AI: The India Stack.
The only Indian sector where funding quadrupled into 2026 ($676M in H1, +317%): a sovereign-model programme (Sarvam at $1.5bn), a $1.2bn Blackstone-led compute bet (Neysa), $67.5bn of hyperscaler pledges, and a first unicorn (Krutrim) already in retrenchment.
The scorecard
India's AI stack is three different investments wearing one label. The application layer (~75-80% of startups) is capital-light and occasionally explosive: Emergent hit $50M ARR on $99M raised and quintupled its valuation in three months. The model layer is state-anchored: Sarvam trained a 105B-parameter sovereign model on government-provisioned H100s, and its $234M Series B (HCLTech strategic) is the only large private model-layer round. The compute layer is where institutional money actually went: Neysa's $1.2bn. The honest constraint: OpenAI raised 166x India's entire H1 AI funding in one round, and IndiaAI has disbursed just ₹400 cr of its ₹10,372 cr outlay in two years. India's edge is deployment at population scale, 22-language coverage and inference pricing an order of magnitude below global APIs, not the frontier.
Market at 25-35% CAGR toward $17bn by 2027 (Nasscom-BCG); ~80% of new GCCs carry AI mandates; Gnani alone processes 30M voice interactions/day.
OpenAI raised 166x India's H1 2026 AI funding in a single round; Sarvam's benchmark claims lack independent replication; Llama is free and Reliance-distributed.
Apps are capital-light (Emergent $50M ARR on $99M); models state-subsidised; compute devours capital (Neysa $1.2bn for 20k GPUs).
Krutrim, the first GenAI unicorn, cut 550→150 staff and pivoted to cloud; sovereign models have no proven demand beyond procurement.
IndiaAI GPUs at ₹65-92/hr (~one-third global cloud) is real; ₹400 cr/2yr disbursement vs outlay is the under-reported gap.
Application-layer entries at reasonable multiples + services cash flows vs $1.5bn pre-revenue model bets; know which side you're on.
Scores are 0-100 favourability. Competition at 35 reflects the structural frontier gap; the sector's investability lives in the layers that don't fight it. Priors reflect the January 2026 read.
The numbers
H1 2026 alone ($676M) ≈ one-third of ALL historical Indian AI funding; scope excludes the Neysa infra megadeal
38,000+ accelerators empanelled (incl. first TPUs); 100k target by Dec 2026: empanelled ≠ deployed
$17bn by 2027 on the Nasscom-BCG path (25-35% CAGR): definitional spread vs IDC's $6bn is wide
Demand · will supply get filled?
The demand question splits by layer: enterprises buy applications and services today; models sell to the state; compute sells to everyone building the other two.
Dollar-weighted, the 'AI boom' is mostly a compute build-out: Neysa's $1.2bn alone was ~81% of the broadest-scope Q1 2026 tally. By company count it's the opposite: ~75% of startups build applications. GCCs are the silent third force, holding ~27% of national AI talent inside foreign-owned centres.
Output, order book & the global gap
Funding charts can't answer the operating question: is anything actually shipping, being used, and contracted? Usage (API calls, GPU onboarding, dataset downloads) is the deployment proof; the sovereign order book is demand depth; the global table is the honest scale check.
The honest read: India ships adoption, not frontier: enterprise deployment and population-scale reach lead the world while capital, compute and research share trail by one to two orders of magnitude. The sovereign order book is real but mostly MoUs; the revenue line behind the $1.5bn model-layer valuation is ₹45 crore.
Competitive dashboard
By dollar the compute layer dominates; by count the application layer does. This chart tracks named private companies, ex-hyperscalers.
Pledges, not booked capex: $67.5bn announced by three hyperscalers in ten weeks. Delivery is the 2027 question.
Capital · unit economics, valuation & deals
Inference & compute economics: The cost wedge is real at both layers: subsidised compute at a third of global cloud, sovereign inference at a fraction of frontier APIs, but it prices smaller models against frontier capability. Services economics (Fractal at 17% margins, 109x P/E at IPO) carry model-layer multiples.
Follow the instrument, not the label: equity went to applications, debt+equity to compute, grants to models. The Reliance-Meta axis is the wildcard: free Llama distribution through India's largest conglomerate caps every domestic model's commercial ceiling.
The public market has already picked its lane: hardware/infra proxies re-rated hard, services flat, AI-adjacent software derated. The listed read matches our private-layer scorecard almost exactly.
Private players & platforms
Where most of the value is still private: startup-, PE- and strategic-backed. Scale, ownership, and the last marker of value.
FY26 revenue ₹45.1 cr: ~275x trailing; the sovereign-AI bet in one multiple
$600M equity + $600M debt: India's largest AI financing
Claims 60-70% of India's GPU capacity; IndiaAI's largest empanelled provider
550 → ~150 staff; LLM + chip paused; ~₹300 cr revenue with first profit (company-claimed)
Coding agents; the application layer's velocity print
Post-training data for DeepMind/Snowflake: India as the world's RLHF supplier
Startups & emerging players · the VC layer
Where venture capital enters the theme.
100k+ developers: picks-and-shovels for agents
IndiaAI voice-model awardee; 30M interactions/day
Three of the 12 sovereign-model grantees
The long tail of the sovereign programme
Sector-vertical sovereign model
The 12 funded foundation-model organisations are the sovereign programme's bench; the venture-backed layer (Composio, Deccan, Emergent) mostly sells globally from day one: the two ecosystems barely overlap.
Public-market exposure index · rules-based, purity-weighted
A screened, exposure-weighted basket: each listed name weighted by its sector-exposure purity score (not naively equal-weighted), after liquidity and quality screens. Selection is rule-driven and set ex-ante.
Real point-to-point anchors: each name rebased to 100 at −3y; the −1y (319) and now (542) levels from its actual 1Y & 3Y returns, purity-weighted. Intra-period linear (daily shape/drawdowns need a price feed).
Rules-based: include a listed name if its AI purity score ≥ 20/100 AND it clears the eligibility screens. Weight by purity (exposure-weighted), single-name cap 25%, overflow redistributed pro-rata. Quarterly reconstitution. Selection is rule-driven, set ex-ante, not a curation of past winners.
- ✓ Liquidity & size: investable free-float, adequate ADTV
- ✓ Quality: positive profitability (excludes loss-makers)
- ✓ Purity: AI revenue-exposure / relevance score ≥ 20 of 100
Rules-eligible, pending verified data: Oracle FS (purity < 20), Zaggle (purity < 20), L&T Technology Services, Cyient, Infosys (too diversified). Purity scores are a documented judgement tier (clean AI-revenue % is rarely disclosed; Netweb's 43.4% AI-systems revenue share is the exception that anchors the scale). Returns are representative point-to-point figures; 3-yr figures partly estimated (E).
Selection-bias caution: built from names already public over the window, which the AI theme has already re-rated (Netweb +139% in a year). Past returns are upward-biased and NOT a forward estimate; the rules, not the hindsight, are what's intended to repeat.
Research / informational only: not investment advice or a recommendation. Baskets are illustrative of the rules, not a managed product.
Externalities & policy footprint
The externalities and strategic stakes a government must price in.
Scenarios to 2030
Frontier price war erases the inference wedge; sovereign models retreat to procurement; services capture the value
Deployment-market thesis: applications + GCC + data services compound; compute build-out delivers partially
The $265bn DC pipeline + $67.5bn hyperscaler pledges convert to services revenue; EY's $359-438bn GDP-add overlay materialises
Three numbers puncture the narrative: Sarvam's ₹45.1 crore FY26 revenue against a $1.5bn valuation (~275x trailing); AIKosh's ~13,000 downloads against 5,500+ catalogued datasets; and IndiaAI's unpublished GPU-utilisation figures. The capacity story is real and accelerating: the demonstrated-usage story is one to two orders of magnitude behind it. Underwrite deployment, not announcements.
Financing · policy · catalysts
Sensitivities · what moves returns
Risks quantified, not just listed: the levers that swing the underwriting. Directional, illustrative.
Technology roadmap · what changes the game
Demand drivers
- ↑ Population-scale deployment surface: UPI-grade digital rails + 22 scheduled languages = the world's largest inference market by users.
- ↑ Cost wedge at every layer: subsidised GPUs at ~1/3 global, inference at 1-2 orders below frontier APIs, engineering talent at GCC rates.
- ↑ Sovereign procurement as anchor demand: the government is the model layer's first paying customer.
- ↑ $67.5bn of hyperscaler pledges since Oct 2025: even partial delivery transforms the compute base.
- ↑ GCC densification: 2,100+ centres, ~80% of new launches AI-mandated.
Risks
- ! Frontier dominance: a 166x ecosystem funding gap cannot be closed by subsidy; India's models compete on price and language, not capability.
- ! The disbursement gap, ₹400 cr released of ₹10,372 cr, means the policy anchor is mostly still a press release.
- ! Monetisation unproven where valuations are highest: Krutrim's retrenchment is the base case, not the tail.
- ! Talent drain in place: GCCs absorb the best AI engineers at premium comp for foreign balance sheets.
- ! Froth markers everywhere: 5x valuation step-ups in a quarter, 30x ARR, pre-revenue unicorns.
What it means · by capital type
The barbell is the strategy: application-layer companies with real ARR (Emergent-class) at the cheap end, data/eval infrastructure (Deccan, Composio) as picks-and-shovels. Leave sovereign-model equity to strategics and the state.
Compute is the institutional lane: Blackstone's Neysa structure (half debt) is the template: real assets, contracted demand, subsidy tailwind. AI services (Fractal-class) offer cash flows but the 109x entry multiple is the warning.
Trade the layer rotation: Netweb's +139% says infra is priced; Fractal-flat says services aren't believed; the gap closes one way or the other. Watch IndiaAI disbursement runs as the sector's true leading indicator.
The mission's binding constraint is disbursement machinery, not ambition: ₹400 cr of ₹10,372 cr in two years undermines the sovereign-AI thesis more than any benchmark dispute. Fix the pipe before adding GPUs to the target.
Data vintage July 2026. Anchored to 2025-2026 industry and official prints; figures across sources differ and are reconciled to the cited ranges. Sources: Inc42: H1 2026 AI funding (+317%)S · Blackstone: Neysa $1.2bn (Feb 2026)P · Sarvam: Series B announcementP · Medianama: IndiaAI ₹400 cr disbursementS · Digital India: 34,000+ GPU milestoneP · TechCrunch: Krutrim cloud pivotS · Layer mix, market path, geography: GW estimatesE
Data confidence. High on funding rounds, GPU tenders and listed prints (P/S); medium on market-size projections (definitional spread is wide); company-reported ARR/profit claims flagged individually (E).
Sourcing. Every figure is sourced and dated. We tier provenance: Primary (official, regulatory, exchange or company filings), Secondary (tier-1 industry research and reputable media), and GW estimate (our own reconstruction or opinion, labelled, never presented as external fact). We prefer primary where it exists, reconcile divergent prints to cited ranges, and hold every number point-in-time: dated, and never silently restated; revisions publish as dated changes.
Fact vs opinion. Facts vs opinion: market sizes, official prints, prices, named deals and agency ratings are sourced facts (Primary/Secondary). Scores, grades, purity weights, scenario paths and indicative sparkline points are Gravitywell's analytical opinion (GW estimate): labelled, not presented as external data.
The sector, each cycle.
AI: The India Stack refreshed every cycle, with the scorecard, dashboard, and capital read. More sectors rolling out.
← All sectors