| Origin | What it is | Who captures it | Time horizon |
|---|---|---|---|
| 1. Reference-design assembly | Bolting together a server/rack to Nvidia's own MGX/HGX specification | Whoever has EMS capacity and NVQual certification — increasingly contestable and margin-compressing, not less | Now, but the industry's own data (§3) shows this margin thinning, not thickening, generation over generation |
| 2. Government-incentivised capability building | ISM 2.0 (OSAT/fab), IT Hardware PLI 2.0 (assembly), IndiaAI Mission (demand) | Whoever wins scheme approval and executes on schedule — CG Semi and Kaynes Semicon are both already in commercial production | Now through the early 2030s, on a government-set, not market-set, timetable |
| 3. Chip/IP design-services exports | ER&D and GCC work for global fabless clients, largely invisible in hardware trade statistics (§8) | Whoever has the engineering talent base and client trust — LTTS, Cyient and Sasken all participate, none discloses how much | Ongoing, but structurally undercounted, so its growth is harder to track than the AI-server headline numbers |
| 4. GPU-scarcity rationing | A hard ceiling on how much of the downstream opportunity can actually be built, regardless of demand | Nobody in this report's universe — the constraint sits entirely upstream, at Nvidia, TSMC and the memory makers | At least through 2027 on current HBM/CoWoS supply commitments |
| 5. GPU-cloud consumption & sovereign-AI demand | Renting out compute on already-built GPU infrastructure, and government-backed subsidised access programmes | E2E Networks and ESDS Software Solution most directly; RailTel and Anant Raj as the facilities/connectivity layer beneath it | Now, but priced at multiples (157-401x trailing) this report's own valuation arithmetic cannot support on either name's currently disclosed earnings |
Global hyperscaler capex is not slowing — four companies alone are guiding to ~$725bn in 2026, up 77% YoY, and every one of them has said the binding constraint is chip supply, not appetite. India's own semiconductor-policy build-out is visibly executing, not just announcing: three OSAT/ATMP facilities are already in commercial production as of this report's research date. The IndiaAI Mission is a genuine, policy-guaranteed domestic demand stream, already generating confirmed nine-figure orders for at least one company in this report. Several of these companies' credit ratings are improving, not deteriorating, reflecting real balance-sheet strength behind the AI narrative.
The industry's own, real (not indicative) margin data shows the assembly layer — where most of this report's original eight companies sit — compressing under Nvidia's own component-integration strategy, a structural headwind none of them can engineer their way around. Several richly-valued names (Netweb, Kaynes, Syrma) trade at multiples this report's own valuation arithmetic cannot support even on optimistic growth assumptions, and in Kaynes's case a real governance deterioration has already shown up in the numbers (a guidance miss, negative operating cash flow). The seven newly-added companies do not fix this picture — they add a second, equally serious risk axis: E2E Networks and ESDS trade at multiples (401x, 157-302x) built on earnings this report's own research found reasons to distrust, and three of the seven (Anant Raj, Adani Enterprises, and to a lesser extent the whole third bucket in §9) carry governance or legal overhangs layered on top of, not instead of, ordinary valuation risk. The chip-supply ceiling (§8) caps the addressable opportunity for every company in this report, regardless of how large any market-size forecast says the demand is. And a single further Nvidia move to internalise more of the reference-design value chain could compress the assembly-layer margin further before any of these companies have proven they can move up the value ladder.
