Calculators

VRAM establishes only a memory floor. Enter a replica degree qualified for the named model, runtime, hardware, context and batch profile; supported degrees need not be powers of two. Purchase quantum is a separate SKU or contract input and is not inferred from rack density. Facility design power follows installed units, while rental cost follows the active fractional fleet. At long context the KV cache can dominate VRAM → Ch 9.7.

This is a preliminary envelope screen, not equipment qualification. Rack kW is one input alongside component heat flux, the named OEM heat split, airflow/inlet limits, supported TCS/FWS or entering-water conditions, room residual rejection, climate, serviceability, redundancy, and the refresh tail. Use product capacity only at its stated air/water temperatures, flow, pressure, fan state, containment, and heat-capture target → Ch 5.4.

Benchmarks: JLL global-average shell/core ≈ $11.3M/MW (50 MW single-tenant, air-cooled; liquid +10%); Epoch's 1 GW model runs facility+land+utility ~$11.8M/MW and network ~$4.9M/MW, with servers ~$21.2M/MW on top → ~$38B/GW up-front. Electrical dominates the facility split → Ch 2.5, Ch 1.8.

Kill gates come first: power or interconnect at ≤2, or water or land/zoning at ≤1, disqualifies a site no matter how the weighted score averages out. Weights reflect the reordered 2026 hierarchy — power availability now leads, ahead of latency and land. Tune to your workload (training tolerates latency; inference doesn't) → Ch 3.13.

Project topology requires named maintenance and fault states, the exact affected path, transfer interruption, post-event loading, path and control independence, common-mode dependencies, a recovery SLO, and workload and contract consequences. These two choices are useful requirements, but they cannot select N, N+1, distributed-redundant, 2N, a Tier, or a capex premium. Compare and price candidate designs only after that project evidence closes → Ch 12.1.

MFU and goodput compound: a 40% MFU × 90% goodput scenario yields 36% of peak theoretical FLOPs as committed training work under the declared definitions. Treat 90% versus 96% only as an illustrative sensitivity; measure the named fleet, job, window, and event accounting, then attribute badput before assigning any gap to checkpointing or cordon policy → Ch 12.2.

Ownership beats the rental rate only above the breakeven utilization computed from your own inputs — below it a debt-financed cluster bleeds cash → Ch 1.8.

Compare ownership and rental on the same throughput and productive-use denominator. Owned node cost is the calendar-hour allocation of capex, energy, and opex; rented node cost is the capacity held for that hour. Market self-serve fell ~$10 → ~$2.50 / M tokens in a year (~4×), so underwrite inference with a price-decline curve → Ch 1.8.

PUE bands: legacy air 1.4–1.6 · direct-to-chip liquid 1.05–1.15 → Ch 15.1.

A real CFADS pro-forma: capex draws over the build with capitalized IDC, a revenue ramp, cash taxes net of the straight-line depreciation and interest shields, and sustaining capex. DSCR is CFADS ÷ debt service — the lender's ratio, stricter than EBITDA coverage; screen against ~1.3–1.5× contracted, ~1.75–2× merchant. Contracted offtake supports more leverage than merchant. Working-capital swings and NOL carryforwards are not modeled → Ch 2.5.