What reached general availability
Grafana Labs announced on August 4 that Adaptive Profiles has reached general availability, closing out the company's Adaptive Telemetry suite by extending intelligent cost optimization to the fourth and final observability signal: continuous profiling, alongside the metrics, logs, and traces capabilities Grafana had already shipped. The pitch, according to staff product manager Steven Dungan, is that the suite inverts the usual telemetry cost model: every signal now carries optimization that learns usage patterns and adjusts automatically rather than requiring engineers to hand-tune retention and sampling rules.
Continuous profiling shows engineers exactly which lines of code are consuming CPU and memory in production, which makes it one of the most useful debugging signals available and also one of the most expensive to run everywhere all the time. Adaptive Profiles is built to change that math by making fleet-wide profiling economically viable instead of something teams reserve for a handful of critical services. Grafana positions this as the last piece of a suite it has been building for roughly a year, meaning customers can now apply the same cost logic uniformly across every telemetry signal they collect, instead of optimizing metrics and logs while leaving profiling as an expensive exception reserved for post-incident deep dives.
How the system actually works
The mechanism is straightforward in concept: collect at a lower baseline resolution during normal operation, then automatically increase collection detail and frequency the moment an anomaly or performance regression appears. That means teams are not paying full ingestion cost around the clock for data they will only ever query during an incident. Grafana's framing is blunt about why this matters now: cost has historically scaled with ingestion volume, not with the insight actually extracted from that data, and that mismatch gets worse as workloads generate more telemetry by default.
That mismatch is precisely what AI agent and LLM deployments are making worse. Grafana's 2026 Observability Survey found that 57 percent of organizations now implement LLM observability and 65 percent name cost as a top factor when selecting tooling, evidence that the telemetry volume generated by autonomous agents and inference pipelines is outrunning teams' ability or willingness to pay for full-resolution collection everywhere. Agent workloads in particular tend to spawn many short-lived processes and nested calls, each one a candidate for profiling, which multiplies the volume problem in a way traditional monolithic services never did.
The numbers behind the pitch
Grafana backs the release with concrete customer figures rather than only architectural claims. Adaptive Metrics has eliminated 28.5 billion active metric series across the customer base, delivering an average 35 percent cost reduction, with video infrastructure company Mux reporting a 60 percent volume reduction. Adaptive Logs has removed 26 petabytes of unused log data, with healthcare logistics company TeleTracking citing a 50 percent volume cut. Adaptive Traces has reduced trace data volume by an average of 82 percent across customers using it.
Those numbers matter because they come from production deployments, not lab benchmarks. Michael Beltz, VP of cloud operations at Upland Software, said Adaptive Profiles lets his team use continuous profiling without worrying about cost overruns while still surfacing where improvements are needed. Kyle Weaver, a staff engineer at Mux, put a dollar figure on it: the optimization saves the company hundreds of thousands of dollars annually while forcing a deeper look at where metrics volume was unnecessary in the first place.
Why this lands differently than a typical pricing update
Observability vendors have spent years competing mostly on breadth of signal coverage and integration count. Adaptive Telemetry is a different kind of bet: that the next competitive axis is cost efficiency per unit of insight, not raw ingestion capacity. That is a meaningful shift for CTOs who have watched observability line items grow faster than the infrastructure they monitor, often becoming one of the largest recurring cloud costs on the books without a proportional increase in usable signal, to the point where some platform teams now review observability spend with the same rigor they apply to compute reservations, and push back hard on any renewal that cannot show volume declining relative to workload growth.
Geoff Schultz, manager of infrastructure engineering at Auditboard, described the before-and-after in stark terms: previously his team chose between budget overruns or insufficient visibility, and with adaptive optimization tracing became usable rather than something rationed to a handful of services. That is the tradeoff most platform teams have been quietly making for years without a good alternative, and it is the specific pain Grafana is targeting with this release. It is also a useful reminder for any vendor evaluation that the right question is rarely whether a tool collects enough data, but whether an engineer can actually afford to look at all of it when something breaks.
What this means for streaming and pipeline architecture
Adaptive Telemetry is itself a data engineering problem: deciding in real time what to keep, what to aggregate, and what to discard is a streaming and stateful processing challenge no different in kind from the tradeoffs teams make in any high-volume event pipeline. Grafana's approach treats telemetry as a data product with its own lifecycle management rather than a firehose that gets dumped into storage and queried later, which is the same architectural shift lakehouse and time-series platforms have been pushing for years.
That framing is useful beyond observability specifically. Any team building high-cardinality event pipelines, whether for security telemetry, IoT sensor data, or application logs, faces the identical tradeoff between collection fidelity and storage cost. Grafana's approach of dynamic, anomaly-aware sampling is a pattern data engineering teams should evaluate for their own pipelines even outside the observability stack, since the underlying cost problem is the same one: most of the value in a high-volume stream is concentrated in a small fraction of events, and paying to store the rest at full fidelity rarely earns its keep.
What CTOs should do with this
The immediate action for platform and data leaders is to audit current telemetry spend against actual usage: how much of what gets ingested is ever queried, alerted on, or referenced during an incident. Grafana's own numbers suggest the gap between ingestion and insight is large enough at most organizations to justify the engineering time to close it, and a majority of that gap tends to sit in exactly the high-cardinality metrics and verbose logging that adaptive sampling targets first.
For teams already on Grafana Cloud, evaluating Adaptive Profiles against current profiling costs is a low-risk next step given the customer results already reported. For teams on other observability stacks, the more durable lesson is architectural: telemetry pipelines should be built with cost-aware sampling as a default behavior, not an afterthought bolted on once a cloud bill becomes impossible to ignore, and that principle should extend to any team standing up new telemetry for agentic AI workloads over the next few quarters.
