This report merges two Dart Consultants primers researched independently — "India's AI Compute Stack" (the chip-to-server-to-GPU-cloud half) and "India's Data-Center Physical Infrastructure" (the power-to- cooling-to-construction half) — into a single, continuous value chain, because that is what an AI data centre actually is: one physical object, not two separate industries. A GPU cannot compute without a transformer stepping grid power down to usable voltage; a transformer earns nothing if there is no rack of GPUs on the other end of it. Treating "AI stocks" and "data-centre infrastructure stocks" as different universes — as most published research does — is the single easiest way to miss how value actually moves through this stack, and where it gets stuck.
Every layer of this stack — chip design, packaging, server assembly, power delivery, cooling, cabling, the building itself — splits the same way: an easy half that any competent operator can execute to a published specification, and a hard half gated by either a scarce allocation (chip supply, global transformer manufacturing capacity) or genuine, hard-to-replicate engineering. This report tested that split independently on two different research efforts covering ten distinct sub-industries and found it every single time. That is not a coincidence of methodology — it is close to the organizing law of how a heavy-capex, physically-constrained industry actually distributes its margin, and it is the one lens applied uniformly, layer by layer, through the rest of this document.
A note on sourcing: every figure in this report carries a source and a date next to it. Where published estimates disagree — market size, power demand, even a single company's own quarterly profit in more than one case — we show the disagreement rather than picking a number silently. Where we could not verify something, we say so. See the closing Notes section for the full methodology and data caveats, including where the two original research efforts' dates and conventions differ.
