How does this compute model calculate hardware TCO and inference cost?

It models compute hardware amortization, colocation rack power, PUE cooling efficiency, network egress, and token throughput per second across modern LLM and AI workloads.

How does on-premise hardware compare to hourly cloud rentals?

Dedicated compute clusters generally reach cost parity and breakeven against hyperscaler on-demand instances within 10 to 14 months at sustained 65%+ utilization.

Can I customize the power consumption and electricity rate?

Yes, power draw per node (kW), facility PUE, and local utility electricity rates ($/kWh) can be customized directly in the input fields.

⚡ Enterprise AI Infrastructure Modeling

NVIDIA H100 Cluster TCO & Token Inference Solver

Model 3-year server capital expenditures, InfiniBand switching fabrics, data center power & cooling PUE, and unit economics per 1,000,000 generated tokens compared to cloud hyperscalers.

Cluster Specifications & Energy

H100 SXM5 GPU Count

Server Capex / 8-GPU Node ($)

InfiniBand / Networking ($/Node)

Data Center Power ($/kWh)

Cooling PUE Multiplier

Avg Tokens / Sec / GPU

Cloud Benchmark Rate ($/GPU/Hour) Industry standard cloud on-demand/reserved range: $2.40 - $3.85/GPU/hr (AWS, GCP, Azure, Lambda).

3-Year Underwriting Summary

Total On-Prem 36-Month TCO (Capex + Power + Facility)

Cost per 1M Tokens: $0.00

Equivalent Cloud Cost (3 Yrs): $0

Net 3-Year Capital Savings: $0

Cluster Peak Load & Cooling: 0 kW

Monthly Power Bill: $0/mo