Two independent Dart Consultants research efforts, covering ten distinct sub-industries between them, each converged on the same finding without being designed to. Here it is, stated once, for the whole stack.
Bolting a GPU board into an Nvidia-designed chassis; pouring a warehouse-grade concrete shell to house a data hall — these are, at their core, the same kind of problem: capital-and-process scaling. Buy the equipment, hire and train the workforce, pass the specified test. Any competent operator — an electronics-manufacturing contractor, a construction firm — can in principle attempt either, using skills India built up over decades in unrelated industries, not invented for this report. Failure here is visible and immediate: a board either boots and passes burn-in, or it does not; a wall either meets code, or it does not.
The same rack, run at sustained full load, and the same building, once it holds fifty-plus megawatts of continuous critical power — both become problems with a layer that is genuinely scarce, not just difficult. On the compute side, that scarcity is chip supply itself, rationed by Nvidia and ultimately by TSMC's and the memory makers' packaging capacity. On the physical side, it is a slower-moving but no less real capital-equipment manufacturing constraint — large power transformers and gas-insulated switchgear facing multi-year lead times in several global markets. Both failure modes share the same signature: invisible and delayed. A marginal power-delivery or cooling design does not fail on a bench test; it fails as GPU throttling or a transformer hot-spot months into sustained real-world load, exactly when it is most expensive to discover.
| The easy half | The hard half | |
|---|---|---|
| Compute layer | Assembling a server/rack to Nvidia's own reference design | The chip/packaging/memory layer (outside India) plus the power-and-thermal engineering that keeps a dense rack alive |
| Physical layer | Civil/EPC construction of the building shell; catalog cabling and generic equipment | Transformers, switchgear and precision-engineered cooling gated by global manufacturing capacity |
| How it fails | Visible immediately | Invisible and delayed, surfacing only under sustained real-world load |
| Where the margin sits, per this report's own company research | Thin and compressing — EMS/server assembly (§3, sister thesis) and civil/EPC construction (KEC 7.0%→5.8%, Ahluwalia ~10%→4.29%) alike | Thick and often expanding — Nvidia/Micron/SK Hynix margins, and Hitachi Energy India's 7%→9.3%→15.4% operating-margin climb |
The shell is the easy part, whether that shell is a server chassis or a building. The chip allocation and the physical engineering that keep whatever is inside the shell actually working are the product. Every company report in Part 4 is really answering one question: how much of this company's revenue sits on the scarce side of this line, at whichever layer of the stack it occupies.
