Provisioned and Observed Power

Two different numbers describe the power of the same set of systems, and Dynamic Power Software (DPS) depends on the difference between them:

• Provisioned power is what you have committed. It is the sum of the power limits currently in force across the systems in scope — the draw those systems are permitted to reach.

• Observed power is what they are actually drawing, measured through telemetry.

Observed power is almost always lower than provisioned power. The gap between them is the headroom that conventional capacity planning has to leave stranded, and closing it safely is what DPS does.

Why the gap exists

Power limits are set for the worst case. A node provisioned at its nameplate TDP is permitted to draw that much, but a real workload rarely holds every component at its ceiling for a sustained period. Inference traffic in particular is bursty: average draw sits well below peak, and the peaks of different nodes do not coincide.

Size a deployment by adding up nameplate limits and you provision for a simultaneous worst case that never arrives. The unused difference is real capacity — power your facility has paid for, that no workload is using.

Why it matters

Three quantities bound a deployment:

Quantity	What it describes
Capacity	The physical limit of the power distribution equipment
Provisioned power	The total of the limits you have committed to systems
Observed power	What those systems actually draw

Static sizing must keep provisioned power below capacity, because nothing prevents systems from reaching their limits at once. That forces provisioned power down, and with it the amount of compute you can install.

DPS enforces limits dynamically and reacts when draw approaches a constraint. That changes which quantity has to stay under capacity: observed power does, rather than provisioned power. You can provision closer to the envelope — install more compute in the same facility — because the system will hold actual draw within the constraint rather than relying on a static sum to guarantee it.

This is why a DPS deployment can appear to be “over-provisioned” on paper and still be operating correctly. Provisioned power above what naive sizing would allow is the intended outcome, not a fault, as long as observed power stays within the binding constraint.

Reading the two numbers

Two ratios are worth watching, and they answer different questions.

Provisioned-load coverage — provisioned power as a percentage of physical capacity — tells you how much of the equipment’s capability you have committed.

Observed-versus-provisioned pressure — observed power as a percentage of provisioned power — tells you how much of what you committed is being used. A value above 100 means measured draw has exceeded the provisioned load.

A persistently large gap between the two under load is a signal, not a comfort. It can mean the fleet is under-utilized, that limits are binding unevenly across nodes, or that a workload is not exercising the hardware the way you assumed.

Where each number comes from

Number	Metric
Physical capacity	topology_service_device_power_capacity_watts
Provisioned load	topology_service_device_instance_power_load_watts
Observed power	device_total_power_watts

Refer to Telemetry and Observability for the queries that build the coverage and pressure ratios, and for the aggregation caveat that matters here: query a PowerDomain directly rather than summing parent and child device types, or topology aggregation counts the same power more than once.

DPS Sample Dashboards covers the Grafana dashboards that plot observed-versus-provisioned deviation over time.

Measuring the difference in a pilot

Run the MaxLPS Power Pilot uses these quantities as the basis of its comparison, and defines the derived figures a pilot reports:

• Capacity efficiency — the power budget minus provisioned power, the headroom between the envelope and what you committed on paper.

• Provisioning efficiency — provisioned power minus observed power, the stranded headroom that does not appear as draw.

A successful pilot shows capacity efficiency shrinking relative to conservative planning — you deliberately provisioned closer to the envelope — while compliance stays clean, meaning observed power never exceeded the binding constraint.