Skip to content
GridNinja

Constraint model

Capacity is bound by the tightest cross-domain constraint

Usable AI data center capacity is bounded by coupled constraints, not a single electrical ceiling. Workload timing changes heat, cooling demand, storage posture, generation reserves, water exposure, network dependencies, and SLA risk. A cross-domain capacity process evaluates those interactions in the same operating window and accepts only the capacity whose binding constraint and margin can be shown.

Definition

What it is

A cross-domain constraint model represents the dependencies that can make an otherwise feasible power action unsafe, unavailable, or commercially unusable.

Category boundary

What it is not

It is not a collection of independent dashboards or a sum of best-case headroom values. Capacity cannot be counted twice across domains that depend on the same reserve or recovery path.

Operational significance

Why operators care

Operators see which constraint actually binds a request and can evaluate repairs—such as timing, placement, or asset commitment—without hiding the cost in another facility domain.

Mechanism

How the decision path works

  1. Step 01

    Normalize constraints into a shared time window and topology context.

  2. Step 02

    Propagate candidate actions across power, thermal, water, storage, workload, and SLA models.

  3. Step 03

    Identify the earliest binding constraint and margin to limit.

  4. Step 04

    Repair the action only when the changed action remains provable across every domain.

Textual diagram equivalent

Decision flow

Requested AI load ↔ workload placement ↔ heat and cooling ↔ power and storage ↔ bridge power and reserves ↔ SLA and recovery → one accepted or refused capacity result.

Synthetic coupled-constraint example

Visibly scoped example

Synthetic illustrative scenario—not a customer or production result. Electrical headroom exists, but a proposed workload placement is rejected because the thermal recovery assumption is unsupported after a cooling-component outage.

Failure and no-proof cases

  • Models use different timestamps, topology versions, or operating windows.
  • One reserve is credited to multiple independent capacity claims.
  • A domain reports a scalar score without the constraints needed for runtime checking.

Limitations

  • The model is only as complete as the declared dependencies and trusted telemetry.
  • Site engineering limits and protective settings remain authoritative.
  • Uncertain or unavailable domains should produce no-proof rather than inferred capacity.

Evidence and sources

Inspect the basis, not just the answer

Local evidence objects show the GridNinja proof contract. Primary sources provide external standards and risk-management context; they do not validate GridNinja performance.

Proof before autonomy

Test the evidence boundary against your capacity question

Start with a Capacity Audit or read-only Shadow Mode discussion. GridNinja does not require control authority to identify where a capacity claim remains unproven.