What is Datadog net revenue retention in 2026?

What net revenue retention actually measures at a usage-priced company

Net revenue retention is the ratio of recurring revenue from a fixed cohort of customers at the end of a period to what that same cohort produced at the start, including upgrades, cross-sells, downgrades, and churn, but excluding revenue from any customer acquired during the period. That definition sounds mechanical until you apply it to a company like Datadog, where the majority of billing is consumption-driven rather than seat-driven. At a seat-based company — a CRM, a helpdesk, an HR suite — expansion is a discrete event. Somebody signs a contract amendment, adds forty licenses, and NRR moves in a single step you can point at on a calendar. At a usage-priced observability platform, expansion is continuous and largely invisible until the invoice arrives. A customer that deploys thirty new Kubernetes nodes, turns on distributed tracing for two more services, or starts routing a noisy application log stream into indexed retention has expanded without a single conversation with a salesperson.

That distinction matters enormously for how you read the number. Datadog's historical NRR strength was never primarily a sales achievement — it was a structural consequence of customers migrating workloads to cloud infrastructure during a decade when nobody was scrutinizing the monitoring line item. Hosts multiplied, containers multiplied faster, microservice architectures fragmented single applications into dozens of independently instrumented services, and the meter ran. The 130%+ era reflected an industry in expansion phase, and the platform captured that expansion automatically.

The corollary is uncomfortable and worth stating plainly: the same mechanism runs in reverse. When customers consolidate services, sample logs more aggressively, drop metric cardinality, shorten retention windows, or move a workload to a cheaper tier, the meter slows without anyone canceling anything. This is why observability NRR compressed faster and more broadly than NRR at seat-based SaaS during 2023 and 2024. Nobody churned. Everybody optimized. A 130% cohort became a 115% cohort while logo retention stayed near the high nineties.

For a RevOps practitioner, the operational lesson generalizes past Datadog entirely. If your company prices on consumption — API calls, data volume, compute minutes, transactions, messages, seats-that-behave-like-usage — your NRR is a leading indicator of your customers' internal budget posture, not just of your product's value. You will feel a spending-discipline cycle in your retention numbers one to two quarters before a seat-based peer feels it in renewals. That is a forecasting advantage if you instrument for it and a nasty surprise if you do not.

One more definitional point that trips people up in board decks. NRR is a cohort ratio, so its denominator is fixed at the start of the measurement window. A company can post a decelerating NRR while growing total revenue faster than ever, because new logos are not in the ratio. Conversely, a company can post flattering NRR while total growth stalls, because the retained base expands while acquisition dries up. Reading either number alone tells you very little. Read them as a pair, always, and read them against gross revenue retention — the same ratio with expansion stripped out — to separate "customers left" from "customers spent less."

Reading the historical arc and what it implies about the 2026 band

Datadog's reported gross and net retention history has a clear shape. Through 2018-2020, disclosed net revenue retention ran well above 130%, riding cloud migration and the multi-product attach motion that took the company from an infrastructure monitoring tool into APM, log management, and synthetics. The 2021-2022 stretch held in a similar band, boosted by pandemic-era digital acceleration and the microservices wave that structurally multiplied the number of billable components per application.

The inflection arrived in 2023. That was the first year enterprise buyers systematically audited cloud spend, and observability was an obvious target because it sits adjacent to the cloud bill, is easy to attribute, and often grew unmanaged. Datadog's disclosed retention fell through the 120s and settled into the low-to-mid teens above 100 by the back half of that year. Through 2024 the company described retention as in the low-120s to mid-110s band depending on the quarter and the disclosure convention. The pattern held into 2025 without a snap-back.

Extrapolating that to 2026 gives you a defensible 115-120% modeled range, and you should be explicit that it is modeled rather than reported. The arguments for the upper half of that band are concrete. AI application workloads generate telemetry at a rate that dwarfs conventional web services — prompt and completion traces, token accounting, retrieval latency, embedding pipeline health, model-cost attribution — and that telemetry lands in exactly the product surfaces Datadog has built out. A team that puts a retrieval-augmented assistant into production does not add ten percent to its monitoring footprint; it frequently adds a new spending category from scratch. Security attach is the second lever: cloud SIEM, application security management, posture management, and sensitive data scanning sell into an install base that already has the agent deployed and the data flowing, which is the cheapest cross-sell setup in enterprise software. Third, selective list-price movement on high-value modules has historically stuck without meaningful churn.

The arguments for the lower half are equally concrete. OpenTelemetry has moved from interesting to default in a lot of platform engineering organizations, and its practical effect is to decouple instrumentation from vendor. A customer that emits OTel data pays for ingest, storage, and the analysis layer, but has removed the switching friction that used to make expansion the path of least resistance. FinOps as a discipline has institutionalized cost review — it is no longer a panic response to a bad quarter, it is a standing function with headcount and quarterly targets. And the competitive field consolidated: the Splunk acquisition by Cisco closed in March 2024 for approximately $28 billion, hyperscaler-native monitoring keeps improving and is bundled with cloud commitments customers have already made, and specialist vendors compete hard on specific high-volume workloads where Datadog's pricing is most exposed.

Net those forces and 115-120% is the honest center of the distribution, with meaningful probability mass at 110-115% if AI monetization proves more substitutive than additive. What is not plausible on any reasonable model is a return to 140%. That number required a one-time infrastructure migration happening simultaneously across the entire economy under conditions of zero-cost capital. Those conditions are gone.

The step-by-step process for building your own NRR estimate

Do not take a vendor's headline retention number and drop it into a model. Build the estimate, because the construction is where the insight lives. The sequence below works whether you are an investor sizing Datadog, a competitor benchmarking against it, or a RevOps leader trying to produce a defensible number for your own board.

Start by fixing the cohort and the window. Pick every customer with recurring revenue as of the first day of the period and freeze that list. Anyone who signs after that date is excluded from both numerator and denominator, no exceptions. Normalize the revenue basis — annualized recurring revenue, or trailing-twelve-month recognized revenue, but one or the other consistently. Consumption businesses create a real trap here, because a customer with a spiky month can look like expansion when it is seasonality. Annualize on a trailing basis rather than multiplying a single month by twelve.

Second, decompose the delta rather than computing a single ratio. Split the change into four buckets: pure usage growth on existing products, new product attach, price and packaging changes, and contraction. Contraction should further separate optimization-driven reduction from partial churn, because those have different fixes. Usage growth is a market-condition signal, product attach is a go-to-market signal, price is a packaging signal, and contraction is a value-realization signal. A blended 117% built on 25% attach and negative 8% contraction is a completely different business from a blended 117% built on 5% attach and positive usage drift.

Third, segment before you conclude anything. Enterprise accounts behave nothing like mid-market accounts, which behave nothing like self-serve. In Datadog's case, the enterprise tier — the several thousand customers above $100,000 in annual spend, which the company discloses in its quarterly materials and which represents the large majority of revenue — carries the blended number. Mid-market is the volatility source: a scaling company can triple its footprint in eighteen months, and a company that hits a funding wall can halve it in two quarters. Self-serve and small business contribute little to blended NRR by revenue weight but dominate logo counts, which is why logo retention and revenue retention tell different stories at every company with a genuine self-serve motion.

Fourth, stress the estimate against two scenarios you actually believe. Not a symmetric plus-or-minus five percent, which is decoration. Pick a specific downside mechanism — say, a large financial services customer completing an OTel migration and cutting indexed log volume sixty percent — and a specific upside mechanism — say, two AI-heavy accounts moving pilot workloads into production. Size each in dollars, then see what they do to the ratio.

Fifth, and this is the step most teams skip, write down what would falsify your estimate. If you model 118% and the mechanism is security attach, then the observable that matters is products-per-customer, not revenue. If that metric stalls for two consecutive quarters while your NRR estimate holds, your model is wrong and you should know it early rather than defending it late.

Costs, timelines, and the ranges practitioners should expect

Retention improvement programs have costs and lead times that get systematically understated, and the numbers below are the ones worth arguing about in a planning cycle.

On timeline: a genuine NRR shift takes four to six quarters to appear in reported numbers, because the metric is a trailing cohort ratio. If you launch a cross-sell motion in the first quarter of a year, the accounts that attach a second product in the second quarter do not fully show up in a trailing-twelve-month retention figure until the following year. This lag is the single most common reason retention initiatives get killed prematurely. Executives fund a program, look at the metric two quarters later, see nothing, and reallocate. The program was working; the measurement window had not caught up.

On cost structure: expansion into an existing account typically runs at a fraction of the cost of new logo acquisition — the frequently cited range is somewhere between a quarter and a half of the fully loaded acquisition cost, though the real number depends heavily on how much technical implementation work the expansion requires. Land-and-expand in observability is cheap precisely because the agent is already deployed and the data is already flowing. Turning on a security module for an account that already ships infrastructure telemetry is a configuration change and a pricing conversation, not an implementation project. That asymmetry is why platform breadth is the dominant retention lever in this category and why single-product observability vendors face a structural ceiling.

On typical ranges for a large-cap infrastructure software company in the current environment: net revenue retention above 130% is now rare and usually indicates either an early-stage business with a small denominator or a company riding a specific one-time adoption wave. The 115-125% band is where strong consumption-priced platforms sit. The 105-115% band is normal for mature enterprise software with solid but not explosive expansion. Below 100% means the installed base is shrinking and new logo acquisition is carrying all growth, which is an expensive way to grow and usually shows up in worsening sales efficiency within a year.

Translate retention into growth arithmetic, because that is what it is actually for. A cohort retaining at 117% contributes roughly 17% growth before any new customer is signed. Layer new logo contribution on top — for a company at Datadog's scale, new customers typically add something in the range of eight to fifteen points annually — and total revenue growth lands in the mid-twenties to mid-thirties. That reconciliation is a useful sanity check in both directions: if someone hands you a retention estimate that does not reconcile to reported total growth net of new logo contribution, one of the two numbers is wrong.

On valuation sensitivity, be careful with precision you do not have. Directionally, forward revenue multiples in enterprise software correlate with retention because retention proxies for revenue durability and pricing power, and a company that slips from the high teens above 100 into the single digits above 100 tends to see multiple compression that exceeds the arithmetic impact on growth. But the specific multiple attached to a specific retention band moves with rates and sentiment, so treat any fixed mapping as a snapshot rather than a rule.

Finally, the internal cost of measuring this well. Building a defensible cohort-based retention model with segment decomposition typically requires cleaning up account hierarchies first, which at a company of any size is a multi-month data project involving parent-child mapping, subsidiary rollups, and reseller pass-through. Most retention numbers are wrong for boring reasons — a customer counted as two accounts because of an acquisition, or a division billed separately — and the fix is unglamorous data work rather than analytics sophistication.

Where teams get retention analysis wrong

The first mistake is treating retention as a customer success metric. It is not. It is a product-market-fit metric with a customer success component. When retention compresses because customers are optimizing their cloud bills across every vendor simultaneously, no amount of quarterly business reviews changes the outcome, and blaming the CS team for a macro spending cycle burns credibility and headcount at the exact moment you need both. The diagnostic question is whether contraction is concentrated in accounts with low product breadth and low feature adoption, or spread evenly across the base. Concentrated contraction is an execution problem you can fix. Even contraction is a market condition you have to price and plan around.

The second mistake is confusing logo retention with revenue retention and then designing the wrong intervention. An observability vendor can hold ninety-eight percent of its customers and still watch net revenue retention fall ten points, because the losses are volumetric rather than relational. Save-the-account playbooks do nothing here. What works is packaging: commitment tiers that trade a spend floor for a better rate, tiered retention and archival options that let customers keep data cheaply instead of dropping it, and cost-visibility tooling that lets the customer self-optimize inside your platform rather than optimizing by leaving.

Third, teams over-index on the blended number and under-invest in segment views. A blended 116% can conceal enterprise at 122% and mid-market at 98%, which is a completely different set of decisions than enterprise at 110% and mid-market at 130%. The first says your growth engine is upmarket and your mid-market motion is leaking; the second says the opposite. Same headline, opposite strategies. If your board deck shows one retention number, you are flying with one instrument.

Fourth — and this is the one that costs the most money — teams respond to compression by chasing the metric directly. They pull back from lower-ARR segments because those segments dilute retention. This is arithmetically effective and strategically expensive, because those accounts are the future enterprise pipeline. A scaling company entering at fifteen thousand dollars a year is the same company that becomes a two-hundred-thousand-dollar account four years later. Optimizing retention by refusing to serve the segment that produces your future retention is the SaaS equivalent of burning furniture for heat.

Fifth, forecasting models treat consumption revenue as if it were contracted revenue. In a usage-priced business, a meaningful share of billings is elastic month to month, and building a forecast that assumes last quarter's run rate persists will be systematically wrong in both directions. The fix is to model committed floor separately from variable overage, forecast them with different methods, and report the split. Commitments give you a defensible baseline; overage gives you the upside distribution. Blending them into one number destroys information at exactly the point where you need it most.

Sixth, analysts treat product-attach counts as a proxy for value delivered. Products per customer is a useful directional metric — moving an install base from roughly three products toward four is a real retention lever — but a customer that turned on a fourth product and never adopted it will not renew it. Track activation depth, not activation count. The distinction between "provisioned" and "in the daily workflow of an engineer" is the entire difference between durable attach and a one-year revenue bump followed by a contraction.

Decision framework: when retention compression demands which response

Not every retention decline calls for the same intervention, and the most common failure is applying an expensive fix to a cheap problem or vice versa. The framework below sorts the response by cause.

Begin by asking whether logo retention held. If customers are actually leaving, retention is a symptom and the underlying issue is competitive displacement or a value gap; expansion tactics are irrelevant until that stops. If logos held and revenue fell, you are in optimization territory, which is where most observability and infrastructure vendors have lived since 2023.

Within optimization, the next split is whether contraction concentrates or spreads. Concentration in a segment or a workload type points at a specific pricing or packaging exposure — an expensive product line, a high-volume data type, a tier boundary customers keep bumping against. That is fixable with packaging inside two quarters. Spread contraction across the entire base is a macro budget cycle, and the correct response is different: protect the commitment floor with multi-year structures, ship the cost-visibility tooling that makes you part of the customer's savings story rather than a target of it, and accept a lower retention plateau in the plan rather than promising a recovery you cannot engineer.

If retention is holding but attach is stalling, the problem is neither pricing nor macro — it is the cross-sell motion itself, and specifically whether product recommendations are triggered by observed usage signals or by a quarterly sales campaign. Usage-triggered attach converts far better because the need is already demonstrated in the telemetry.

Apply the same framework to adjacent categories and it holds. A data warehouse vendor facing query-cost optimization, a communications API platform facing message-volume tightening, a CDN facing egress renegotiation — all of them are running the same play against the same customer behavior. The specific meter differs; the structure of the problem does not. That is why the Datadog retention question is worth understanding in detail even if you never own a share or a contract: it is the cleanest public case study in what happens to consumption-priced revenue when customers get disciplined, and every RevOps team building a usage-priced motion will face a smaller version of it.

Related questions

Is 115-120% net revenue retention good?

Yes. For a company at multi-billion-dollar scale in enterprise infrastructure software, sustained retention in that band is strong and implies meaningful organic growth from the installed base alone. It sits well above the roughly 100-110% typical of mature enterprise software.

Why does NRR compress without customers churning?

Because consumption pricing lets customers reduce spend without canceling. Sampling logs, cutting metric cardinality, shortening retention windows, and consolidating services all lower the bill while logo retention stays intact. Revenue retention falls; gross logo retention does not move.

Does OpenTelemetry adoption hurt observability vendors?

It reduces instrumentation lock-in and shifts the value contest to storage, analysis, and workflow. Vendors that compete on the analysis layer adapt; vendors whose moat was proprietary agents face structural pressure on expansion economics rather than immediate revenue loss.

How do AI workloads change monitoring spend?

They add new telemetry categories — traces, token accounting, retrieval latency, model cost attribution — that did not exist in conventional application stacks. For teams putting AI features into production, this often creates a new spending line rather than shifting an existing one.

Where does Datadog disclose retention figures?

In quarterly earnings materials and annual filings published through its investor relations site. Definitions and disclosure conventions vary between companies, so read the stated methodology before comparing any two vendors' retention numbers directly.

FAQ

What is net revenue retention and how is it calculated?

Net revenue retention takes a fixed cohort of customers measured at the start of a period and compares their recurring revenue at the end of that period to the start, capturing expansion, downgrades, and churn while excluding all newly acquired customers. Expressed as a percentage, anything above 100% means the existing base grew on its own. The precise convention — annualized recurring revenue versus trailing recognized revenue, monthly versus annual cohorts — varies by company, which is why cross-vendor comparisons require reading each disclosure's methodology rather than lining up headline numbers.

Has Datadog published a 2026 net revenue retention number?

No. Retention figures are disclosed in quarterly earnings materials and annual filings, and a 2026 full-year figure would only be available after that fiscal year concludes and results are reported. Any 2026 number circulating now is a model, not a disclosure. The defensible framing is a range grounded in the reported trajectory — above 130% through 2022, compressing through the 120s in 2023, settling into roughly 110-115% across 2024 — which supports a modeled 115-120% for 2026 with real downside risk if AI monetization proves substitutive.

What drove the compression from 130%+ to the 110-115% range?

Cloud cost discipline, primarily. When capital got expensive, enterprises audited cloud spending, and observability was an obvious target because it grows automatically with infrastructure and sits adjacent to an already-scrutinized cloud bill. Customers sampled logs, trimmed metric cardinality, shortened retention, and consolidated services. Layer on OpenTelemetry reducing agent lock-in and hyperscaler-native monitoring bundled into existing cloud commitments, and the expansion that used to happen automatically now has to be earned.

Could AI workloads push retention back above 130%?

Very unlikely. AI telemetry is genuinely additive for teams shipping production AI features, and per-account expansion in those cases can be substantial. But the 130%+ era rested on an economy-wide cloud migration happening simultaneously under zero-cost capital, and AI adoption is more staged and more scrutinized. Realistically, AI-driven expansion offsets ongoing optimization pressure and holds retention in the high teens above 100 rather than restoring prior peaks.

How should a RevOps team apply this to its own consumption-priced product?

Model committed floor and variable overage separately, forecast them with different methods, and report the split rather than a blended run rate. Decompose retention into usage growth, product attach, price change, and contraction so you know which lever moved. Segment by revenue tier, because a blended figure routinely hides opposite trends in enterprise and mid-market. And measure activation depth, not just how many products a customer has provisioned.

How long does it take for a retention initiative to show up in reported numbers?

Four to six quarters, because the metric is a trailing cohort ratio. An expansion motion launched in the first quarter does not fully register in a trailing-twelve-month retention figure until well into the following year. Kill decisions made two quarters in are almost always premature. Track leading indicators instead — product attach rate, activation depth, commitment coverage — and let the retention figure confirm what those already told you.

Sources

• https://investors.datadoghq.com/
• https://www.datadoghq.com/product/llm-observability/
• https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=DDOG&type=10-K
• https://opentelemetry.io/
• https://www.cncf.io/projects/opentelemetry/
• https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2024/m03/cisco-completes-acquisition-of-splunk.html
• https://investors.snowflake.com/
• https://ir.crowdstrike.com/
• https://investor.mongodb.com/
• https://www.finops.org/

Related on PULSE

• Should I hire a fractional CRO if my net revenue retention is below 100 percent?
• Why is net revenue retention compressing and how do you fix it?
• How do you operationalize net revenue retention?
• What is NRR and what is a healthy benchmark?
• What is the net revenue retention benchmark for B2B SaaS?
• How do you structure RevOps OKRs that tie to net revenue retention?