If qualification is this demanding, why does a conventional EMS operation get any AI-server work at all? Because in a reference-design build, Nvidia has already done the hard engineering for you. It has picked the chassis geometry, defined the cooling-loop topology, specified the power-delivery components and carries the system-level qualification cost through its own NVQual programme. The manufacturer's job shrinks to precision execution — the "free input" here is Nvidia's own systems- engineering, handed over as a specification. This is why a commodity-grade EMS capability still finds a real market in AI hardware: someone else has already solved the hard half of this specific design, and is renting out the easy half.
The alternative — custom engineering — inverts this. A hyperscaler with its own hardware team designs its own rack, cooling loop and interconnect topology from scratch, owns the resulting intellectual property, and captures the performance edge that design delivers over an off-the-shelf reference build. It is a smaller, harder door, and — reading across this report's eight companies in §9 — essentially none of India's listed AI-compute-stack universe has walked all the way through it yet: every one of them sits, to varying degrees, on the reference-design side of this line. That is not a criticism of any individual company; it is a structural fact about where India's listed compute-stack industry currently stands, and it is the single most important qualifier to carry into every company report that follows.
When a company report later in this document describes a company as an "NVIDIA-certified AI server manufacturer," read that as: mostly reference-design building, margin will be thin and structurally under pressure from Nvidia's own component-integration strategy. When it says a company holds a scarce packaging capability or a chip-design-services practice with real client IP embedded in it, read that as closer to the hard half, and worth asking whether the valuation already assumes that scarcity persists.
