Adjust utilisation, power, electricity prices, and amortisation to see yearly TCO for a 512-accelerator cluster. Defaults follow public pricing disclosures from NVIDIA, Google Cloud TPU, AMD, and Groq.
| Accelerator | Yearly capex amortisation | Yearly energy + cooling | Support/licensing | Total annual TCO |
|---|
Capex estimates: NVIDIA H100 SXM5 $32k (list), Google TPU v5e subscription $2.25/hr, AMD MI300X $29k, Groq LPU $18k. Power draws sourced from vendor datasheets. Maintenance assumed at 12% of capex annually unless cloud subscription bundles support.