# The India AI Stack, graded: built from both ends toward the middle > TCS just committed $7.4B to one Hyderabad campus. Map India's GenAI push onto the four layers we track and a clear shape emerges: strong at the bottom (compute) and the top (population-scale apps), thin in the middle (models, middleware) — the inverse of how the West built it. A research-based read, layer by layer. Source: https://theaidaily.in/analysis/india-ai-stack.html Published: 2026-09-08 --- The signal behind the AI headlines — **ranked, not recapped**[Get the daily brief →](/subscribe.html) [********The AI Daily](/)[Subscribe →](/subscribe.html)☰ 🇮🇳 Deep dive · The India AI Stack India is building the AI stack from both ends TCS just committed **$7.4 billion to a single Hyderabad campus**. Map India’s GenAI push onto the four layers we track and a clear shape appears: strong at the bottom (compute) and the top (population-scale apps), thin in the middle (models, middleware) — almost the inverse of how the West built it. The AI Daily · Analysis · **September 8, 2026** ## The one-scroll version - **One deal contains the whole thesis.** A services company — TCS — is spending **₹70,000 cr (~$7.4B)** to become a compute *landlord*. India isn’t buying its way up the stack; it’s building the ends first. - **The shape is a barbell, not a pyramid.** Compute (bottom) and applications (top) are lit. Foundation models and middleware — the profitable middle the West monetised first — are the **thinnest part of India’s stack**. - **Compute is loud but leased.** Gigawatt campuses and a ₹1.28 lakh cr chip mission are real — but they run on **imported GPUs at a 15%+ premium**, and on power and water India hasn’t secured yet. - **The models layer is a one-name bet.** Sarvam is a genuine sovereign champion; behind it the bench is thin and the logic is **sovereignty, not yet revenue**. - **India’s strength is also its exposure.** Its best AI-adopting sector — IT services — is the one AI is **deflating 25–30%**. The infra land-grab is partly a hedge against its own core business shrinking. Start with the deal everyone read as an infrastructure headline. In early September, TCS’s subsidiary **HyperVault** acquired 264 acres in Hyderabad for a **1 GW, ₹70,000 crore (~$7.4–8.4 billion) AI data-centre campus** — India’s largest single AI-infrastructure commitment, aimed squarely at hosting global hyperscalers and frontier labs ([MediaNama](https://www.medianama.com/2026/09/223-tcs-hypervault-ai-data-centre-telangana/), [ET](https://economictimes.indiatimes.com/tech/information-tech/tcs-subsidiary-hypervault-to-invest-rs-70000-cr-to-develop-hyderabad-ai-data-centre/articleshow/133797995.cms)). Look past the number and it’s a tell. India’s most famous *services* company is spending eight figures a day to become a compute *landlord*. That single move — a giant of the old application layer racing to pour the new foundation layer — is the whole story of India’s AI stack in miniature. To see why, map the country’s GenAI push onto the four layers we track across every other market, from [the Killer App Watch](/analysis/killer-app-watch.html). What emerges isn’t a ladder being climbed rung by rung. It’s a stack being built **from both ends toward the middle**. The India AI Stack · graded ### Where India’s build-out actually stands, layer by layer Bars = our read of India’s progress at each layer, demo → dollars → default. Read bottom-up, the way it’s being built. Applications ⚡ India’s strength — population scale 68 Landing: Bajaj Finance runs **71% of DIY service** on AI (₹2,500 cr in disbursements); SEBI’s SUDARSAN watches markets; UPI & Aadhaar are agent-ready rails. The catch: …yet the flagship app — **IT services** — is being AI-deflated 25–30%. Real ROI at the top, thin breadth below it. Middleware & platforms The missing middle — the real gap 22 Landing: Razorpay **Vulcan** (a payments foundation model), Pine Labs’ agentic-UPI protocol, Gnani’s Plexus agent platform. The catch: No horizontal orchestration, serving or dev-tooling layer at scale. One genuine edge: **payments data** is a moat Western labs don’t have. Foundation models Nascent — one leader, thin bench 35 Landing: **Sarvam** — India’s 130th unicorn ($234M + $74M with NVIDIA on the cap table), a trillion-parameter roadmap, government-commissioned cyber models. The catch: Behind it: Gnani’s 30B, Soket, a hedging BharatGen, a shrinking Krutrim. The whole layer rests on a **sovereignty** bet, not yet on revenue. Compute & infrastructure ⚡ The loudest layer — capital first 72 Landing: TCS HyperVault’s **1 GW, ₹70,000 cr (~$7.4B)** Hyderabad campus; Andhra’s ~$34B of approvals; Amazon +$13B; Semicon 2.0 at ₹1.28 lakh cr. The catch: But it’s scaling on **imported silicon** — NVIDIA’s 15%+ server hike lands here directly, GPUs are on a waiting list, power and water are the ceiling. The West monetised this stack bottom-up over a decade — infra, then middleware, then apps. India is lit at the **ends** and hollow in the **middle**. That middle is the sovereignty gap it’s now racing to close. Grades are The AI Daily’s editorial read of publicly reported activity across May–September 2026, not a market-share index. Sources are linked in each section below. ## Bottom layer: compute is loud — and leased No layer is louder right now. Inside a single four-month window, India moved multiple **gigawatt-class campuses** from policy to committed capital: TCS HyperVault’s 1 GW in Hyderabad, Google’s 1 GW in Anakapalli, a Colt–RMZ gigawatt build in Vizag. Andhra Pradesh alone cleared **₹2.83 lakh crore (~$34B)** of data-centre projects and ~3,498 MW of power ([MediaNama](https://www.medianama.com/2026/08/223-andhra-pradesh-ai-data-centres/)); Amazon added **$13B** for India cloud and AI infra ([TechCrunch](https://techcrunch.com/2026/06/25/amazon-ups-india-bet-with-fresh-13b-ai-infrastructure-investment/)); Microsoft opened its largest India data-centre hub. On chips, the Cabinet notified **Semicon 2.0 at ₹1.28 lakh crore**, targeting 100,000 new chip engineers ([Inc42](https://inc42.com/buzz/centre-notifies-%e2%82%b91-28-lakh-cr-semicon-2-0-targets-1-lakh-more-chip-engineers/)). But read the fine print and the layer is **building on borrowed silicon**. The GPUs inside these campuses are imported, scarce, and getting pricier: NVIDIA’s warned 15%+ server price hike lands on India directly ([Inc42](https://inc42.com/buzz/nvidias-ai-server-price-hike-threatens-to-inflate-indias-compute-costs/)), and domestic buyers describe a GPU “waiting list.” Semicon 2.0 is real money, but it funds chip *design* and packaging (Paras Defence’s $644M unit, Marvell’s $250M expansion) — not yet a fab printing AI accelerators. And the true ceiling isn’t land or money; it’s **power and water**. India is already reaching for small-modular-reactor law and BIS energy standards to feed these campuses ([MediaNama](https://www.medianama.com/2026/08/223-govt-nuclear-power-law-bis-standards-ai-data-centre-strains/)). India is pouring the foundation faster than almost anyone. It just doesn’t yet make the thing that goes inside it. India’s most famous services company is spending eight figures a day to become a compute landlord. That is the whole story in one deal. ## Second layer: foundation models — a one-name bet Go up a layer and the volume drops sharply. India has exactly one foundation-model champion with real capital and momentum: **Sarvam**, which became the country’s 130th unicorn on a $234M round led by HCLTech, then extended it by **$74M with NVIDIA on the cap table** ([Inc42](https://inc42.com/buzz/sarvam-to-raise-74-mn-in-series-b-extension-from-nvidia-others/)). It has announced a trillion-parameter roadmap and a home-hosted inference service, ships a coding agent, and has been *commissioned by the government* — alongside BharatGen — to build sovereign cybersecurity models ([Inc42](https://inc42.com/buzz/centre-asks-sarvam-ai-bharatgen-to-develop-mythos-like-cyber-ai-models/)). HCLTech is even building it a dedicated **₹14,257 cr data centre in Odisha**. Behind that single name, the bench thins fast. Gnani.ai has a credible 30B Indic model inside its Artha stack; Soket is early; BharatGen’s own CEO argues [“models alone can’t help India win”](https://inc42.com/features/bharatgen-ceo-on-why-models-alone-cant-help-india-gain-in-the-ai-race/); and Ola’s Krutrim just cut nearly half its remaining staff. Crucially, the layer’s logic is **sovereignty, not revenue**: MeitY has told central ministries to *pause* OpenAI and Anthropic over data concerns ([MediaNama](https://www.medianama.com/2026/07/223-government-asks-ministries-pause-deployment-openai-anthropic-ai-models/)), and the state may take a **1–2% stake in Sarvam** directly. The trigger is fresh in memory: a US export ban on frontier models earlier this year “rattled” India’s tech community and turned sovereign AI from a talking point into a policy. This is a real bet — but it is a bet on one horse, funded by conviction about independence rather than a market that pays for Indic tokens yet. ## The hollow middle: middleware barely exists This is the thinnest part of India’s stack, and the most revealing. In every mature market, the middleware layer — orchestration, serving, agent platforms, dev tooling — is where durable margins pool; it’s the layer that [just lit up globally](/analysis/killer-app-watch.html) with Palantir, Databricks and ServiceNow. In India it is almost empty. What exists is **vertical, not horizontal**: Razorpay’s **Vulcan**, a foundation model trained on payments data rather than text; Pine Labs’ protocol for agents making autonomous UPI payments; Gnani’s Plexus agent platform bundled inside its model. There is no Indian orchestration or model-serving company of scale in the record. The upside hides in that same fact. India’s one defensible middleware asset is built on something Western labs can’t easily replicate: **payments data at population scale**. Vulcan understands “how money moves” across UPI in a way GPT-class models trained on the open web simply don’t. If India builds a globally interesting middle layer, it will likely grow up out of its rails — payments, identity, language — not down from generic models. For now, though, the middle is a gap. ## Top layer: applications are India’s strength — and its exposure Here India looks most convincing, and the reason is structural: it can deploy AI onto rails that already reach a billion people. The proof points are the rare kind tied to *disclosed outcomes*, not pilots. **Bajaj Finance** now runs **71% of its DIY customer service** on AI and credits it with ₹2,500 crore in quarterly disbursements ([MediaNama](https://www.medianama.com/2026/08/223-bajaj-finances-71-customer-service-2500-crores-disbursements/)). **Naukri** says its AI tools lifted ARPU ([MediaNama](https://www.medianama.com/2026/08/223-ai-arpu-new-monetisation-lever-naukri/)). The market regulator **SEBI** runs an AI system, SUDARSAN, that scans regional-language social media and filings for fraud. WhatsApp ships on-device scam detection to hundreds of millions. This is population-scale deployment, and it’s the layer where India can genuinely leapfrog. Then comes the twist that makes the whole stack cohere. India’s single biggest AI-*exposed* business is the same one that built its tech economy: **IT services**. And AI is deflating it. Persistent’s CEO put a number on it — AI can [**“deflate the effort by 25% to 30%”**](https://economictimes.indiatimes.com/tech/information-tech/ai-to-deflate-it-revenue-past-30-persistent-ceo/articleshow/132869284.cms) — and the majors are already rejigging contracts from billed-hours to outcomes. LTIMindtree booked its first **$150M of standalone AI revenue** even as Info Edge watched Shiksha billings fall 23% to AI-changed search. That is precisely why TCS and HCLTech are sprinting into data-centre *landlording*: the infra land-grab at the bottom of the stack is, in part, a **hedge against the application business at the top shrinking**. India’s strongest layer and its most exposed layer are the same layer. ## The four forces that decide it A stack shape is a snapshot; these four cross-cutting forces are the physics that will move it. None of them sits on a single layer — each one bends the whole thing. 🇮🇳 #### Sovereignty MeitY paused OpenAI & Anthropic across central ministries; the government may take a 1–2% stake in Sarvam. The whole push was lit by the US export ban on frontier models that “rattled” India’s tech community. ⚡ #### Power & cost The concrete is easy; the electrons are not. India is invoking small-modular-reactor law and BIS energy standards to feed gigawatt campuses — and paying NVIDIA’s 15%+ hardware premium on top of imported silicon. 👥 #### Talent India’s real edge. GCCs already steer **$55B** of IT exports; OpenAI poached Uber’s India chief to run its biggest market outside the US; Semicon 2.0 targets **100,000** new chip engineers. ⚖ #### Governance The gaps are widening as fast as the capex. UIDAI updated the biometrics of **~2 crore children** with consent questions; deepfake rules go largely unenforced; there is still no health-data law under the AI being deployed on patients. ## The honest counter-case #### Three ways this read is too neat First, **“both ends” can become “stuck at both ends.”** A country can pour concrete and deploy chatbots for years without ever building the profitable middle — and the middle is where the compounding value and the sovereignty actually live. Second, **imported everything is a policy, not an accident, and it has a ceiling.** If GPU access tightens or prices keep climbing, the gigawatt campuses become expensive shells. Third, **the app-layer “strength” is thinner than the headline names suggest** — Bajaj, Naukri and SEBI are real, but most population-scale plays (DigiLocker/UMANG agents) are still at tender stage, and the sector carrying the flag is the one being deflated. The optimistic read is a barbell that fills in. The bearish read is a barbell that stays hollow while the electricity bill comes due. ## So what — for people placing bets on India #### Buy the rails, watch the middle India’s durable edge is at the ends — population-scale distribution (payments, identity, language) and now compute capacity. The asymmetric opportunity is the empty middle: orchestration and serving built *on top of* those rails. Whoever builds India’s middleware from the UPI/Aadhaar layer up, rather than importing generic tooling, owns the scarce part. #### Underwrite power, not just capex The gating constraint on the infra story isn’t money or land — both are flowing. It’s **electrons and imported silicon**. Stress-test any India data-centre thesis against secured power (and its price) and GPU supply, not announced megawatts. #### Separate sovereignty from returns The models layer is a national bet with government capital behind one champion. That can be a good bet and still not be a venture return. Price Sarvam and its peers on the sovereignty mandate that funds them — and don’t assume a market that pays a premium for Indic tokens exists yet. ## The bottom line The West built its AI stack bottom-up and monetised the middle first. India is building **from both ends toward the middle** — pouring compute at the base and deploying AI onto billion-person rails at the top, while the foundation-model and middleware layers between them stay thin and state-backed. TCS spending $7.4B to become a landlord is the tell: the country’s AI ambition is real, well-capitalised, and genuinely distinctive — but it is **racing to fill a hollow middle before the imported silicon gets more expensive and the power runs short**. Watch that middle. It’s where India’s AI decade will actually be won or lost. Related: [Killer App Watch](/analysis/killer-app-watch.html) · [The Nvidia Moat](/analysis/nvidia-moat.html) · [Enterprise AI Costs 101](/analysis/llm-cost.html) · [The $50B Club](/analysis/fastest-to-50b.html)