Planning
GPU Utilization and Data Center Occupancy: Different Denominators
GPU utilization and data-center occupancy describe different things. A GPU activity metric concerns a device over a sample period; cabinet occupancy concerns space under an operator’s reporting definition. Neither percentage alone establishes model throughput, saleable compute or available facility power. Keep the metric, denominator, scope and time window beside every number.
Identify what the GPU counter measures
NVIDIA’s current NVML utilization structure defines its GPU field as the share of a sample period with at least one kernel executing. The memory field measures the share of that period with device-memory reads or writes. These are activity-time measures, not directly a model’s tokens per second or a percentage of theoretical FLOPS.
NVML separately documents memory byte fields. In the cited v1 structure, used memory includes reserved and allocated device memory. A memory activity percentage and a used-bytes figure therefore answer different questions. Preserve the exact field, API/version and device scope when reporting either.
Read the operator’s occupancy definition
Equinix’s 2025 annual report defines cabinet utilization in terms of cabinet space billed versus total cabinet capacity, taking power limitations into account. That is a dated example of a facility operator’s reporting measure. It does not measure whether a tenant’s GPU kernels are busy, and other operators can report occupancy differently.
When a presentation says occupancy, ask whether it means billed cabinets, leased area, contracted power or another quantity. Keep portfolio and site figures separate. A portfolio statistic cannot establish spare space, power or compute for one tenant at one location.
References: [E2]
Work through four percentages and quantities
Imagine a fictional reporting pack containing the following observations. Each can be valid at the same time because it describes a different population or denominator.
Scroll the table horizontally on smaller screens.
| Observation | Calculation or measurement basis | Question it cannot settle | Source |
|---|---|---|---|
| 80% cabinet occupancy | 80 billed cabinet equivalents / 100 available equivalents under the stated example definition | How busy are the GPUs in those cabinets? | Planning analysis |
| 20% GPU activity | One device executing kernels during 20% of its stated sample period | What fraction of peak model throughput was delivered? | Planning analysis |
| 70% memory activity | Device-memory reads/writes during 70% of that sample period | How many memory bytes are allocated? | Planning analysis |
| 60 of 80 example memory units used | Used-byte count / total-byte count = 75% | Was memory traffic active during 75% of the interval? | Planning analysis |
Fictional observations, with simplified units. No actual GPU, facility performance, customer result or recommended operating target is represented.
Keep aggregation visible
A fleet average needs a defined population: all owned devices, only reporting devices or only the devices assigned to one job. State the interval and how samples were combined. Without that information, a comparison can change because the population changed rather than because the work became more efficient.
For example, an unweighted mean of 10% and 90% is 50%. That arithmetic alone says nothing about unequal device capabilities, different sample durations or missing observations. Retain the underlying scope and aggregation rule before using the average in a planning discussion. Missing telemetry should remain missing rather than quietly becoming zero.
Ask a narrower question when a number looks low
A low sampled activity figure can justify a question to the platform owner; it does not establish the cause or prescribe a configuration change. Request the metric definition, measurement window, device support and relevant workload observations. Throughput, latency and completed work need their own definitions and acceptance criteria.
Likewise, a facility occupancy percentage cannot tell you which cabinets have suitable unallocated capacity. Ask the operator for the actual offered allocation and applicable constraints. Do not turn unused activity time into a claim that equipment, power or a customer service is immediately available.
Use each measure in its own handoff
A platform acceptance record can name workload-specific observations and the team that accepts them. A facility allocation can name space and service boundaries. A physical installation handoff can identify equipment, cable/label records and agreed defects. Keep those records connected without substituting one for another.
Leviathan can discuss physical installation and documentary handoff within the agreed site scope. It does not claim GPU performance tuning, hosted compute capacity or facility-capacity certification. If the requirement is physical work, describe the equipment and installation evidence needed; keep workload diagnosis with its qualified platform owner.
References: [L1]
Common planning questions
Is 100% sampled GPU activity proof of maximum useful performance?
No. Kernel activity time does not by itself establish model throughput, latency or useful completed work.
Can low cabinet occupancy prove GPU cloud availability?
No. The cabinet definition, actual allocation and compute-service inventory are separate evidence.
Primary sources
- [N1] NVIDIA NVML utilization structureChecked 2026-10-15. Current reference last updated September 9, 2026: sampled kernel-execution and device-memory-activity time. Not delivered FLOPS, model throughput or a universal monitoring configuration.
- [N2] NVIDIA NVML memory structureChecked 2026-10-15. v1 byte fields: used includes reserved and allocated device memory. Metric/device/API applicability must be confirmed; not a tuning instruction.
- [E2] Equinix 2025 annual report: cabinet utilization definitionChecked 2026-10-15. Historical reporting definition: billed cabinet space versus total capacity, taking power limitations into account. No 2025 operating percentage is presented as current; other operators may define occupancy differently.
- [L1] Leviathan colocation installation scopeChecked 2026-10-15. Customer-side physical installation subject to agreed scope, facility rules and available resources. No hosting, GPU rental, cloud provisioning or facility capacity offer.