Datadog Is Becoming the Eyes of the Agentic Stack, and the Selloff Missed It

Background

Datadog (DDOG) is a cloud observability and security platform that monetizes through usage-based consumption across logs, metrics, traces, database monitoring, and a growing AI product suite. In Q2 2026 the company reported $1.1B in revenue, up 35.6% year over year, with gross margin of 78.6% and GAAP net income of $44.6M, alongside the launch of a remote MCP server that lets third-party AI agents pull live telemetry directly into their workflows. Despite the reacceleration, the stock sold off to recent lows on concerns about valuation and revenue concentration in a handful of AI-native customers.

Key Findings

The Incident That Explains the Next Decade of Ops

This week, a team pushed a badly tuned query to production and Postgres CPU pinned at 100%. No human opened a dashboard. Their coding agents called PlanetScale's MCP server, which exposes Insights data including a tool built specifically to surface queries sorted by CPU usage, and identified the offending pattern. The agents then called Datadog's MCP server to pull CPU time-series for the host, traces from the calling service, and logs around the spike, correlated the regression with the deploy, and proposed the fix. The entire investigation ran through two MCP endpoints, with Datadog supplying the ground truth about what was actually happening in production.

This is not a demo. Datadog's remote MCP server ships tools like get_logs, list_spans, and get_trace, plus Database Monitoring tools for query metrics and samples, and Agent Observability tools for LLM traces. PlanetScale explicitly markets its MCP tools so "your AI assistant can quickly find which queries are driving CPU spikes." The prevailing assumption is that observability is a human interface, screens for engineers. The finding is that observability just became a machine interface, an API that agents must call to operate infrastructure. That changes who the customer is and how often the product gets used.

The Quarter Was Strong and the Market Sold It Anyway

The prevailing narrative after Q2 was that Datadog is expensive and its AI-lab revenue is lumpy. The numbers tell a different story. Revenue hit $1.1B, up 35.6% year over year, a clear reacceleration from the 29.2% growth posted in Q4 2025. Gross margin held at 78.6%. On a non-GAAP basis, per company reporting, EPS came in at $0.65 with a 23% operating margin, free cash flow was $279M, and management raised the full-year guide to $4.45B to $4.47B. Growth plus profitability put the company at roughly a Rule of 40 score of 61.

Yet the stock sold off to recent lows. The bear logic centered on two things: concentration in a handful of AI-native customers, with OpenAI the name investors keep asking about, and guidance conservatism relative to the print. Both concerns are real, and both are the wrong frame. A company reaccelerating to 36% growth at this scale, with expanding cash generation, is not being punished for its results. It is being punished because the market has not repriced what the results represent: agentic workloads moving into production and pulling Datadog consumption with them.

MCP Tool Calls Grew 22x, and That Is the Leading Indicator

The single most important disclosure in the quarter got the least attention. Tool calls against Datadog's MCP server grew 4x quarter over quarter in Q2, and 22x versus Q4 2025. Third-party agents, including Claude Code, Codex, Cursor, and Factory's Droid, along with Hermes connectors, now natively pull Datadog telemetry as part of their standard investigation workflows. When an agent needs to know why latency spiked, whether a deploy caused an error rate change, or which query is burning CPU, it calls Datadog.

The assumption being challenged here is that AI agents threaten seat-based software by reducing the number of humans who need licenses. Datadog does not sell seats. It sells consumption. An agent investigating an incident does not run one query the way a tired on-call engineer does; it runs dozens, testing hypotheses in parallel, pulling traces, correlating logs, checking feature flags. Agent-driven query multiplication is a direct revenue tailwind under usage-based pricing, and 22x growth in two quarters is what the front edge of that curve looks like.

Datadog Is Running Both Sides of the Agentic Trade

Datadog is not just the data feed for other companies' agents. Its Bits AI portfolio ships first-party agents, an SRE Agent, a Dev Agent, and a Security Analyst, that investigate alerts, test hypotheses, isolate root cause, and propose fixes inside the platform. That is "AI for Datadog." The complementary side is "Datadog for AI": LLM observability, agent observability, GPU monitoring, AI Guard, and Agent Console, the tooling customers need to run their own AI systems in production.

Adoption data says this is real demand, not roadmap theater. LLM observability framework adoption roughly doubled year over year, from about 9% of organizations in early 2025 to about 18% by early 2026. That doubling is the clearest available proxy for AI workloads graduating from prototypes to production, and every production AI workload is a new surface Datadog monitors and a new consumer of Datadog telemetry.

The Bear Case Is Real, and the Data Layer Wins Anyway

Two bear arguments deserve a straight answer. First, revenue concentration: a handful of AI-native customers and AI labs drive lumpy consumption, and a single renegotiation or workload migration could dent a quarter. True, and it is why guidance stays conservative. But the MCP data shows the demand base broadening beneath the labs, into every team whose coding agents touch production.

Second, the structural risk: if agents own the workflow and the customer relationship, observability data could be commoditized into a context feed that someone else's agent monetizes, and Bits AI has to compete against every horizontal coding agent that can already read Datadog through MCP. This is the strongest version of the bear case, and the counter is positional. Whoever's agent acts, it still needs Datadog's telemetry to act correctly. The company collecting, indexing, and serving the ground truth about production systems gets paid on every query regardless of which agent asks. Per-seat competitors get squeezed as agents replace human operators; a consumption-priced data layer gets amplified. Datadog does not need Bits AI to beat Claude Code. It needs Claude Code to keep calling get_trace.

Implications

Datadog

The strategic position is picks-and-shovels for autonomous operations, and the Q2 reacceleration to 35.6% growth confirms the model is compounding rather than decaying. The priority is to keep the MCP surface the deepest and most complete telemetry API available to agents, because every third-party agent that standardizes on Datadog's tools deepens the moat. Concentration in AI-native accounts remains the execution risk to manage, but the broadening MCP usage base is diluting it each quarter.

AI Coding Agent Vendors (Anthropic, OpenAI, Cursor, Factory)

These vendors are simultaneously Datadog's largest new distribution channel and its most credible long-run competitive threat. Their agents already treat Datadog as the default source of production truth, which accelerates Datadog consumption today. If any of them attempts to own the full incident-response workflow, they still depend on Datadog's data, which caps how much value they can extract without it.

Incumbent and Seat-Based Observability Competitors

Vendors monetizing through per-seat licensing or thin dashboard layers face structural pressure. Agents reduce the human headcount that seat models bill against, while multiplying the query volume that consumption models bill for. Competitors without a first-class MCP surface risk being invisible to the agents that increasingly run operations.

PlanetScale and the MCP Ecosystem

The PlanetScale incident shows the emerging pattern: specialized MCP servers for domain data, composed with Datadog for cross-service correlation. Infrastructure vendors that expose their telemetry through MCP become nodes in agent workflows; those that do not get routed around. Datadog benefits from every node added, because it sits at the correlation layer where the investigation resolves.

Conclusion

The findings converge on one conclusion: the market sold a 36% growth reacceleration because it is valuing Datadog as a dashboard business with concentrated AI-lab revenue, while the evidence, 22x MCP tool call growth, doubled LLM observability adoption, and agents already resolving production incidents through Datadog's API, describes something structurally different. Observability has become the interface through which software runs itself, and Datadog is the ground truth every production agent must query. The bear case about agent-layer commoditization fails on the same fact that powers the bull case: whoever's agent acts, it pays to see. The selloff is not a verdict on the quarter; it is a mispricing of the shift from human-operated to agent-operated infrastructure, and Datadog is the consumption-priced toll road running through the middle of it.