Why does a monitoring bill double when nothing about the infrastructure changed? Ask anyone who has run observability at scale and they will tell you the same story: the tool did not get more expensive, the way it counts did. Infrastructure monitoring pricing is rarely a simple per-server number. It is a bundle of hosts, custom metrics, log volume, retention windows, and seats, each with its own meter running.
That is why a straight list-price comparison between Datadog, Grafana Cloud, New Relic, CloudWatch, and self-hosted Prometheus misleads almost everyone. The numbers only mean something once you define the workload. This comparison breaks down how each model actually charges, where the hidden costs hide, and how to build a bill you can predict. (Cloud load balancing tools practical comparison 2026)
Why list prices for monitoring tools mislead
Vendors publish prices per host, per gigabyte, per million spans, or per seat. None of those units describe the same thing. A “host” in one product may include basic metrics and nothing else; in another, the same host costs more because application performance monitoring is bundled in whether you use it or not.
Datadog’s model is host-based plus per-feature. You pay for infrastructure hosts, then separately for APM hosts, log ingestion, and custom metrics beyond an included allowance. That structure is fine when you know your workload, but it makes a headline price comparison almost meaningless.
AWS CloudWatch is usage-based. It bills per metric, per API request, and per gigabyte of logs ingested and stored. There is no host concept at all. A quiet account can cost almost nothing; a chatty one with high-cardinality custom metrics can cost more than a dedicated observability platform.
Prometheus is free under the Apache 2.0 license, but the license is not the cost. The official documentation states plainly that Prometheus is not a long-term, durable, clustered storage system. Retention is local and bounded by disk. The moment you need durability, long retention, or high availability, you are paying for storage, compute, and people.
The lesson: compare total cost of ownership at your workload, not list prices. The same discipline applies across the stack, which is why a careful look at how SIEM pricing is explained and how it drives security monitoring cost is worth reading alongside any observability budget.
The five pricing models, side by side

| Tool | Primary billing unit | What drives the bill up | Best fit |
|---|---|---|---|
| Datadog | Per host, plus per feature | Custom metric cardinality, log volume, APM hosts | Teams wanting one integrated platform |
| Grafana Cloud | Usage tiers (metrics, logs, traces, users) | Active series, log GB, retention | Teams already using Grafana OSS |
| New Relic | Per user plus data ingested | Seats, data volume, full-fidelity retention | User-heavy teams with moderate data |
| AWS CloudWatch | Per metric, request, GB | Custom metrics, API calls, log storage | AWS-native workloads, light usage |
| Prometheus + Grafana OSS | Infrastructure you run | Storage, compute, engineering time | Teams with platform skills and predictable scale |
Read that table as a map of incentives, not a price list. Each model rewards a different kind of discipline. Host-based pricing rewards consolidation. Usage-based pricing rewards cardinality hygiene. Seat-based pricing rewards automation over headcount.
The hidden cost of “free” open source
Self-hosted Prometheus is the cheapest option on paper and sometimes the most expensive in practice. The software costs nothing. Everything around it does.
Storage is the first bill. Prometheus keeps time series on local disk with a bounded retention window. Extend retention and disk grows. Add high availability and you are running duplicate scrapers and duplicate storage. Add long-term retention and you need a remote backend, which means either operating one or paying for a managed service.
The second bill is engineering time. Someone has to size the servers, manage upgrades, tune retention, watch for cardinality explosions, and get paged when the monitoring system itself falls over. That person is not free, and monitoring systems fail at exactly the worst moments.
The third bill is the one nobody budgets for: cardinality. A single careless label, such as a user ID or a request path with parameters baked in, can multiply one metric into hundreds of thousands of time series. Prometheus will happily ingest them until memory and disk give out. Managed platforms charge for the same explosion, just more visibly.
Where open source genuinely wins is predictability. If your metric volume is stable and your team already runs infrastructure well, self-hosting can undercut any SaaS bill by a wide margin. The same logic shows up in cloud infrastructure tools that actually earn their place: the tool is only cheap if the operational burden lands on people who were already there.
Hybrid architectures: cheap local, paid long-term
The most cost-effective pattern in 2026 is not “open source or SaaS.” It is both.
Prometheus supports remote write, a specification that lets a local Prometheus ship samples to a remote backend. Version 1.0 of the remote-write spec landed in 2023, and most managed observability vendors now accept it. That single feature changes the economics: keep high-resolution data locally for fast alerting, and ship a downsampled or filtered copy to a managed store for long-term retention and cross-team querying.
Practical tips from teams that do this well:
- Keep local retention short, often 15 to 30 days, and let the remote store hold the long tail.
- Filter at the source. Drop debug-level metrics and noisy labels before remote write, not after.
- Use recording rules to pre-aggregate high-cardinality data into a handful of useful series.
- Set a cardinality budget per team and alert when a team exceeds it.
The same hybrid thinking applies to compute. Running functions for glue work instead of dedicated instances often beats both extremes, as a function-as-a-service tools practical comparison lays out.
OpenTelemetry changes the pricing equation
OpenTelemetry is a vendor-neutral standard, graduated within the cloud native ecosystem in 2023, covering metrics, logs, and traces through one set of APIs and SDKs. Its pricing significance is simple: it removes lock-in.
When instrumentation is vendor-specific, switching tools means re-instrumenting everything, so teams stay put and accept price increases. When instrumentation is OpenTelemetry-native, the backend becomes a commodity decision. You can send the same telemetry to a cheap store, an expensive one, or both, and change your mind next quarter.
That optionality has a real dollar value, even though it never appears on an invoice. It is the strongest negotiating position a platform team can hold.
A worked comparison framework
Since list prices go stale within months, build the comparison yourself with a repeatable method. This is what actually works in practice.
- Inventory your workload. Count hosts, active time series, log gigabytes per day, trace spans per month, and the number of people who need dashboards.
- Measure cardinality before you shop. Pull your top 50 metrics by series count. This one step predicts most bill surprises.
- Model three scenarios. Current volume, 2x growth, and a bad month where a deploy doubles your log output.
- Price the operational overhead. For self-hosted options, add storage, compute, and a realistic fraction of an engineer’s salary.
- Check the exit. Ask what it costs in engineering time to leave each vendor. That number belongs in the comparison.
A spreadsheet with those five inputs beats any vendor comparison page, because it uses your numbers instead of a marketing department’s.
Common mistakes that inflate the bill
The same errors appear across almost every organization that gets surprised by a monitoring invoice.
Ignoring custom metrics. Included allowances are small relative to real-world cardinality. Teams enable a new exporter, and a thousand new series appear overnight.
Logging everything at full fidelity. Logs are usually the largest line item. Most of that volume is debug output nobody reads. Sampling and structured logging cut costs without cutting insight.
Buying seats for people who only look at dashboards. Read-only viewers are often cheaper than full users, and many teams overprovision full seats out of habit.
Forgetting retention tiers. Hot storage for 30 days plus cold storage for a year costs a fraction of hot storage for a year. Vendors rarely lead with this.
Skipping the load balancer and network layer. Monitoring traffic is not free to move, and egress between regions adds up. A cloud load balancing tools practical comparison is a useful reminder that the plumbing has its own price tag.
Where AI-assisted monitoring fits the budget
AI features are now standard in monitoring platforms, and they carry their own pricing logic. Some vendors bundle anomaly detection into existing tiers; others charge per host or per query. The practical question is whether the feature replaces work a human was doing.
If anomaly detection catches a real incident a month, it pays for itself. If it generates noise that engineers mute, it is a line item with no return. The broader pattern, visible across the AI tools pricing guide covering what 50 top tools really cost, is that AI pricing is still unsettled and varies wildly between vendors for similar capability. Treat AI add-ons as a separate purchase decision, not a bundle default.
There is also a macro story here. The current wave of AI infrastructure spending echoes earlier build-outs that overshot, as an analysis of the AI infrastructure race and the 2000 telecom bust argues. Monitoring budgets ride the same cycle, which is one more reason to build cost visibility into your tooling now rather than during the next correction.
How to negotiate and right-size
Once you have your workload model, negotiation gets easier. Vendors price against alternatives, so your strongest lever is a credible plan to walk away.
- Ask for committed-use discounts in exchange for a multi-year term, but only after modeling growth. Overcommitting is worse than paying list.
- Push for custom-metric allowances rather than percentage discounts. The allowance is where the bill actually moves.
- Request a proof of concept with your real data. Vendors that will not ingest your cardinality are hiding something.
- Review quarterly. A monitoring bill is a living thing. What fit in January rarely fits in October.
Conclusion
Infrastructure monitoring pricing is not a number you look up. It is a function of your workload, your discipline, and your willingness to operate software yourself. Host-based platforms like Datadog reward consolidation. Usage-based systems like CloudWatch reward hygiene. Self-hosted Prometheus rewards engineering skill and punishes complacency. OpenTelemetry quietly shifts power back toward the buyer by making the backend replaceable.
The teams that win are not the ones that found the cheapest list price. They are the ones that measured their cardinality, modeled their growth, priced their own time honestly, and built a hybrid architecture that keeps fast local alerting while paying only for the long-term storage they actually query. Do that, and the invoice stops being a surprise.
FAQ
Is Prometheus really free?
The software is free under the Apache 2.0 license, but running it is not. You pay for storage, compute, and the engineering time to operate it. Prometheus’s own documentation notes it is not a long-term, durable, clustered storage system, so long retention and high availability require additional infrastructure or a managed backend.
Why is my Datadog bill higher than expected?
Usually custom metric cardinality. Datadog charges per host plus per feature, and custom metrics beyond the included allowance are a common surprise. A single high-cardinality label can multiply series counts dramatically. Audit your top metrics by series count before assuming the vendor raised prices.
Which is cheaper, CloudWatch or a dedicated monitoring platform?
It depends on volume and workload shape. CloudWatch bills per metric, per API request, and per gigabyte of logs, which is cheap for light, AWS-native usage and expensive at high cardinality or heavy log volume. Dedicated platforms often win at scale, while CloudWatch wins for small, simple accounts.
What is OpenTelemetry and why does it matter for pricing?
OpenTelemetry is a vendor-neutral standard for metrics, logs, and traces. It matters because it makes the backend replaceable. When instrumentation is not vendor-specific, you can move telemetry between providers or run a hybrid setup, which gives you real leverage in pricing conversations.
How do I reduce monitoring costs without losing visibility?
Start with cardinality hygiene: drop noisy labels and pre-aggregate with recording rules. Then sample logs and send only what you query to hot storage. Use retention tiers for older data, right-size user seats, and filter at the source before shipping data to a paid backend.
Should I self-host monitoring or buy a managed service?
Self-host if your metric volume is stable, your team already runs infrastructure well, and you can absorb the operational work. Buy managed if your cardinality is unpredictable, your team is small, or you need long retention and high availability without hiring for it. Many organizations do both, keeping local Prometheus for alerting and shipping to a managed store for the long tail.


