Jio Arogya · AI Cost Explorer
This site is access-restricted.
Please enter the password to continue.
Incorrect password. Try again.
← Home
Jio Arogya · AI Cost Model

JioCare AI Investment Decision Explorer

Interactive cost analysis: Rent vs Build postures | Confidential — May 2026

Member Scale (actual user count per year)

Slider-driven
Manual edits override the growth slider for that year. Touch the growth slider to reset.

Engagement Multiplier by Year

Slider-driven
Manual edits override the global engagement slider for that year. Touch the engagement slider to reset.

Scenario Controls

Stress-test the baseline model assumptions shown in the Reference section below. Changes here are scenario overrides only and do not alter the underlying source definitions.

Usage Growth SpeedMultiplier on the baseline member ramp (10K → 50M over 7 years). 1.0x = baseline forecast. 1.0x
Slow (0.5x)Fast (2.0x)
API Price Decline Rate (per year)Annual percentage drop in per-token and per-minute prices. Applied equally to third-party API rates and self-hosted inference costs, on the assumption that hardware and efficiency gains compress both at a similar rate. 0% = prices stay flat. 0%
0% (no decline)70%/year
Build Start YearThe year proprietary model training begins. Before this year, the Build posture runs on rented APIs and accumulates consented, de-identified usage data that can lower later annotation and corpus construction costs. Y3
Y1 (day-zero anchor)Y5
Engagement IntensityMultiplier on baseline AI interactions per member per year (496 events a year at 1.0x). An event is a single AI-mediated action — user-initiated (chat, voice query, document scan) or system-initiated on behalf of the member (CHW prep briefs, telehealth pre-consult summaries, clinic follow-up triggers). 1.0x · 496 ev/yr
Low (0.25x · ~124 ev/yr)High (3.0x · ~1,488 ev/yr)

Scenario Presets

Annual Cost Comparison

Total cost per year across Rent vs dynamic Build ($M)

Cumulative Cost Over Time

Running total spend — crossover points highlighted ($M)

Per-Member Inference Cost

Unit economics: $/member/year by posture

CapexRENT capex: PII gate, safety pipeline, STT/TTS, orchestration, R&D team. BUILD capex adds training compute, GPU cluster, corpus, annotation, distillation, eval infra. vs OpEx Breakdown

Investment structure by year — Rent vs Build grouped stacked view ($M)

Data Asset Accumulation

Shows data generated by member activity. Not priced as an asset — shown to illustrate the compounding data flywheel that underpins the Build thesis.

Cumulative Platform Data Over Time

Raw data exhaust vs training-usable corpus — cumulative tokens across all members

Reference: Total publicly available Indian medical corpus is estimated at ~70–80B tokens. The platform surpasses this by Y2 in cumulative raw exhaust.

Data Composition Breakdown

Annual token volume by source — scaled by member count and engagement

Agent conversations dominate volume (~46%), but clinical encounters (doctor, specialist, hospital) carry highest per-token diagnostic value for model training.

Detailed Year-by-Year Breakdown ($M)

Decision Insights

Reference: Definitions, Assumptions & Methodology

These panels describe the baseline assumptions behind the model. Controls above test alternate scenarios without changing the underlying source definitions. All figures are Y2+ steady-state unless noted.

Source documents: 1_Usage_and_Inference_Demand.md · 2_AI_Light_Rent.md · 3_AI_Full_Build.md (incl. Data Dividend for Delayed Build Starts) · 4_AI-Build-Posture_Full-Context.md — AI Cost Model, May 2026.

1

What Drives These Costs: Per-Member AI Engagement Breakdown

← 1_Usage_and_Inference_Demand.md
▼

Every cost figure in this model traces back to a concrete set of AI engagement events per member per year. The baseline is ~496 personal agent events/member/year (Y2+ steady-state), plus clinical encounter events layered on top.

A. Personal Agent Events (496/member/year)
Feature Category Events/member/yr LLM calls/event Tokens/call Routing (Frontier / Mid / Flash) Voice & OCR per event
Passive / proactive
Daily briefings, wearable alerts, weekly patterns
247 1 500 0% / 10% / 90% None
Admin / reminders
Medication adherence, appointments
109 1 500 0% / 20% / 80% 30% voice: 1 min ASR + 700 chars TTS
Wellness / diet / medication coaching 113 2 1,500 0% / 30% / 70% 50% voice: 1.5 min ASR + 1K chars TTS; 20% meal photo: 1 OCR page
Symptom triage / care navigation 14 5 3,000 20% / 60% / 20% 50% voice: 2 min ASR + 1.2K chars TTS
Document / lab / prescription help 13 3 6,000 20% / 70% / 10% 2 OCR pages / event
Total personal agent events 496
B. Clinical Encounter Events
Encounter Type Rate/member/yr (Y5) LLM calls/encounter Key Workflows
CHW visits 5.0 6 Pre-visit prep, screening, escalation summary, follow-up, voice capture
Doctor consults
Telehealth + clinic
13.0 5 Pre-consult summary, differential + safety check, care plan + documentation
Specialist referrals 2.0 6 Referral package, post-visit reconciliation, patient summary
Hospital admissions 0.15 7 Admission brief, discharge reconciliation, discharge instructions
Pharmacy events 12.0 3 Drug interaction, substitution, adherence plan
Post-encounter reconciliation fires after CHW + doctor 1 Longitudinal record update
C. Y5 Per-Member Demand Summary
Modality Annual Volume per Member
LLM calls (all tiers) 855
LLM tokens (all tiers, incl. 1.4x overhead) 2.56M
   — frontier tokens 329K
   — mid-tier tokens 1,395K
   — flash tokens 834K
ASR minutes 213
TTS characters 113,373
OCR pages 64
2

User Segments: Who Uses What

← 1_Usage_and_Inference_Demand.md
▼

Not all members engage equally. The 496 events/year is a blended average across three engagement segments. This distribution is critical — it means 20% of users drive disproportionate cost, and 40% barely use the personal agent at all.

Segment Share Events/year Profile
Power users 20% 1,383 Chronic disease (diabetes, cardiac, pregnancy), daily wearable, highly motivated
Moderate 40% 399 Engaged 2–3x/week, responds to nudges, logs meals occasionally
Light 40% 150 Enrolled via insurance, minimal self-initiated use, engages around CHW visits
Blended 100% ~496 Weighted average across segments

Verification: (0.20 × 1,383) + (0.40 × 399) + (0.40 × 150) = 276.6 + 159.6 + 60.0 = 496.2 ≈ 496

3

Pricing Assumptions: API & Self-Hosted Rates

← 2_AI_Light_Rent.md · 3_AI_Full_Build.md
▼
API Prices (May 2026, Rent posture)
Service Unit Price Use
Frontier LLM (Opus / GPT-4.5) $10.00 / M tokens Differential diagnosis, safety checks
Mid-tier LLM (GPT-4o / Sonnet) $4.38 / M tokens Pre-consult summaries, care plans
Flash LLM (Gemini Flash) $0.26 / M tokens Reminders, admin, low-risk follow-ups
ASR (Indic + English) $0.006 / minute Voice input
TTS (Indic + English) $0.015 / 1K chars Voice responses
OCR / document extraction $0.03 / page Prescriptions, lab reports
Self-Hosted Prices (Build posture, by model phase)
Model $ / M tokens What It Serves
Phase 1 (8B active, MoE 30B) $0.30 First proprietary model
Phase 2 (60B active, MoE 300B) $0.50 Mid-scale clinical workloads
Phase 3 / GP-level (100B active, MoE ~900B) $0.80 Full GP model
Cloud distilled clinical $0.12 Mid-tier equivalent
Cloud distilled flash $0.02 Routine / admin
Local / JORO / on-device $0.00 Eligible workloads on user hardware

API prices are as of May 2026 list rates. The "API Price Decline Rate" slider above models year-over-year compression of these rates. Self-hosted inference cost assumes amortized GPU cluster (768× H100/H200) over 5 years.

4

Build Investment Breakdown: Where the Capex Goes

← 3_AI_Full_Build.md
▼
Phased Model Training — Kill Gates

Each phase has a binary kill-gate: if the answer is "no", stop investing and revert to Rent. This is the core risk-management mechanism of the Build posture.

Phase Timeline Model Cost Kill-Gate Question
0 Months 1–6 Fine-tune OSS 70B $0.3M Does fine-tuning improve Indian clinical performance?
1 Months 6–12 30B MoE (8B active) $7.4M Beat frontier APIs on Indian clinical evals?
2 Months 12–18 300B MoE (60B active) $39.7M Achieve task-shift multipliers in pilot?
3 Months 18–24 800B–1T MoE (100B active) $199M Full production deployment
Full 5-Year Capex — Y1 Build Anchor
Component Cost ($M)
Training compute (Phases 0–3 + retraining) $288M
Corpus construction $22M
Manual annotation (1,776 peak FTE) $30M
Synthetic data pipeline $13M
Eval & safety infrastructure $11M
Distillation + local modalities $18M
GPU cluster (768 H100/H200) + clinical validation $66M
PII gate + orchestration $4M
R&D team (45→95 headcount) $91M
Total 5-yr capex $546M
Data Dividend for Delayed Build Starts

When Build starts after Y1, the platform has already accumulated consented usage data. The model credits only the portion that survives consent, de-ID, clinical relevance, and quality filters, reducing annotation and corpus construction costs while leaving training compute, GPU cluster, synthetic data, safety, distillation, and team costs unchanged.

Build Start Clean Data Clinical Annotation Pool Annotation Discount Annotation Cost Corpus Discount Corpus Cost Data-Creation Savings
Y10.0B0.0B0.0%$30.1M0.0%$22.0M$0.0M
Y26.4B1.2B2.0%$29.5M1.6%$21.6M$1.0M
Y360.1B10.8B18.6%$24.5M15.0%$18.7M$8.9M
Y41,824.1B328.3B55.0%$13.5M25.0%$16.5M$22.1M
Y520,548.1B3,698.7B55.0%$13.5M25.0%$16.5M$22.1M
5

Overhead Multiplier: Why Raw Demand × 1.4

← 1_Usage_and_Inference_Demand.md
▼

The per-member demand figures (855 LLM calls, 2.56M tokens) already include a 1.4x overhead multiplier applied to raw engagement volumes. This multiplier captures the production reality that every user-facing AI call triggers additional system calls.

Overhead What It Covers Multiplier
Safety pipeline Input guardrails, output hallucination check, clinical audit log 1.15×
PII filtration model Separate NER scrub per cloud call +2% calls
Data validation model Input quality checks on health data +1% calls
Real-world overhead Retries, cache misses, multilingual expansion, A/B testing 1.20×
Combined effective multiplier ~1.4×
6

Member Ramp with Context

← 4_AI-Build-Posture_Full-Context.md
▼

The member ramp is the single largest driver of total cost. The 500× growth from Y1 to Y5 is aggressive but assumes Jio distribution activation — the same channel that scaled JioPhone to 100M+ users.

Year Members What's Happening
Y1 10,000 Jamnagar pilot — Reliance employee families
Y2 50,000 Expanded pilot — Gujarat / Mumbai metro
Y3 1,500,000 National launch — first mass enrollment
Y4 15,000,000 Rapid scale — Jio distribution activated
Y5 50,000,000 Full national scale

Y6–Y7 hold at 50M (steady state). The "Usage Growth Speed" slider above scales all member counts linearly — at 2.0x, Y5 = 100M members.