Kubernetes cost terms, defined

A

Abandoned workload

A workload that uses almost none of what it requests over a long window. KubeOn flags CPU at or below 2% of request over 14 days.

Amortized cost

Cost with upfront and recurring commitment fees (Savings Plans, reservations) spread across the hours they cover, so each day carries its share.

Anomaly

A day of spend well above its recent baseline. UnitVane compares each day with the median of the prior 14 days and requires both an absolute and a percentage increase.

C

Cached tokens

Prompt tokens a provider serves from its prompt cache at a lower price. UnitVane AI prices them separately from input and output tokens.

Chargeback

Billing internal teams or cost centers for the infrastructure they use, usually through the finance system.

Cost allocation

Assigning spend to the teams, cost centers, applications or labels responsible for it.

Cost per 1M tokens

A unit cost for language models: allocated model spend divided by tokens processed, times one million. UnitVane AI includes PTU and reservation cost in it.

CUR 2.0

The current format of the AWS Cost and Usage Report, delivered as a Data Export. It is the only AWS export with split cost allocation data.

D

DCGM

NVIDIA Data Center GPU Manager. Its Prometheus exporter publishes GPU utilization, GPU memory and, where profiling is enabled, streaming-multiprocessor activity per GPU.

Direct cost

The part of a node's cost charged to a specific pod, before shared and idle cost are applied.

E

Effective-dated mapping

An ownership mapping that records when it started and ended. A change applies from the date you choose, and days before it keep their allocation; to correct past days, date the fix back and re-run them, up to 92 days.

Efficiency

The share of paid-for capacity that was used. UnitVane weights each resource's efficiency by its cost.

F

Fallback

Sending a request to another target when the first returns a throttling error, a server error or times out.

FinOps

The practice of bringing financial accountability to variable cloud spend, so engineering, finance and product make cost trade-offs together.

FOCUS

The FinOps Open Cost and Usage Specification, a common schema for billing data across providers.

G

gp3

The current general-purpose SSD volume type on Amazon EBS. It costs less per GB-month than gp2 and includes 3,000 IOPS and 125 MiB/s at any size.

GPU burst

Attaching GPU nodes from other regions, clouds or your own data center to a cluster when its home region has no GPUs left. KubeOn uses only the locations a burst policy lists, and only for pods that opt in.

GPU optimizer

KubeOn's optional in-cluster component that shares, scales, bursts and switches off GPUs. It is installed separately from the read-only agent and acts only through the policies you enable, in Recommend, Approve or Automatic mode.

GPU time-slicing

Letting several pods take turns on one GPU through the NVIDIA device plugin. It works on any NVIDIA GPU but gives no memory or fault isolation between the pods.

H

Headroom

Capacity added on top of measured usage when sizing a request, node or volume. KubeOn adds 20% to requests and node groups, and sizes volumes to 1.5 times their peak.

I

Idle cost

The cost of capacity no workload used. Infrastructure idle is node capacity nobody requested; workload idle is requested capacity that went unused.

Idle GPU time

GPU capacity on a node that no pod requested, priced at the node's rate. It is a scheduling problem, separate from low utilization of GPUs pods do hold.

IRSA

IAM Roles for Service Accounts: lets a pod on EKS assume an IAM role through the cluster's OIDC provider, without static keys.

K

Karpenter

An open-source Kubernetes node autoscaler. Its NodePool resource defines which instance types a cluster may launch.

L

LLM gateway

A single endpoint in front of model providers that routes requests, applies keys, budgets and limits, and records each request's tokens and cost.

M

max(request, usage)

The allocation rule that charges each pod for the larger of what it reserved and what it used, per resource and per hour.

MIG

Multi-Instance GPU: partitions a supported NVIDIA GPU, such as A100 or H100, into isolated instances with their own memory and compute, in fixed profiles per GPU model.

N

Net unblended cost

The cost of a line item after discounts such as EDP and credits are applied, before amortization.

O

Orphaned volume

A cloud disk that still bills after the Kubernetes PersistentVolume that used it is gone.

Ownership coverage

The share of billed cost in scope that resolves to a team or cost center, including shared and idle cost once a sharing policy assigns it. A cost can be reconciled before it has an owner.

P

p95

The 95th percentile: the value usage stays at or below 95% of the time. Used for sizing because it covers busy hours without chasing rare spikes.

Pay-as-you-go

Model usage billed per token, with no commitment.

Provisioned throughput

Reserved model capacity billed by the hour or by term, such as Azure OpenAI PTUs, Bedrock model units or Vertex AI GSUs.

PTU

Provisioned Throughput Unit: Azure OpenAI capacity bought in fixed units, deployed in Global, Data Zone or Regional pools.

R

Rate card

A price list for capacity you own, such as cost per vCPU-hour or GiB-hour, used to price on-premises clusters.

Reconciliation

Checking that all billed cost in scope appears in the cost ledger, with an owner or listed as unallocated. KubeOn reports it per cluster as a share of billed cost, separately from ownership coverage.

Request

The CPU and memory a container asks the scheduler to reserve for it. Requests decide how many pods fit on a node.

Rightsizing

Adjusting requests, node sizes or volume sizes to match measured usage plus headroom.

Route

In the UnitVane AI gateway, a named list of targets and rules that apps call instead of a vendor model.

S

Shared cost

Cost of platform services many teams use, such as kube-system, monitoring and ingress controllers.

Showback

Reporting each team's costs to it without moving budget, as a step before or instead of chargeback.

Spillover

Sending traffic beyond a provisioned deployment's capacity to pay-as-you-go instead of throttling it.

Split cost allocation data

AWS's per-pod cost for EKS, included in CUR 2.0, computed from requests and usage.

T

Tokens per minute (TPM)

The rate of tokens a deployment or key may process each minute; used for capacity sizing and rate limits.

U

Unallocated cost

Billed cost that cannot be joined to a cluster resource or an owner. KubeOn lists it with resource IDs instead of hiding it in idle.