… %!s(
Dynatrace was already shipping evaluation before this deal. Its AI Observability app traces gen_ai spans, scores live production responses with LLM-as-a-judge evaluators and detects drift in those scores over time. What it did not have is a position with the AI engineers who choose an evaluation harness. Those choices get made while an application is still being written, months before anything reaches an operations team.
The terms are roughly $815 million in cash plus replacement equity awards for Arize employees joining Dynatrace. The company plans to fund it from cash on hand or its existing credit facility. Co-founders Jason Lopatecki and Aparna Dhinakaran both join at closing, with Lopatecki continuing to lead the team and reporting to Rick McConnell, chief executive of Dynatrace. Closing is expected this quarter or early next, subject to regulatory review.
What Dynatrace Already Shipped
The existing product is more complete than the deal coverage suggests. In June, the company open-sourced dt-evals, a command-line tool that pulls recent gen_ai spans and scores them with an LLM judge. Results are written back as business events linked to the source trace. The documentation lists more than 10 built-in judge evaluators plus statistical drift detection against a rolling baseline of earlier scores.
That is a working production evaluation loop. It does not cover the half of the lifecycle that runs before an application has production traffic to score. Experiments, datasets, prompt iteration and pre-release evaluation sit on that side, and Arize is strongest there. The purchase buys lifecycle position rather than feature parity.
MORE FOR YOU
How Arize Scores An Agent Trajectory
Arize reaches developers through Phoenix, a self-hostable tracing and evaluation project, and enterprises through the commercial AX platform. A single agent run may involve a model request, a document retrieval, several tool calls and a final response. Each operation lands as a span within a single trace. An engineer can inspect the whole trajectory rather than the final answer alone.
Evaluators then attach themselves to that telemetry. An evaluator tests whether a response stayed grounded in the retrieved context, whether the correct tool was selected or whether the task was completed at all. It can be deterministic code, a human annotation or another model acting as a judge. The output is a score against a particular rubric and evaluator, which is not the same thing as a verdict on truth.
An architectural wrinkle sits underneath all of this. Phoenix uses OpenInference as its native semantic format rather than the OpenTelemetry conventions for generative AI. Traces arriving from other libraries get translated into OpenInference so they display consistently. Arize AX now normalizes compatible gen_ai attributes into OpenInference fields during ingestion, removing the need for a client-side conversion processor. Arize treats both conventions as first-class and expects them to converge as the OpenTelemetry specification stabilizes.
Datadog And Splunk Are Already There
Datadog traces LLM and agent applications, tracks token usage and cost and supports managed and custom LLM-as-a-judge evaluations attached to individual spans. Splunk has gone further than most observability buyers realize. Its AI Agent Monitoring runs platform-side and instrumentation-side evaluations covering hallucination, bias, relevance, sentiment and toxicity. The documentation says an agent gets flagged when fewer than 80% of evaluations pass for a metric. New Relic has its own AI monitoring across models, traces, cost and performance.
The decisive difference is where the tooling decision starts. Datadog, Splunk, and Dynatrace all sell into operations and platform engineering, and their evaluation features emerged as extensions of those relationships. Arize built from the opposite end, with Phoenix as a free local project that AI engineers adopt long before a procurement conversation exists.
By the time an application reaches production, the instrumentation library, the trace schema and the evaluator definitions have already been chosen. Dynatrace is paying to be in the room when that happens.
The Gaps
Dynatrace did not disclose Arize’s revenue, which makes the multiple impossible to calculate from public information. The company guided to roughly 200 basis points of accretion to ARR growth in the coming fiscal year. It also guided to a 175 basis-point dilution in the non-GAAP operating margin, with expansion expected the year after. Reverse-engineering an Arize ARR figure from that accretion guidance does not work. The number describes an effect on Dynatrace's own growth rate, including the timing of the deal.
The base is the useful comparison here. Dynatrace reported $2.14 billion in ARR and a 29% non-GAAP operating margin for the June quarter. Against that base, it is accepting a year of margin dilution and spending close to a billion dollars while the category is still forming.
The harder problem is that the evaluations are themselves probabilistic. When one model judges whether another has hallucinated or completed a task, the monitoring system has a second model embedded in its control loop. That evaluator can disagree with a human reviewer, and it can drift when its underlying model version changes. An evaluation score behaves like a sampled quality indicator rather than an HTTP status code. Running these in production calls for versioned evaluators, held-out test sets, periodic human calibration and an explicit threshold before a score blocks a deployment.
Arize describes Phoenix as open source, and the main repository ships under the Elastic License 2.0. That license permits broad use and self-hosting. It restricts anyone from offering the software itself as a hosted or managed service, and it is not approved by the Open Source Initiative. The distinction matters to exactly the developers whose trust Dynatrace is paying for.
The Enterprise Implication
The first question for a buyer is ownership of instrumentation. Establish whether the application emits OpenInference attributes, OpenTelemetry gen_ai attributes or vendor extensions, and where the translation happens. That answer sets the cost of a future platform change.
The second question is about evaluation economics, and the answer runs counter to intuition. Arize AX lists evaluations, experiments and human annotations as unlimited across its Free, Pro and Enterprise plans, metering span volume and ingested data instead. Running an LLM judge still costs model tokens paid to the provider, and tracing the evaluator's own execution consumes the same span allowance. Model a representative agent trajectory rather than pricing the system on request counts.
The third question is ownership of the quality signal. AI engineering may own the evaluator, platform engineering the trace pipeline, operations the incident and the business unit the definition of an acceptable outcome. Merging evaluation into an observability platform does not resolve those boundaries, though it does put them on one screen for the first time.
Dynatrace is taking a calculated risk on lifecycle position rather than on features. In a category where every incumbent already has the features, position is the right thing to buy. For enterprises, whether an AI system ran and whether it produced an acceptable result are now two separate operational questions. The vendors are finally competing on the second one. %!s(
