AI Portfolio & ROI Framework
Delivered 3.2X ROI on a ₹65 Cr AI portfolio across 5+ product lines by modelling rigorous GPU training vs. inference unit economics. Established the financial evaluation framework later adopted for ₹150 Cr+ AI initiatives at JioMedia.
The Challenge
AI investments at large enterprises often lack rigorous financial evaluation. Scattered investments across multiple product lines made it difficult to measure true ROI, prioritize effectively, or make data-driven decisions about GPU allocation and model selection. There was no unified framework for evaluating and comparing AI initiatives.
The Solution
- Modelled rigorous GPU training vs. inference unit economics across 5+ product lines
- Consolidated scattered AI investments into a unified portfolio strategy with standardized evaluation criteria
- Built a financial evaluation framework that accounts for GPU costs, inference latency, model accuracy trade-offs, and business value
- The framework was adopted as the standard for evaluating the majority of ₹150 Cr+ AI initiatives at JioMedia
Impact
Key Decisions
- Unit economics over vanity scale: Evaluated each AI initiative on GPU cost per inference vs. business value generated, rather than just model accuracy or user volume.
- Unified portfolio view: Consolidated all AI investments into a single dashboard with standardized metrics, enabling leadership to compare apples-to-apples across diverse AI products.
- Training vs. inference split: Separately modeled training costs (one-time/periodic) from inference costs (ongoing/scaling) to prevent misleading ROI calculations.
Try the AI Cost Calculator
An open calculator built on the same unit economics: token costs across 16 current models, the request volume at which self-hosting an open-weights model beats an API, and GPU sizing for serving and training. Prices are dated and sourced, and the cost engine is unit-tested, including a check against Meta's published Llama 3.1 training compute.