Opportunity

What can potentially be improved?

Filter replacement, Coil cleaning, CHWS optimization, Equipment yard, roof reflectance — every plant item conditional on an inferred as-found state.

How much physical capacity?

6,099kWModeledstrongest single intervention, facility peak

How much energy?

28.9MkWh/yrModeled

How much compute?

7,006kW ITInferred≈ 5,495 accelerator-equivalents campus-wide; worth $0 as modeled, up to $16.7M/yr marginal if power-constrained

How much environmental impact?

Pending Impact lensescarbon waits on the environmental layer; water is modeled

What does it cost?

$334kModeledclass cost prior (assumed)

What is the confidence?

Bands, not scalars: coherent favorable / unfavorable worlds on every number (open any figure).no mechanical fault observed; BAS trends would move rows to observed

What is observed vs assumed?

Observed roof reflectance, thermal, weather, prices Assumed IT load, facility cap, as-found states, costs

The whole-system model puts the recoverable opportunity in heat rejection and air distribution, not the roof. But every row is conditional: if the plant is in the assumed as-found state (fouled coils, loaded filters, scaled fill, conservative setpoint), then the modeled recoverable opportunity is the figure shown. No mechanical fault has been observed at this campus.

Bands (open any number) are the engine's coherent favorable/unfavorable worlds. Costs are class priors from the knowledge layer (ASSUMED). BAS trends — condenser approach, filter ΔP, tower approach — are what would move these rows from modeled to observed.

Recovered facility power becomes IT capacity becomes accelerator-equivalents. Whether that capacity is worth anything depends on one thing the model cannot see: whether the facility is actually power- or cooling-constrained. So the chain is shown twice — as modeled under the facility-cap prior, and as the engine's sensitivity with the facility constrained every hour.

Combined interventions and interaction value

Joint counterfactuals versus the sum of the individual ones. Negative interaction value means the interventions partially overlap (a cleaner condenser leaves less for the fill to recover).

Roof heat is a small term in a data-center heat balance, so restoring reflectance recovers little facility power. Its value is different: the roof state is observed from orbit, a wash test is cheap, and the result is visible within weeks — it validates the observation → model chain that the mechanical rows depend on. Dirt causation remains a hypothesis until that test.

Per roof — engine counterfactuals (R-30, $0.05/sf)

Observed albedo is the engine's 60-day window (Sentinel-2 v2, qa pass). Targets: peer median (observed), own best 30 days (observed upper bound), manufacturer age-adjusted prior (via an ASSUMED S2→SR scale).

If every inferred as-found degradation state exists at once, this is the facility overhead they cost relative to expected-at-age, and what that overhead would be worth as IT capacity. Constrained-hours fraction is 0 as modeled, so the compute lens is a sensitivity, not a realized value.

SIL = D − ΣA: the combined penalty of all degraded states minus the sum of their individual penalties. Positive means degradations compound.