AI in India is, for now, an electricity and construction business more than a software one. The layers below run from the land and the powered shell up through generation, grid connection, transformers, cabling, cooling and fibre before reaching silicon, servers, cloud and finally applications — and the weight of listed value sits far lower down that stack than the word 'AI' suggests. Two things are worth carrying as you read. The honest gap is layer 9: no Indian company designs or fabricates AI-class silicon, and the largest node there is unlisted. And the strongest genuine position is layer 5, where Indian transformer and switchgear makers compete on cost and capacity rather than proximity. Tap any layer to open it, then any company for detail.
At the head of this chain
Highest return on capital and fastest sustained sales growth among the 87 listed companies in this chain above ₹1,000 Cr market cap and ₹500 Cr revenue. Sustained growth is the lower of the 3-year and 5-year sales CAGR, so a single strong year cannot win it — Netweb Technologies compounded 69.9% over three years and 72.5% over five. Same rule on every map. Snapshot 8 Sep 2026.
How India's AI build-out works
Twelve layers from land and power up to applications. Mostly an electricity business.
- Margin pool
- Highest in electrical equipment and cooling, where global shortage meets cheap Indian capacity — a 500 MVA transformer costs about $2.5 mn to build here against $14–22 mn in the US. Thinnest at semiconductors and servers, where India assembles and tests but does not design, so value added is a small share of system cost.
- Bargaining power
- With whoever is scarce. Today that is transformer, switchgear and high-density cooling suppliers, plus NVIDIA at the silicon layer. Data centre developers have less power than the headlines suggest — they are landlords leasing to a handful of very large tenants.
- Demand or supply led
- Supply-constrained. Demand for compute exceeds what can be powered, cooled and connected. The binding constraint is grid connection and equipment lead time, not customer appetite.
- Who owns the customer
- The hyperscaler — Amazon, Microsoft, Google — sits between Indian infrastructure and the end user. Indian companies lease them capacity or sell them equipment; almost none has a direct relationship with whoever ultimately uses the AI.
- Barriers to entry
- Lowest in construction and fit-out. Highest at semiconductor fabrication, which India does not do at all, and at land with grid access, which cannot be manufactured.
- Threat of substitutes
- At the compute layer, efficiency is the substitute: a better model needs fewer GPUs. Further down, none — nothing substitutes for electricity, copper or cooling water, which is why those layers are the more defensible place to sit.
This is a generalist read of how the industry typically works, not a rule. Anomalies exist at every stage — a well-run company in a poor part of the chain routinely beats a badly-run one in a good part, and structure changes with the cycle. Use it as a starting frame, not a conclusion.
No company matches that search.