AI Datacenter Infrastructure Simulator Physical MW ($IREN) Software Stack ($NBIS)

Dynamic engineering tradeoff analysis: Physical power & liquid cooling scale vs Software orchestration & inference efficiency

Presets:
Model A: Physical Infrastructure First ($IREN)
Substation & Thermal Heavy
Grid Substation Capacity 50 MW
Raw physical utility power delivery at high-voltage substation site.
Power Usage Effectiveness (PUE) 1.12
Total facility power divided by IT equipment power (1.12 = direct liquid cooled).
Cooling Technology & Delta T
Liquid cooling prevents thermal throttling under dense GPU rack loads (up to 120kW/rack).
Orchestration & Virtualization 82%
Base cluster scheduling efficiency and active GPU job slot utilization.
Inference Quantization Gain 1.00x
Model weight compression (FP16 standard vs INT8/INT4 software optimization).
Model B: Cloud Software Stack First ($NBIS)
Orchestration & Virtualization Heavy
Grid Substation Capacity 50 MW
Physical grid connection limit for cloud datacenter deployment.
Power Usage Effectiveness (PUE) 1.32
Standard air/mixed cooling footprint in leased compute facilities.
Cooling Technology & Delta T
Lower facility capex compensated by aggressive software layer performance gains.
Orchestration & Virtualization 96%
Advanced multi-tenant vGPU slicing, zero-idle packing & batch scheduler.
Inference Quantization Gain 1.45x
Custom vLLM / FlashAttention-3 kernels, FP8 dynamic quantization yield boost.
Global Electricity Cost ($/kWh) $0.065 / kWh
Determines baseline annual utility power expenditure for both campus sites.
D3 Power-to-Compute Flow Stream
Substation Utility MW → Cooling/Transformer Overhead → Active GPU Power → Effective Token Throughput
Effective Compute Output
44.6 EFLOPS
52.1 EFLOPS
Model A ($IREN) Model B ($NBIS)
Cost per 1M LLM Tokens
$0.0084
$0.0072
$IREN TCO/Token $NBIS TCO/Token
Infrastructure Efficiency Ratio
73.2%
72.7%
Grid MW converted to productive compute FLOPs after thermal & software losses.
Thermal Throttling Risk Index
Model A: 3.2% (LOW)
Model B: 14.8% (MED)
Risk of GPU frequency downclocking during sustained peak compute workloads.
Comprehensive Architectural Comparison Telemetry
Telemetry Parameter Physical Infra Heavy ($IREN) Cloud Software Heavy ($NBIS) Variance / Edge
Raw Substation Capacity 50.0 MW 50.0 MW Parity
Facility Power Overhead (PUE Loss) 5.36 MW (PUE 1.12) 12.12 MW (PUE 1.32) $IREN +6.76 MW Saved
Net IT Power Delivered to Racks 44.64 MW 37.88 MW $IREN +17.8% Power
Effective GPU Compute Efficiency 0.820 (82% Slot Occ.) 1.392 (vGPU + Quant) $NBIS +69.7% Yield
Annual Utility Power Bill $28.47M / yr $28.47M / yr Equal Input Cost
Annual Token Output Yield 3,389 B Tokens 3,954 B Tokens $NBIS +16.7% Tokens
Unit Cost per Effective MW Delivered $0.638M / MW $0.752M / MW $IREN -15.1% Cost/MW
Calculations execute deterministically in client JavaScript runtime. No API dependencies.
SIMULATION ACTIVE: 50MW Benchmark | Model A: 44.6 EFLOPS | Model B: 52.1 EFLOPS
Ready

Physical MW vs Software Orchestration Tradeoffs

Read the explanation

AI datacenters convert raw high-voltage grid megawatts into compute throughput, balancing physical infrastructure efficiency against software orchestration. Direct liquid cooling in Model A minimizes thermal loss to five point three six megawatts, directly routing forty-four point six megawatts to GPU racks. In contrast, Model B with hybrid cooling wastes twelve point one two megawatts in thermal overhead, shrinking rack delivery to thirty-seven point nine megawatts. Yet software orchestration and FP8 quantization multiply Model B's yield by nearly one point four, propelling compute to fifty-two point one exaflops.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.