confident number for the India AI data center market reflects one house’s definition rather than a settled figure.
The government’s own Budget 2026-27 documentation puts “cloud data centre capacity” at about 1,280 MW. CBRE counted roughly 1,530 MW of operational IT load in September 2025. JLL says 1.6 GW at mid-2026. KPMG says 1,900 MW for FY26. Savills says 1.8 GW at H1 2026. Wood Mackenzie says 2.2 GW.
| Source | Installed capacity | As of |
|---|---|---|
| Government of India (Budget 2026-27) | ~1,280 MW | Feb 2026 |
| CBRE | ~1,530 MW (IT load) | Sep 2025 |
| JLL | 1.6 GW | Mid-2026 |
| Savills India | 1.8 GW | H1 2026 |
| KPMG India | 1,900 MW | FY26 |
| Wood Mackenzie | 2.2 GW | 2025 |
None publish a comparable methodology. The differences look definitional: IT load versus total facility load, third-party colocation only versus including enterprise captive and hyperscaler self-build, and the narrower “cloud” subset the government appears to be counting. It matters commercially, because the 2030 forecasts inherit the base. Cushman & Wakefield lands near 4.2 GW, CEEW at 4.5–6.5 GW, JLL at 6 GW by 2029, Rubix at above 6.5 GW, KPMG at 7–7.5 GW, IEEFA at 9 GW and Wood Mackenzie at 12 GW. Compare Wood Mackenzie’s 12 GW against CBRE’s base and you will overstate the growth rate by a wide margin.
The honest version, and the one worth planning against: India roughly triples to quadruples its data centre capacity by 2030, from a base of somewhere under 2 GW. India currently holds around 2–3 per cent of global installed capacity while generating something close to 20 per cent of the world’s data. That imbalance is what the capital is chasing.
Where the Money Actually Is
Four different “India data centre investment” numbers circulate, and they are routinely quoted side by side as if comparable. They are not — one is announced commitments, one is cumulative since 2019, one is capital required, and one is a ministerial aspiration.
| Figure | What it actually measures | Source |
|---|---|---|
| $70bn underway + $90bn announced | Government framing of the project pipeline | PIB, Budget 2026-27 |
| $94bn | Cumulative commitments since 2019 | CBRE, Sep 2025 |
| >$120bn | Cumulative commitments to Mar 2026 | KPMG, 2026 |
| $110bn | Capital required to reach 6 GW by 2029 — of which $88bn is IT equipment | JLL, Sep 2026 |
| $200bn | Stated national ambition, no published basis | Ministerial statement, Feb 2026 |
The named commitments are firmer ground. Amazon has committed $48 billion to India across 2026–2030, of which $21 billion is cloud and AI infrastructure, and broke ground in Hyderabad in July 2026. Microsoft committed $17.5 billion across CY2026–2029 with a new India South Central region. Google, AdaniConneX and Airtel Nxtra broke ground in April 2026 on a $15 billion, roughly one-gigawatt campus at Visakhapatnam. Reliance Intelligence has put ₹10 trillion behind a seven-year AI programme with 120 MW targeted at Jamnagar by the end of 2026. And on 5 September 2026, TCS subsidiary HyperVault announced ₹70,000 crore for a one-gigawatt liquid-cooled campus in Hyderabad.
Note what JLL’s $110 billion number is telling you: eighty per cent of the capital required is IT equipment, not buildings. This is a silicon story wearing a real-estate costume.
Announced Is Not Energised
Around 3.5 GW of Indian data centre capacity was announced between March 2025 and April 2026 across roughly thirty projects. CRN Asia’s April 2026 analysis put less than twenty per cent of that announced capacity live, with phased commissioning running to 2028 and beyond. Meanwhile, of the installed base that does exist, KPMG puts the AI-capable share at about five per cent.
These are two different gaps and they are constantly collapsed into one headline. A press release announcing 200 MW is a statement of intent about 2028. The capacity that can take a liquid-cooled 130 kW rack today is a much smaller number, and the capacity that is both AI-ready and unsold is smaller still — JLL put colocation vacancy at a record-low 2.8 per cent in H1 2026, with 82 per cent of absorption pre-committed to hyperscalers.
This is not a contrarian reading of the market. CRN Asia reached the same conclusion in April 2026, and KPMG’s July 2026 India data centre report frames execution — not demand — as the binding question on the buildout. The operators say it themselves: asked where the margin now sits, three of India’s largest told CRN that the scope for differentiation in simply provisioning space and power is narrowing, and that the higher value is emerging above the infrastructure layer.
KPMG projects the AI share of Indian capacity reaches around 55 per cent by FY30 and 65 per cent by FY35. That transition — from a fleet built for 8 kW racks to one built for 30–80 kW racks, with specialised deployments at 120–150 kW — is not an upgrade path. It is largely new build.
The Map Is Being Redrawn
Mumbai holds 49–53 per cent of India’s installed data centre capacity but only 35 per cent of the forward pipeline. Hyderabad at 24 per cent and Visakhapatnam at 14 per cent now exceed it between them. That is the most defensible “what is changing” datapoint available, and it comes from JLL’s September 2026 pipeline analysis.
| City | Share of installed base | Share of forward pipeline |
|---|---|---|
| Mumbai | 49–53% | 35% |
| Hyderabad | within the top five by facility count | 24% |
| Visakhapatnam | negligible | 14% |
| Chennai | ~20% | 9% |
The drivers are specific rather than general. Visakhapatnam has coastal subsea landings and active Andhra Pradesh state support. Hyderabad is absorbing gigawatt-scale hyperscaler self-build. Chennai has eight subsea cable landings, though the Tamil Nadu Data Centre Policy 2021 was written on a five-year term — worth confirming the current incentive position before underwriting a Tamil Nadu site.
State policy is now the live frontier. Gujarat notified its Viksit Gujarat Data Centre Policy 2026-29 on 7 August 2026, targeting 7.5 GW of green AI capacity and becoming the first state to mandate at least 51 per cent renewable sourcing, with a ₹1 per unit tariff subsidy for twenty years and support for captive desalination. Uttar Pradesh approved a policy in July 2026 with GPU-data-centre-specific incentives and a two-gigawatt target.
Three Constraints Capital Cannot Buy Past
Power, water and operating capability are the three things that money alone does not resolve on an India AI data center project — and only one of them gets discussed honestly.
Power, where the national number is comfortable and the local one is not
A written reply to Parliament on 29 July 2026 put the additional power demand from AI and data centres at 26.3 GW by FY32. Against installed generation capacity of 520.5 GW, that does not break the grid in aggregate. Locally is another matter: the CEA has told states that data centres need 132 kV-and-above transmission connections and has warned Andhra Pradesh, Telangana, Tamil Nadu and Maharashtra about voltage fluctuation and tripping risk. The same constraint is playing out in the United States, where interconnection queues have become the binding limit — we covered that in the AI datacenter power crisis.
Water, where the honest framing is local, not national
CEEW puts Indian data centre water consumption at roughly 150 billion litres a year, projected to more than double by 2030. That sounds alarming until you see the same study’s other number: data centres account for 0.02–0.03 per cent of national water demand. The problem is not national scarcity. It is that a 100 MW facility on evaporative cooling draws around two million litres a day, and more than half of India’s data centres sit in water-stressed catchments. Digital Edge’s Navi Mumbai campus is the reference case for the alternative — ten million litres a day of recycled water with a WUE target below 1.75.
Operating capability, which nobody puts in a press release
The third constraint has no capex line. Running a multi-tenant AI service is not the same job as running a colocation hall. It means allocating shared GPU capacity between customers without one customer’s job starving another’s, proving to an auditor which tenant’s data sat in which region, and producing an invoice that a finance team will pay without a dispute. It is a genuinely new job, and almost no operator in any market has had to do it at this scale yet. India is about to have a great deal of capacity that needs it done well from the first tenant onward.
The Missing Layer
The day after the racks energise, everything that decides margin on an India AI data center sits in software: a service catalogue, tenant isolation, metering, rating, invoicing and a self-service portal. Three of the four things below the line are capital expenditure. The fourth is on nobody’s capex line at all.
Concretely, here is what an Indian operator needs that a rack and a GPU do not provide:
- A catalogue with offers and SKUs. Not a rate sheet emailed as a PDF — a catalogue a customer can order from, with entitlements attached, so that ordering and provisioning are the same action.
- Tenant isolation that survives an audit. Role and scope per tenant, region pinning where the workload is regulated, and an access trail per tenant rather than per cluster. This is a harder requirement in India than most markets, and it is the subject of the seven rules that actually bind on AI data residency in India.
- Metering that reaches the units you sell. GPU-hours, tokens, storage and egress, attributed to a tenant, continuously — not reconstructed from logs at month end. See usage-based and metered billing.
- Rating in rupees, with channel margin. A price list per customer and per reseller tier, in INR, because the channel is how Indian capacity reaches mid-market buyers. See pricing and rate management.
- Invoicing that is correct the first time. Eighteen per cent GST on cloud services under SAC 998315, e-invoicing above ₹5 crore of aggregate turnover, and a 30-day IRP reporting window above ₹10 crore. An invoice that fails these is not a billing inconvenience; it is a compliance event.
- A portal the customer actually uses. Order, change, watch spend accrue, download the invoice — without a ticket. See the customer self-service portal.
Without that layer, an operator sells raw hours at whatever price the market sets, and the market for raw hours is already crowded and already discounting. With it, the same rack becomes a catalogue of priced services sold through a channel. The arithmetic of that difference — what Indian GPU capacity actually rents for, and why utilisation rather than price decides the outcome — is the subject of GPU-as-a-Service in India.
Watch the two-minute version. We walk through what sits between a rack and a customer’s invoice — catalogue, metering, rating, invoicing — using the actual product rather than a diagram. See how the service layer works →
Who This Is For
Three groups face this India AI data center gap from different sides, and all three need the same layer.
- Colocation and hosting operators moving up-stack. You have the power, the space and increasingly the GPUs. Selling capacity as a metered, multi-tenant service rather than as space and power is the entire margin difference, and it is a software problem.
- Indian MSPs and Microsoft CSPs adding AI to an existing book. You already bill seats and subscriptions. Adding GPU-hours and tokens to the same invoice, under the same reseller margin structure, is how AI becomes a line of business rather than a project. See Microsoft CSP billing.
- Enterprises and GCCs running internal AI platforms. India hosts 2,117 global capability centres, of which more than 1,200 have embedded AI capability. Charging shared GPU capacity back to business units is the same metering-and-rating problem with an internal customer.
Where Hybr® Fits
Hybr® is the service layer above the infrastructure — the catalogue, tenancy, metering, rating, invoicing and portal that turn capacity into a billable product. It registers the Kubernetes and GPU clusters you already run rather than provisioning them, connects to the LLM gateway you already operate rather than hosting models, and puts GPU-hours, tokens, Microsoft CSP seats and managed services on one invoice per customer, in their currency, under their price list.
If you are building or operating AI capacity in India, the useful next step is narrow: take one service you intend to sell, and model the offer, the meter, the rate card and the invoice it produces. That conversation takes an hour and tells you more than a capacity forecast will. Start at AI Factory, or see how the metering works in AI application and token billing and Kubernetes and GPU-as-a-Service.
Frequently Asked Questions
How big is the India AI data center market in 2026?
India’s installed data centre capacity sits between roughly 1.28 GW and 2.2 GW depending on the source and what each counts, with the government’s own Budget 2026-27 figure at about 1,280 MW of cloud data centre capacity. Forecasts for 2030 range from about 4.2 GW to 12 GW. Around $90 billion of investment is announced and $70 billion underway on the government’s framing.
How much of India’s data centre capacity can actually run AI workloads?
KPMG puts the AI-capable share of India’s FY26 installed base at roughly 0.1 GW out of about 1.9 GW — close to five per cent — and projects that share reaching around 55 per cent by FY30. Separately, less than twenty per cent of the roughly 3.5 GW announced between March 2025 and April 2026 is operational today.
Why does an India AI data center need a billing and metering layer?
Because capacity only earns margin once it is sold as a service. That requires a catalogue a customer can order from, isolation between tenants sharing GPUs, metering of GPU-hours and tokens per tenant, rating against that customer’s price list in rupees, and a GST-compliant invoice. None of that comes with the rack, and none of it appears on a construction capex line.
Which Indian cities are attracting AI data centre investment?
Mumbai still holds roughly half of installed capacity but only 35 per cent of the forward pipeline. Hyderabad at 24 per cent and Visakhapatnam at 14 per cent together now exceed it, driven by hyperscaler self-build and coastal subsea landings respectively. Chennai holds about 20 per cent of installed capacity, though its 2021 state data centre policy was written on a five-year term and the current incentive position is worth confirming.
Is power a real constraint on India AI data center growth?
Locally, yes; nationally, not yet. The government puts the additional demand from AI and data centres at 26.3 GW by FY32, against installed generation capacity of 520.5 GW. The Central Electricity Authority has warned specific states about grid stability and requires 132 kV-and-above connections for data centres.
What does an operator need before selling AI capacity to multiple tenants?
Six things: a service catalogue with offers and SKUs, per-tenant isolation and role-based scoping, continuous metering of the units being sold, rating against per-customer and per-reseller price lists, invoicing that meets Indian GST and e-invoicing rules, and a self-service portal so routine changes do not become support tickets.
References
- Press Information Bureau — Budget 2026-27 data centre and cloud measures
- Business Standard — India’s installed data centre capacity to rise four-fold by 2030 (Rubix Data Sciences)
- KPMG India — The India data centre opportunity (July 2026)
- JLL — India data centres market dynamics
- CEEW — How is data centre infrastructure in India shaping power and water use
- Amazon — India investment announcement, June 2026
- Microsoft — $17.5 billion India investment
- Adani — Adani and Google partner on Visakhapatnam campus
- ThePrint — TCS subsidiary HyperVault to invest ₹70,000 crore in Hyderabad
- Business Standard — AI data centres to add 26.3 GW to India power demand by FY32
- ANI — Gujarat notifies Data Centre Policy 2026-29
- nasscom–Zinnov — India GCC Landscape 2026
- Digital Edge — Navi Mumbai solar PPA and recycled-water data centre
- CRN Asia — India’s AI data centre boom is real, but execution, not announcements, will decide outcomes (April 2026)
- CRN Asia — India’s AI data centre reset is reshaping infrastructure economics and partner roles (April 2026)
- Government of Tamil Nadu — Tamil Nadu Data Centre Policy 2021