Type: Independent product-strategy case study (PM assignment format: market sizing, pricing, operating model, investment thesis) · Domain: On-device/edge AI · India GTM · Enterprise monetization

Company context: Sarvam AI, an IndiaAI Mission-selected sovereign LLM company. Sarvam Edge (launched Feb 2026) runs compact models (speech, TTS, translation) fully on-device — the strategic question is where the Year-1 paid revenue actually comes from.

Core discipline: a well-reasoned Rs 50 Cr beats an unsupported Rs 500 Cr. The exercise explicitly rejects an industry-first model (50 use cases → a headline number nobody can defend) in favor of a device-capability funnel that only asks what can run the product today, with a payer, and a reachable channel.

Problem

On-device AI can't be sized like cloud SaaS — there's no per-query revenue, hardware compatibility gates who can even run the product, and "India = 1.4B people" produces a fantasy TAM. The real question: of devices in India, how many can run Sarvam Edge today, how many sit with a payer who has an Indic-or-sovereignty need, and how many can Sarvam reach through a real channel in 12 months?

Approach & Framework

Device-capability funnel (not industry-first): Total device base (179.7M: laptops + smartphones + edge appliances) → AI-capable share (~37M, gated by NPU/chipset) → Sarvam-relevant SAM (4.82M, Indic-language need + sovereignty driver + reachable channel) → Year-1 capture (~104K units) → Rs 23.3 Cr core Year-1 SOM. A second, independent payer-segment buildup (banks, government, education, health, MSME, insurtech) triangulates to Rs 25.5 Cr — within 10%, which is the actual validation, not a single model's precision.

Value-anchored pricing, not benchmark-anchored: e.g., the laptop-seat price is derived bottom-up from a BPO agent's fully-loaded cost, time saved per call, realization discount, and a 15% value-capture rate — landing at Rs 2,000/seat/yr, 15x cheaper than Copilot, which is the point (anti-anchoring: the comparison should resolve in 30 seconds).

Hybrid monetization by unit, matched to each payer's natural budget logic: per-seat subscription (laptops), per-device royalty (OEM channel, no direct end-user relationship), per-asset subscription (ATMs/cameras — SLA-bound), per-deployment platform fee (5K+ seat rollouts), rev-share (telco/insurer partner-led, since direct B2C fails on CAC against bundled Big Tech AI).

Output

• Named-account GTM: 15 specific target accounts with named buyer titles (Head of Contact Centre Ops, CISO, GM ATM Operations) — converts "enterprise BPO and banks" from a slogan into something a sales team can execute against
• v1 product spec: Sarvam Edge for BPO Agent Assist — one form factor (Windows/macOS laptop), one buyer, one workflow, with named failure modes (thermal throttle, transcript-confidence circuit breaker) and a concrete pilot success metric (handle-time reduction ≥15%)
• Engineering architecture defended, not just modeled: INT4-quantized 30B model fits ~5GB RAM budget; paged-attention KV cache bounds memory on long calls; signed model bundles with delta OTA for enterprise IT trust; PII redaction runs on-device before any cloud sync
• Sensitivity + break-your-model: single pivot lever (enterprise pilot-to-paid conversion) tested at ±30%, moving core SOM Rs 19.3–27.3 Cr — a 30% miss compresses revenue 17%, not by half, which is the actual test of whether a model is fragile
• Three named ways the whole thesis fails (Big Tech bundles free Indic AI; pilot conversion stalls on integration friction; hardware floor stays too narrow) — each with a 90-day leading indicator and a structural hedge, not a hope

Outcome (Illustrative — This Is a Strategy Submission, Not a Live Deployment)

Model output, not a result: Rs 23.3 Cr core Year-1 SOM (Rs 19.8–21.0 Cr net of cloud cannibalization), ~85% blended gross margin (on-device inference = near-zero marginal cost), CAC payback <12 months on multi-year contracts. Recommendation: bet Year 1 entirely on enterprise laptops in BPO/IT/private banks — same buyer, same channel motion, same surface — and explicitly do not chase direct consumer subscription, smart-city CCTV tenders, or macOS coverage in Year 1.

Why This Case Study

Demonstrates the discipline of naming your own model's fragility before someone else does — sensitivity analysis and break-your-model sections aren't hedging, they're the credibility mechanism. Also demonstrates hybrid-unit monetization design and India-grounded unit economics (COGS, CAC, integrator margin splits) rather than a generic SaaS pricing template.