Most of the AI-and-India narrative for the last three years has been written about applications — language models, agents, consumer surfaces. Underneath all of it is a much quieter story about kilowatts. India's installed data-centre capacity went from 158 MW in 2014 to 942 MW by the first half of 2024. That is a 6× run, and it makes India the second-fastest-growing DC market in APAC at 28% growth. The build is real, and it changes what AI-native companies in India can credibly attempt.
India installed data-centre capacity (MW)
Composite from industry reports and operator disclosures. Figures rounded.
The trajectory has accelerated in the last four years, driven by hyperscaler buildouts and the entry of Indian developers like Yotta, CtrlS, and AdaniConneX. The slope from 2020 to 2024 is sharper than anything the previous decade saw.
Capacity build by category (estimated, MW)
- Hyperscaler-leased
- Enterprise + BFSI
- Other / colocation
Hyperscaler-leased capacity is the fastest-growing slice; managed-services DC capacity is the steadiest. Independent estimates vary; this is our composite.
Why the build matters for AI-native builders
Three first-order implications. First, latency. Inference compute now exists physically inside the demand market. An AI-native consumer surface built for an Indian audience can roundtrip a tokenised response in single-digit milliseconds rather than the 80–150 ms it took to hit a Singapore region. For voice and agent products, this is the difference between feeling local and feeling laggy.
Second, data residency. RBI, SEBI, IRDAI, and a half-dozen other regulators have made on-shore storage a condition of doing business. Any AI workload touching financial, health, or telecom data now needs to be served from Indian capacity. The build means it can be.
Third — and this is the under-discussed one — cost. India is the cheapest large market in APAC to build a data centre in. The build-cost-per-MW is roughly a third of Japan's and 40% of Singapore's.
Capex to build 1 MW of data-centre capacity ($M)
Singapore
$11.2 · $11.2
Japan
$12.7 · $12.7
Australia
$9.3 · $9.3
Indonesia
$7.6 · $7.6
India
≈ 36% of Japan, 41% of Singapore
$4.6 · $4.6
$M per MW
Composite estimates across operators, 2024. Includes shell, power, cooling, security; excludes land variability.
Cheaper capex compounds into cheaper inference for end users. The operator who builds an Indian-served AI consumer product in 2026 has a structural cost advantage over a peer running on Singapore-based capacity, before any product or distribution effect.
Power, the actual constraint
The 942 MW is installed IT capacity. The total electrical load at the wall, including cooling, is roughly 1.5×. That means India's DC industry currently draws ~1.4 GW. By 2030, sector forecasts call for 5+ GW. India's grid can absorb this — barely, and only if renewable additions keep their current pace. The next bottleneck for the DC build is not concrete or chips; it is megawatts at the substation.
“The interesting questions about AI in India over the next five years are mostly energy questions. Whoever signs the right PPAs in 2026 owns the inference economy of 2030.”
What this means for our deal flow
We are seeing — and increasingly funding — three kinds of company catalysed by this build:
- AI-native consumer surfaces that simply could not run on the previous-generation cost-and-latency stack. Voice agents in Indic languages are the cleanest example.
- Vertical compute orchestration — companies whose product is moving model workloads between Indian regions for cost, latency, or compliance reasons.
- DC-adjacent infrastructure — cooling efficiency, on-site renewables, demand-shaping. Not all venture-scale, but the few that are will be very large.
The shorthand we use: if a company you are looking at structurally depends on Indian compute being abundant and cheap, the timeline for that being true is now, not in three years. The MW have shown up.