## INDUSTRY OUTLOOKCapacity trajectory:* 2007: 75 MW → 2020: 597 MW → 2025: 1,650 MW → 2030E: 5,000 MW* 26% CAGR over next 5 years* 3 GW active pipeline | ~$25Bn capex neededThe S-curve story:2G/3G (2007–14) → smartphones + Jio 4G (2016–20) → COVID digital shift → 5G + cloud (2022–24) → AI inflection (2025+)═══════════════## 🎯 TRENDSDual demand engine:1️⃣ Foundational: cloud adoption, data localization, 5G, BCP2️⃣ AI workloads: IndiaAI Mission, enterprise AI, sovereign computeStructural shifts underway:* MW → GW scale competition (AI factories)* Training → Inference dominance (global GPU spend: 34% inference in 2023 → 36% in 2027, growing 5x faster)* Air-cooled → Liquid- cooled standard* Colocation → Vertical integration (DC + GPU cloud + software)* Domestic-only → Global capital + global operators entering═══════════════## AI ADOPTION INSIGHTS* India = 2nd largest ChatGPT user base globally (9% share, behind US at 18%)* AI market: $13Bn (2025) → $130Bn (2032) at 39% CAGR* 45% of enterprises already deploying AI; only 6% haven’t started* 64% want in-house AI on cloud GPUs — the demand bedrock for domestic Neoclouds* 1,800+ GCCs (500+ AI-focused); 89% of new startups AI-nativeSegment-wise AI market (2025, $13Bn):BFSI $2.5b | Startups $1.8b | Media $1.6b | Mfg $1.4b | Tech Svcs $1.3b | Public $1.2b | Others $3.3bIndiaAI Mission ($125Bn / 5 yrs):* 38,000+ GPUs committed | 22,000 (~58%) allocated* 3,000 datasets, 243 AI models across 20 sectors* Selected LLM developers: Sarvam, Gnani, Soket, Gan.AI═══════════════## 🧱 KEY BUILDING BLOCKS (Value Chain)Bottom-up stack:1. Data Centres (Sify, NTT, Equinix) — physical infra2. GPU Hardware (Nvidia 90–95% share, AMD ~5%, Intel <1%)3. AI Cloud Service Providers (AWS, CoreWeave, Yotta) — the gateway4. Compute Software (GPT, Gemini, Sarvam, DeepMind)5. End-user Apps (Copilot, ChatGPT, Gemini)GPU roadmap (Nvidia): Hopper ‘24 → Blackwell ‘25 → GB300 ‘26 → Rubin ‘27 → Feynman ‘28Useful life logic:* Latest gen → Training (12–15 months per LLM)* N-1/N-2 gen → Inference (perpetual)* 6-year financial life; longer physical life═══════════════## 💰 GPU CLOUD UNIT ECONOMICSPer Nvidia H200 8-GPU server:| Metric | Value || Capex | ₹27.3 M || Revenue | ₹12.5 M/yr || EBITDA | ₹9.8 M/yr || EBITDA margin | 78.6% || Payback | ~2.8 yrs || Pricing | ₹195/GPU/hr (blended) || Utilization | 88% blended || PUE | 1.4 (liquid- cooled) |Sized cluster (3,000 GPUs / 375 servers):* Project IRR: 20.3% | Equity IRR: 28.4% (HTM basis)* Leverage: 60% | Debt cost: 10% | Tenor: 5 yrs | DSCR: 1.5x* Contract mix: 75% take-or-pay / 20% merchant / 5% spotPricing tiers: Take-or-pay ₹300/mo | Merchant ₹225 | Spot ₹300 (with low utilization)═══════════════## 🗺️ DC HUB MAP (1,650 MW total)| Hub | Capacity | Vacancy | U/C | Pipeline | Mumbai | 801 MW 2.9% 448 MW 893 MW| Chennai |268 MW | 12.4% | 153 | 273 || Delhi-NCR | 161 | 10.2% | 47 | 270 || Hyderabad | 138 | 9.7% | 106 | 200 || Bengaluru | 119 | 6.9% | 19 | 107 || Pune | 111 | 2.0% | 30 | 160 || Kolkata | 17 | 3.5% | 15 | 84 |Takeaways:* Mumbai = 50% of capacity, 47% of incremental supply (cable landing moat: 12 stations)* Chennai = 15% incremental, 3 new subsea cables landing 2026–27* Hyderabad = 11% incremental, hyperscaler self-build hub* Mumbai now ranks 6th globally in under- construction capacity