Datadog vs Grafana Cloud vs New Relic: what a Kubernetes estate costs

Three platforms, three billing units

Datadog publishes units rather than one rate. Infrastructure Pro is $15 per host a month on annual billing, custom metrics $5 per 100, logs $0.10 per GB with indexed events from $1.06 per million, and APM Pro $35 per host plus $1.70 per million indexed spans. Seats are extra.

Grafana Cloud bills by signal. Pro costs $19 a month with 10,000 active series included, then $6.50 per 1,000; the enterprise rate falls to $3 per 1,000 on a $25,000 annual commit. Application Observability adds host hours, and a host is an operating system instance, not a container.

New Relic gives away 100 GB of ingest a month and unlimited basic users, then charges $0.40 per GB for the original data option or $0.60 for Data Plus, with $0.05 per GB a month extra for the EU data centre. Full platform users start at $10 for the first and $99 for each additional on Standard, capped at five, or $349 a user on the annual Pro rate.

Cardinality is what produces the invoice

kube-state-metrics reads Kubernetes object state from the API server and republishes it as metrics: pod phase, replica counts, restarts, requests and limits, tagged with labels copied from the objects. Every distinct combination of label values is its own series, and all three vendors bill the series rather than the samples. One metric per pod across 300 pods is 300 series; add namespace, node, container, image and owner labels and the count becomes multiplicative.

Prometheus instrumentation guidance is blunt: label values that vary per request, user or pod are the ones to keep out. The fix is deletion before ingestion. Drop labels nobody queries, leave identifying detail in logs or trace attributes, and pre-aggregate the rest with recording rules. A cheap first pass is to ask which series anyone has queried in the last 90 days; your own query history answers it.

The worked example: 50 nodes and 300 pods

Take 50 nodes, 300 pods, 20,000 active series, 100 GB of logs a month and five engineers who need full platform access. On Datadog's annual rates the infrastructure line is 50 hosts at $15, or $750; 20,000 custom metrics are 200 units at $5, or $1,000; logs are $10; and APM on the same hosts is $1,750. The total is $3,510 a month before seats.

On New Relic the 100 GB sits inside the free allowance and five full platform users on the annual Pro rate are $1,745 a month. At 150 GB the bill is $1,765. The Standard edition prices five users at $10 plus four at $99, or $406, but caps at five users.

Grafana's metrics bill for 20,000 series is $19 plus 10 units at $6.50, or $84 a month. Host hours are separate: 50 hosts running continuously accumulate 36,000 in a 30-day month, and customers who started before 13 February 2026 get 1.5 active series and 0.02 GB of traces per host hour, or 54,000 series and 720 GB at that scale.

The totals are not like-for-like: Datadog includes APM and log indexing, New Relic includes five full users, and Grafana's figure is metrics only. Build the comparison from your own telemetry.

Logs and traces: retention is the lever

Log volume tracks how chatty the workloads are, not how many nodes exist. Ingest is the floor at every vendor, and Datadog charges indexing and retention separately, so a log stored but not searchable costs a fraction of one that is indexed. New Relic folds retention into the per-GB rate and sells Data Plus at $0.60 a GB for longer retention.

Thirty days hot on everything is rent on data nobody reads. A short indexed window with a cheap long-term tier often moves the number more than any filter. The cheapest gigabytes are at the source: namespaces nobody on call reads, debug lines in production, health-check chatter from a sidecar, all removed by a collector filter before they are metered. Sampling belongs there too: head sampling keeps the bill predictable, tail sampling keeps errors and slow requests but needs a collector tier that buffers spans.

OpenTelemetry is the exit route

Instrument with OpenTelemetry SDKs and route telemetry through the collector, and the backend becomes a configuration target rather than an architecture. The collector is a receiver, processor and exporter chain, so one pipeline can fan the same telemetry into two backends during a migration. Datadog publishes an OpenTelemetry integration path, and Grafana Cloud's ingestion is Prometheus-compatible.

The question for a vendor call is what still emits if the agent is removed tomorrow. Standard OTLP metrics, logs and traces reach any OTLP endpoint; proprietary span attributes, agent-side tag enrichment and dashboards built on vendor-only metric names do not travel, and rebuilding them is the real switching cost. Self-hosting Prometheus, Loki and Tempo pushes the licence towards zero and the cost into storage, retention, upgrades and the rota. For a three-person platform team that is usually the wrong trade.

The decision rule

Choose Datadog when a single agent, one interface and a per-host line finance can forecast matter more than the ceiling, and the team will enforce label discipline. Choose Grafana Cloud when engineers already write PromQL, active-series pricing matches the telemetry and the open-source fallback has value. Choose New Relic when per-gigabyte ingest and per-user seats match how the organisation buys.

Price the commitment against measured telemetry, not list rates, and get a written answer on what a doubling of pods does to it. Most platform teams end up with metrics in one place and logs in another, because the pricing models reward different shapes of data. A 12-month term quoted against your own numbers is the cheapest diligence on offer.