Is Datadog worth it for a 100–1,000 person engineering team?

The money, named

What you pay. Datadog publishes a list-price comparison table, and the numbers that matter for an org this size are these: Infrastructure Pro is $15 per host per month billed annually ($18 on-demand), Infrastructure Enterprise is $23 ($27 on-demand), and APM alongside Infrastructure is $31 per host per month billed annually ($36 on-demand). Multiply by your host count and twelve months:

Three things fall straight out of that table. Running on-demand instead of committing costs 20% on Infrastructure Pro and 16% on APM, for nothing. Taking Enterprise instead of Pro across 500 hosts is $48,000 a year. And putting APM on every host rather than only the ones on a user-facing request path is $37,200 a year for every 100 hosts you did not need to trace.

What it has to beat. There is no per-engineer break-even for an observability platform, so here is the one that is actually decidable:

Infra Pro + APM, 100 hosts        100 x $46 x 12  =  $55,200 per year
$55,200 / $90 per hour ( = $180,000 / 2,000)      =  613 engineer-hours per year

On 100 instrumented hosts, Datadog pays for itself if it saves your engineers about 613 hours a year — roughly twelve hours a week across the whole org, or one four-hour Sev-1 averted a month with change to spare. For an org of 100–1,000 engineers running production on someone else's cloud, that is not a close call, and anybody arguing the licence on principle is arguing the wrong line. The argument worth having is about what you meter, not whether you meter.

What it can cost you. The two meters that produce surprise invoices are logs and hosts you did not know you had. Logs are billed twice — once to ingest, again to index — and the rates are very different: ingest is $0.10 per GB, while standard 15-day indexing is $1.70 per million events per month on an annual commit and $2.55 on-demand. Ingest is the cheap half; indexing is where a log line becomes a budget line:

Ingest   100 GB/day x 365 days = 36,500 GB x $0.10             =   $3,650 / year
Index    300M events/month x 12 x $1.70 per million (annual)  =   $6,120 / year
Index    the same volume at the on-demand rate of $2.55       =   $9,180 / year

Substitute your own volumes — those are round numbers chosen to show the shape of the meter, not a claim about your estate. The shape is the point: ingesting everything is affordable, indexing everything is not, and the gap between those two sentences is the entire log-management decision.

The reason, at this size

1. The bill compounds, and Datadog publishes the rate.

You do not have to guess what happens to a Datadog contract over five years. Datadog reports a net dollar-based retention rate every quarter in its own supplemental financial information: it was about 120% for four consecutive quarters through the end of 2025, and in the low-120%s for the quarter ended 31 March 2026. That is the company's own published statement that the median existing customer's spend grows roughly a fifth every year.

Read it honestly: that is a cohort-wide average across a whole customer base, and it includes customers buying new products as well as customers whose usage grew. It is not a forecast for your account, and we are not presenting it as one. But it is the best published evidence available about what happens to a bill like yours if nobody is watching, and it is published by the vendor:

A $250,000 contract that follows that curve costs $982,480 more over five years than a flat $250,000 line would. Nobody approves that; it arrives one autoscaling group at a time. Which is why the useful thing to take from this table is not fear — it is that the compounding is your negotiating document. Print it, take it to the renewal, and ask what the committed rate looks like if you sign for the growth you can already see.

2. The host meter does not count what you think it counts.

Datadog's billing documentation is explicit that usage is measured on a 99th-percentile basis over the month, not on your average and not on your peak: its APM billing page states that the APM bill is calculated using the top 99th percentile of active agents sending traces every hour of the month, and its usage documentation reports host counts on the same 99th-percentile basis.

Datadog frames that as protection, and it genuinely is — the top 1% of hours is discarded, so a single scaling spike does not set your bill. But the corollary is what catches orgs this size: anything that runs for more than about seven hours in a month is, for billing purposes, permanent. A nightly batch fleet, a load-test environment somebody left up for a week, a blue/green deploy that doubles your host count for an afternoon each day — none of those are spikes at the 99th percentile. They are your host count.

So the host number to commit to is not the one in your architecture diagram and not the one your cloud bill implies. Pull the hourly concurrent host count from the last ninety days, take the 99th percentile of it yourself, and negotiate against that. This is a twenty-minute exercise that routinely moves a commit by double digits in percent, in either direction — and being under-committed is the expensive direction, because it prices the overage at on-demand rates.

3. Pro versus Enterprise is a per-host allocation decision, not a features decision.

The $8 per host per month between Infrastructure Pro and Infrastructure Enterprise mostly buys headroom rather than capability: Datadog's published plan comparison gives Pro 5 containers and 100 custom metrics per host, and Enterprise 10 containers and 200 custom metrics per host, alongside features like live process monitoring and machine-learning-based alerting.

That framing changes the conversation usefully. Custom metrics and container density are things you can measure before you buy, and both are the classic sources of an invoice line nobody recognises. A Kubernetes estate running twenty pods per node is not a Pro-versus-Enterprise question, it is a per-host allocation question with a specific answer. Count your containers per host and your custom metrics per host first; the tier then picks itself, and at 500 hosts the difference between guessing and counting is $48,000 a year.

The trap in the other direction is real too: buying Enterprise across the whole estate for a capability two teams need. If the machine-learning alerting is what you are after, ask whether it needs to be on all 500 hosts or on the forty that carry the revenue path.

4. Logs are two meters, and only one of them is cheap.

At $0.10 per GB, ingesting your logs into Datadog is close to free at any volume an org this size produces. Indexing them at $1.70 per million events is not, and the difference is entirely under your control: Datadog's log pipeline lets you ingest, process and archive everything while indexing only a subset, and the subset is the only part you pay the index rate for.

The policy that works is boring and takes an afternoon: index what you alert on and what you search during an incident; archive the rest. Debug-level logs from a service nobody has paged about in a year do not need to be searchable in fifteen milliseconds; they need to be retrievable in fifteen minutes, which is a different and much cheaper product. Write the exclusion filters before you onboard the second service, because retro-fitting them across forty services after the first big invoice is a quarter's work that nobody has budgeted.

5. The failure mode at this headcount is nobody owning the number.

At twenty engineers, one person knows what is instrumented. At five hundred, the platform team owns the contract and forty product teams own the hosts, and the incentive structure is exactly backwards: adding a monitor is free to the team that adds it and expensive to a budget they never see. Every observability overspend we have looked at is that sentence, not a bad procurement decision.

Which is why the controls below are about attribution rather than about restraint. You do not want fewer dashboards; you want each team to see the monthly cost of the ones they own. That is a tagging job and a recurring calendar invite, and it is worth more than any discount you will negotiate.

What to do instead of just buying it

Four controls. Together they are about a week of one platform engineer's time, and they are the difference between the list price above and the compounding curve above it.

1. Take the 99th-percentile host count yourself, before the call.

Ninety days of hourly concurrent host counts from your cloud provider, 99th percentile, per environment. That is the number Datadog's meter will produce, so it is the number to commit against and the number to sanity-check the first invoice with. Bring it to the negotiation rather than receiving it from it.

2. Write the indexing policy before the first log arrives.

One page: which services index at full fidelity, which sample, which are ingest-and-archive only, and who approves a change. Exclusion filters on day one. The rule of thumb that survives contact with an incident: if you would not put it in an alert or a runbook, it does not need an index.

3. Tag every host and every custom metric with an owning team, and publish the monthly split.

Not to punish anyone — to make cost visible to the only people who can change it. A single dashboard showing spend by team, reviewed monthly for fifteen minutes, is the highest-return meeting in this whole purchase. Teams that can see their line reduce it without being asked; teams that cannot, cannot.

4. Put the renewal date and the compounding table in the calendar now.

Ninety days before renewal, not ninety days after. Take the growth curve, your own usage trend, and a written list of the products you are actually paged by — and be genuinely willing to drop the ones you are not. A renewal you have prepared for is the only leverage that exists in a contract with no published discount schedule.