AI Data Center Per-MW IRR Dashboard v3

10Y horizon · per-year price paths · scenario save/load · GB300 NVL72 with BTM gas · View source on GitHub
Saved scenarios: | |

Preset scenarios

Capex (US$M / MW)

Revenue & Operations

Y1-Y10 Price Path (per-year YoY %)

Price change per year (vs prior year)

Effective Revenue/MW per year

GPU Refresh option

0 = no refresh; Y5-Y7 = mid-cycle Vera Rubin upgrade

Residual & Discount

Key metrics

Total Capex (incl Refresh)
$59M
Effective Y1 Revenue
$15M
Cumulative Revenue 10Y
$150M
10Y Unlevered IRR
10%
NPV at chosen rate
$0M
Payback (years)
4.2

Year-by-year cash flow (Y0-Y10)

Year$/GPU-hr equivRevenueOpexRefresh CXResidualNet CFCum CF

Sensitivity: Y1 Revenue × Avg Annual Decay → 10Y IRR

📋 Methodology + sources
Methodology: NPV = Σ CF_t / (1+r)^t for t=0..10. IRR via bisection. Y0 = -capex; Y_n = revenue × util × Π(1 + price_change_yr_i) × (1+managed) - opex. Refresh year: subtract refresh capex, then revenue path resets with boost. Y10 includes GPU residual + DC shell residual.

Sources: Crusoe slide (Stanford MS&E 435 Class #3, Spring 2026) — base capex/revenue assumptions; SemiAnalysis — gas plant cost benchmarks, GPU rental decay history; CoreWeave 10-K FY2025 (SEC EDGAR CIK 1769628, filed 2026-03-02) — 6-yr GPU life, 850 MW running, $5.1B revenue; 4-AI external review (Gemini 3.1 Pro, GPT-5.5, Grok 4.20, DeepSeek V4 Pro) — opex underestimation, residual realism, Vera Rubin transition.

Per-MW physical assumptions: 1 MW IT load ≈ 7-8 GB300 NVL72 racks ≈ 540 GPUs. Implied $/GPU-hr = revenue / 540 GPUs / 8760 hours.