PDF ↗
Home/Primer/The parts the data misses
Section 8

The parts the data misses

Three pools of value are structurally undercounted, or structurally mis-modelled, by the headline statistics used to size this opportunity. This is where the report's own original synthesis, rather than a repackaging of public data, actually lives.

1. Chip design work booked as generic "services exports," not as "AI hardware"

India's global capability centres (GCCs) and engineering-R&D (ER&D) firms do real chip-level design, verification and embedded-software work for global fabless semiconductor clients today — Nasscom projects India's ER&D services revenue will cross US$100bn by 2030. None of that work shows up in any "India AI hardware market" or "India semiconductor market" statistic, because India's trade and industry data book it as a generic IT/BPM services export, indistinguishable in the headline numbers from an unrelated back-office contract. Two of this report's eight companies (L&T Technology Services, Cyient) earn real revenue from exactly this kind of work — see their individual reports for how much of it is disclosed at the semiconductor/AI-specific level (in both cases, less than an outside analyst would like).

2. Pre-committed capacity that is not yet "operational" in anyone's MW count

JLL's own India data-centre research found that pre-committed hyperscale capacity made up 82% of total data-centre space absorbed in the first half of 2026 — meaning the great majority of new capacity changing hands today is locked in by land and power agreements well before it is commissioned and counted as "operational" in the capacity statistics §6 and §7 quote. Every published "India data-centre capacity, MW" figure is, by construction, a lagging count of what has already been switched on — it structurally understates the demand for the chips, servers and packaging this report's companies actually sell, because that demand is contracted years ahead of the MW figure catching up.

3. A GPU import ceiling that market-size forecasts do not model as a hard constraint

Reporting aggregated in this research (not independently confirmed against a primary US Commerce Department or Indian customs document, and flagged here as such) suggests India may be permitted to import on the order of 50,000 H100-class-equivalent GPUs through 2027 under US export-control tiering. If a figure anywhere near this is accurate, it functions as a hard rationing ceiling on the entire downstream opportunity — the servers, racks, cooling systems and packaging this report's companies sell are all, ultimately, gated by how many chips are allowed into the country, independent of how large any "India AI market, $bn" forecast says the addressable opportunity is. This is exactly the kind of supply-side constraint a demand-side market-sizing exercise structurally cannot see.

Why this section matters for §9

Every company report that follows should be read against this backdrop: a company's current, disclosed revenue from "AI" or "semiconductor" work is very likely a lagging and understated signal of its real position — both because genuine chip-design contribution is often buried inside a larger, undifferentiated services line, and because the entire industry sits downstream of a chip-supply ceiling no company in this report controls. Conversely, that same ceiling means near-term revenue growth for several of these companies is capped by an input none of them can source more of by trying harder.

Educational material only — not investment advice. Dart Consultants is not a SEBI-registered Investment Adviser or Research Analyst.