In the 2026 AI landscape, GPU compute is the single largest expense for most enterprises. This project provides a real-time FinOps Command Center designed to monitor the unit economics of AI inference. By linking hardware telemetry with business revenue, this engine identifies "Zombie GPUs" and calculates the true profit margin per token.
- Star Schema Data Model: Successfully integrated disparate data sources including GPU hardware catalogs, live utilization telemetry, and inference revenue streams.
- Zombie GPU Alerting: Implemented conditional logic to flag under-utilized hardware (Utilization < 15%), enabling immediate cost reclamation.
- Unit Economics Tracking: Developed custom DAX measures for Profit Ratio and Idle Waste Dollars to provide transparency into cloud spend.
- ESG Reporting (Carbon Impact): Built-in carbon footprint tracking (kg CO2) for 2026 sustainability compliance.
| Data Model Architecture | Operational Intelligence |
|---|---|
 |
 |
| Star Schema: Optimized for high-frequency telemetry logs. | Conditional Alerting: Identifying $2.85/mo in total fleet waste and flagging specific "Zombie" instances for decommissioning |
To demonstrate technical proficiency, here are the primary calculations used in this engine:
Profit Ratio:
Profit_Ratio = DIVIDE([Total_Revenue], [Total_GPU_Cost], 0)
Purpose: Measures the return on investment for specific GPU clusters.
Projected Monthly Waste:
Idle_Waste_Dollars = [Total_GPU_Cost] * (1 - AVERAGE(Utilization_Telemetry[Avg_Util_Pct]))
Purpose: Quantifies the financial impact of idle hardware.
Carbon Impact:
Carbon_Footprint_kg = SUM(Utilization_Telemetry[Power_Watts]) * 0.0004
Purpose: Aligns infrastructure spend with corporate ESG targets.
Using this engine, an organization can:
- Reduce Waste: Identify and terminate idle H100/A100 instances immediately.
- Optimize Pricing: Adjust token pricing models based on actual hardware costs versus revenue.
- Multi-Cloud Arbitrage: Compare profitability across regions (e.g., us-east-1) to find the most cost-effective provider.
- Why a Star Schema? I chose a Star Schema over a flat table to ensure the model can scale to millions of rows of telemetry data without slowing down report performance. This ensures stakeholders get 'Instant-On' reporting even with high-frequency GPU logs.
Note: Technical source files (.pbix) and data schemas are available upon request for verified recruiters.