The Canyon ReviewThe observability bill, torn down line by line
Where Datadog, Grafana Cloud and self-hosted stacks actually charge you. Hosts, custom metrics, ingest and the cardinality trap, priced.
The observability bill is never one number. It's four: hosts, custom metrics, log ingest, and the retention you forgot you set. Vendors quote the first and grow the other three. This teardown prices a 20-host, 200 GB-of-logs-per-month shop across three stacks, list prices as of April 2026.
Verdict up front: Grafana Cloud for most teams this size, Datadog if someone else pays, self-hosted only with a person who wants to own it.
Where the money goes
| Line item | Datadog | Grafana Cloud | Self-hosted |
|---|---|---|---|
| Infrastructure/hosts | ~$15 to $23/host/mo | usage-based | your VMs |
| Custom metrics | billed per 100 series | included allotments, then usage | your disk |
| Log ingest | per GB, tiered | per GB past free tier | your disk, your grief |
| The surprise | cardinality | graduating off free tier | upgrade weekends |
The number that matters is custom metrics. A single histogram with a
user_id label can mint tens of thousands of billable series, and no
sales call will warn you. This is the cardinality trap, and every
per-series pricing model shares it. Before comparing vendors, run
your own count; both vendors' calculators accept it.
The 20-host math
Datadog: infrastructure plus APM on 20 hosts lands in the four figures monthly before logs, using nothing but the public pricing page. It is a genuinely excellent product and the only one here we'd call delightful. Delight is a line item.
Grafana Cloud: the free tier is real (10k series, 50 GB logs at the time of writing, their numbers), and our modeled shop lands in the low hundreds monthly. The catch is migration gravity: dashboards-as-code helps you arrive, nothing helps you leave.
Worth stating the floor of the market too: a brochure site has no observability bill at all. Something like a technology leadership portfolio needs an uptime ping and nothing else, and "do we actually emit anything worth $400 a month" is a fair opening question for smaller shops.
Self-hosted Prometheus, Loki and Grafana: the software is free and the correct comparison is a salary. Two 8 GB VMs (call it $80/month) run this shop's telemetry fine until the first retention crunch or Prometheus upgrade weekend. If nobody on the team is excited to own that, the "free" stack is the most expensive option on this page.
The skip
Skip per-host APM pricing for anything containerized and dense. Ten services per node bills like a monolith farm, and the per-host meter doesn't care that they're 50 MB Go binaries. If your platform is Kubernetes with high pod density, price everything usage-based or you will fund a vendor's sales kickoff personally.
Prices checked April 2026 against public pages; when they move, this entry moves.