EngiMetric
Medium-voltage switchgear, transformer and cable trays in an electrical plant room

Stage 2 of 14 / Calculation

Model the AI workload and rack capacity

Turn a heterogeneous rack schedule into installed nameplate, operating average, operating peak, concurrent peak, annual IT energy and growth headroom.

  • Advanced
  • 14 stages
  • 20 min
  • Mission-critical electrical and thermal systems
Development preview AI Data Center: Full Engineering Workbench
Development preview. The calculations on this journey run on existing EngiMetric engines, but the journey text and scope have not completed independent engineering review. The route is excluded from search while it remains a preview.
Professional review required.Engines and workbench are implemented and tested. Defaults, sources and engineering assumptions still require independent review before publication or use on a real project. Current verification state: Needs engineering review.

After this stage you can

Learning objectives
  • Build rack groups with density, count and average/peak load fractions.
  • Calculate nameplate, average, peak, concurrent peak, annual energy and growth headroom.
  • Identify whether the target capacity is limited by installed racks, concurrent load or future growth.

The engineering

AI racks are not a single density. A training hall may run at 50 kW per rack while an inference hall runs at 12 kW and a network row at 8 kW. The model keeps a separate group for each density and applies an average and a peak load fraction to the provisioned power.

Installed nameplate is not operating demand. A 50 kW rack provisioned for a future GPU generation may draw 35 kW average today and 47 kW at peak. The workbench requires both fractions and rejects a peak below the average. Concurrency is applied to the summed peak, not to each rack independently.

Annual IT energy comes from the operating average multiplied by 8,760 hours. Growth is a compound scalar over the stated horizon. It is a planning allowance, not a deployment schedule, and it must be compared with the owner target before electrical and cooling capacity are fixed.

EngiMetric / AI data-center workbench

One project, computed from rack to report

Loading the saved project state...

Preview screen. Inputs are saved in this browser only. Climate, grid, water and cost values are illustrative. Failed checks stay visible; a clean summary is not a stamped design and no Tier rating is claimed.

Your deliverable for this stage

Capstone evidence

A rack schedule and IT capacity statement with explicit nameplate, average and peak values.

Stage assumptions, limits and sources

Provenance

Assumptions

  • All racks in a group share the same provisioned power and load fractions.
  • Load fractions are assumptions, not measured telemetry.
  • Annual IT energy uses a flat 8,760-hour average unless the energy stage applies its hourly shape.

Constraints

  • The rack schedule is a capacity model, not a bill of materials for servers.

Limitations

  • No GPU workload trace, scheduler, power capping or DVFS model.
  • No measured PDU branch-circuit or rack PDU data.
  • Growth is not a phased deployment schedule.

Review status

  • Needs engineering review
  • Professional review is required before publication or use as a design basis.

Stage progress is stored in this browser only. The explanation and every limit on this page remain readable with JavaScript disabled; the calculators need it to run.

Stage kind: Calculation / Journey: ai-data-center / Stage: it-capacity