Dynatrace already grades its customers’ AI. Its AI Observability product runs a second model on live traffic to assess whether the customer’s own model answered well, and that capability ships today. In June, the company open-sourced a command-line version of the same scoring, called dt-evals. Teams point it at their traces, the step-by-step records of what an AI application did on each request, and quality scores come back. Nine weeks later, Dynatrace agreed to pay $915 million for Arize. Building the evaluator was clearly not the hard part. Reaching the developers was.
Phoenix Is What the Money Actually Bought
Arize ships two products, and the free one is why this deal happened. Phoenix is the free one. Its code is public, and anyone can run it. Developers use it to record what their AI application did, score the results, and test prompt changes against a fixed set of examples. AX is the paid platform that does the same work at production scale. Both are built on OpenTelemetry, which is open and controlled by no single vendor. It sets one common format for recording what software does so that any tool can read the result. Arize wrote OpenInference, which extends OpenTelemetry to AI.
Phoenix clears two million downloads a month and eleven thousand stars on GitHub. A developer picks a recording library in week one of a project, before there is anything to monitor and usually before operations knows the project exists. Whatever they pick sets the data format and the quality checks from then on. Almost nobody rips it out later. Dynatrace can build a better evaluator whenever it likes. It cannot get the people already running Phoenix to switch.
The founders matter for the same reason. Jason Lopatecki keeps running Arize inside Dynatrace and reports to the CEO, and cofounder Aparna Dhinakaran joins with him.
Dynatrace built its business on the opposite habit. OneAgent installs once and discovers the environment on its own, so no developer ever has to pick anything. AI evaluation does not work that way. The developer wires up the application by hand and chooses the library that records it, and that conversation was happening without Dynatrace in it.
The Deal Covers the Months Before an Application Has Any Traffic
Dynatrace has always been strongest where traffic already exists. Davis AI works backward from a failure to determine its cause on a running system. Grail stores the logs, metrics, and traces that pour off live software. Even dt-evals needs AI traffic already flowing in Dynatrace before it can grade anything. All of it starts after something reaches production.
Arize also scores production traffic, and it covers the months before any of that traffic exists. A developer runs an agent against a set of test questions and reads back every step the agent took. Then they edit the prompt and run it again to see whether the answers improve. None of that generates a single real request. That work still settles the choices that decide what production monitoring can see later.
A developer building an agent today writes the quality check twice. The first version runs on a laptop and evaluates a fixed set of examples, scoring whether the answers stay true to the documents the agent was given. The second runs in production against live traffic, in a different tool, scoring the same thing. The check ends up in two places and the two versions drift. Dynatrace could not sell a single check that follows the application from the laptop to production. Now it can.
Evaluation on Both Ends Still Leans on One Model Grading Another
dt-evals and Arize both score quality by having a large language model judge another model’s output. The whole category works that way today, which puts an AI in charge of catching another AI misbehaving. Those judges drift when the provider updates the underlying model, and they disagree with human reviewers more often than teams expect. Every judgment also costs real money, since each one is another paid call to a model. So most teams score a sample of their traffic and leave the rest unchecked.
Smaller models would take most of that cost out. One model answers a single narrow question, whether an answer stuck to its sources or whether it turned abusive. It runs for a fraction of what a frontier model charges for the same judgment. At that price, nobody has to sample. Scoring runs on everything, the way authentication already does. I would like to see Dynatrace build them. The company has the production traffic to justify that work and the infrastructure people to do it.
A Steep Price for a Choice That Developers Make Only Once
Dynatrace was not missing AI observability, and it is not buying from weakness. It posted a strong quarter eight days before this announcement. Recurring revenue reached $2.14 billion, up 17%, and the organic part of new recurring revenue rose 41%. What it lacked was any presence on the machine where AI applications get written. Presence like that does not arrive with a feature release, because developers are choosing their tools right now.
The price reflects that urgency. Dynatrace expects the deal to add about two points to its recurring revenue growth this year. That puts the acquired revenue somewhere in the tens of millions. Against $915 million, the multiple is steep, and I am not going to call a price fair on revenue nobody has disclosed. The deal also costs Dynatrace about 1.75 points of operating margin this year. The company expects that margin to start growing again after that. A company that takes a margin hit on purpose is buying position rather than revenue.
A big observability vendor absorbing an evaluation company looks like the category folding in, which is the opposite of what I’ve been arguing since the spring. And I’ll keep arguing it isn’t folding in. Not yet anyway. A model that confidently invents a citation is nothing like a latency spike, and the tools built to catch the second one never see the first. The engineer who notices the invented citation is usually not the person renewing the monitoring contract, either. Agent evaluation remains separate for at least another two years. Dynatrace bought in while there was still something separate to buy.
Protecting that purchase starts with Phoenix’s license, which I would watch most closely. Phoenix ships under the Elastic License, which the Open Source Initiative does not recognize as open source. That sits awkwardly alongside the open-standards strategy Dynatrace has built through Bindplane and OpenTelemetry. The following stays only as long as the public releases keep coming.
Right now, a developer installs Phoenix with a single command and never speaks to anyone at Arize. Nobody sold it to them, which is exactly why they picked it. Dynatrace sells to large enterprises, where trying the software usually starts with a form and a sales call. If that process lands in front of Phoenix, the developers stop showing up, and they were the whole point of the $915 million. I would leave the install exactly as it is.
