Best viewed on desktop

This financial model has detailed inputs and multi-year tables that require a larger screen to use effectively. Open it on your laptop or desktop for the full experience.

Get the desktop link
AI Infrastructure Economics · 2025

Cloud vs. On-Premises
for AI Workloads

GPU economics are fundamentally different from general compute. Training is bursty; inference is steady-state. Utilization rate is the variable nobody models. This calculator gets that right.

GPU pricing: H100 · A100 · L40S · A10G Models: Training · Fine-tuning · Inference After-tax NPV with 2–3yr hardware refresh
Scenarios → Load a scenario or set custom inputs ready
⬛ GPU Cluster Configuration
GPUs
%
Loading GPU pricing…
⬛ Workload Split
%
%
☁ Cloud GPU Pricing
%/yr
$/yr
$/yr
$
▪ On-Premises GPU Infrastructure
$
↑ Auto-calculated from GPU type × count. Override if needed.
$
$
$
%
W
$/kWh
%
⬛ MLOps Staffing (On-Prem)
FTE
$/yr
FTE
⬛ Financial Parameters
%
%
Workload Split — Training vs. Inference
Training 30%
Inference 70%
Training — bursty, episodic. Cloud almost always wins. Inference — steady-state. On-prem wins at scale past Yr 2–3.
AI Infrastructure Verdict
—
Load a scenario or adjust inputs.
Cloud GPU Total NPV
—
present value, all-in
On-Prem GPU Total NPV
—
net of depreciation tax shield
NPV Advantage
—
winner saves X in NPV
Effective Cost / GPU / Month
—
at your utilization rate
NPV Cost Comparison — Proportional Scale (lower = better)
☁ Cloud GPU infrastructure —
▪ On-Premises GPU infrastructure —
GPU Utilization Impact — Why This Number Changes Everything
—
Your Utilization
—
Cloud pays for actual use
—
On-prem break-even util.

—

Cost Component Yr 0Yr 1 Yr 3Yr 5Yr H
Nandeep Nagarkar · SVAG Labs · AI Infrastructure Economics Series
Compute runs server-side. Methodology, GPU pricing data, and scenario library are proprietary to SVAG Labs (Single View Agile Governance).
Power calculation: GPU count × watts × PUE × 8,760 hrs × $/kWh. Depreciation straight-line over refresh cycle. Validate with vendor quotes before board presentation.
Reference material for the Machine Capital Brief.