Ownership vs Rental: An Interactive Cost Model

Companion to Ireland's AI Research Gap by Conor Igoe.

This model estimates the annualised cost per GPU-hour of owning and operating a national AI GPU cluster in Ireland. Hardware is amortised over a 5-year replacement cycle. Facility construction and land are amortised over 20 years. It compares this against reserved commercial cloud rental (e.g. AWS, CoreWeave, Lambda). Ireland's current interim arrangement — renting time on Luxembourg's MeluXina through ICHEC — may have different pricing, but the contract terms are not public; commercial cloud provides a transparent benchmark.

Key Parameters

Number of GPUs800
Utilisation rate65%
Cloud rental rate (€/GPU-hour, 1-year reserved)€2.50

Assumptions

Defaults are calibrated to Irish parameters. Hover any row for an explanation. Click values to edit.

GPU unit cost
Intra-server overhead×
Networking overhead×
Bulk discount×
Hardware lifetime yr
Land acquisition
Facility cost / MWM
Facility lifetime yr
Chip TDPW
System power overhead×
PUE
Staff team size
Avg staff salary
Electricity rate/kWh

The Math

Every step from inputs to cost-per-GPU-hour. Hover any value for an explanation.

Sources: Epoch AI GPU Clusters Dataset and methodology (Pilz et al., 2025); Cottier et al. (2024) for hardware cost decomposition; Epoch AI ML Hardware Dataset and Chip Sales Documentation for GPU pricing; Turner & Townsend Data Centre Construction Cost Index 2025–2026; Eurostat electricity price statistics, H1 2025; SCSI/Teagasc land market data, 2025; Silicon Data GPU rental indices; cloud pricing from Lambda, CoreWeave, GMI Cloud, RunPod, and AWS (early 2026).