OpenTelemetry's Maturing Ecosystem and Cloudflare's Switch to Per-Gigabyte Trace Billing: What Sources Show About Observability Costs

Background

OpenTelemetry (OTel) is the open-source framework for collecting traces, metrics and logs. According to The New Stack, it began in 2019 as the merger of two competing efforts, OpenTracing and OpenCensus, and the Cloud Native Computing Foundation (CNCF) announced its graduation, the foundation’s highest maturity level, in May 2026. The same report, published on May 21, 2026, says the CNCF describes OpenTelemetry as among its “highest-velocity” projects, second only to Kubernetes, with over 12,000 contributors and participation from more than 2,000 companies. The Machine Herald covered the milestone at the time in an earlier article.

The New Stack reports that observability vendors historically relied on proprietary agents, SDKs and telemetry pipelines that made it difficult to move between platforms, and that major cloud providers and observability vendors now support OpenTelemetry in various forms. It quotes Chris Aniszczyk, CTO at the CNCF, as saying that “Before OpenTelemetry, vendors competed on proprietary agents/collectors and data formats, which often led to customer lock-in,” and that today the competition is on “developer and user experience” and on “cost effective solutions,” among other things. The same article notes that adoption has had operational costs: it cites a 2025 proposal from the project’s Governance Committee acknowledging “complexity and lack of stability,” and it states that OpenTelemetry “doesn’t solve observability economics on its own.”

That framing, in which standard instrumentation moves the contest to cost and experience, is a claim by a foundation executive reported by one outlet. The developments below are a way to test it against more recent events.

Current situation

Interoperability is improving, but most teams run a mix

The New Stack reported on September 25, 2026 on a survey of Prometheus and OpenTelemetry interoperability, written up by OTel contributors. According to that report, nearly half of respondents mix Prometheus-style and OTel-style instrumentation for infrastructure metrics, and 30.7% use both for application metrics. The average ease-of-use rating for using the two together rose from 3.1 to 3.6, and the share who find them hard to use together fell from 29% to 10%. Respondents still asked for better alignment of data models, better handling of resource attributes and metadata, and fewer naming and formatting issues.

The same weekly column reports on a New Relic survey, the 2026 Observability Forecast, of 2,575 IT and engineering leaders and practitioners. According to the column, 73% are standardized on OTel, actively migrating to it, or testing it, and engineers report spending 37% of their time addressing disruptions. New Relic is an observability vendor, so this is a vendor-run survey rather than independent measurement, and The New Stack relays its figures without further verification.

A large user migrates for compatibility, and reports cost effects

In a post on the CNCF blog dated September 17, 2026, Atlassian engineers describe replacing the company’s gostatsd metrics pipeline, which handled metrics from roughly 100k hosts across 14 regions. Their stated reason was that the community had converged on OpenTelemetry and more of what fed the pipeline was emitting OTel data that gostatsd did not support. They kept the interface that service teams use and rebuilt the pipeline behind it with OTel Collector distributions.

The authors report several cost-related results. Folding metrics into the tracing sidecar saved about 3.9% CPU on average per service across what they call their priciest Micros services, which they put at roughly a 30% cut in sidecar cost at fleet scale. Their aggregation tier takes in about 4.8 billion datapoints a minute and lands about 220 million, roughly a 96% reduction, and now runs on about half the CPU for the same traffic. The authors write that the design lets them “drop wasteful datapoints at ingest, the cheapest place to do it.” Their next step, they say, is to move instrumentation off vendor and in-house clients, naming Datadog/DogStatsD and StatsD libraries. This is a first-party account from the company that did the migration, and the post is a self-report that no cited independent source has checked.

Cloudflare moves tracing from per-span to per-gigabyte billing

On October 2, 2026, Cloudflare announced Cloudflare Traces in open beta, which records supported steps of a request’s path through its platform as spans and can export them to any OTLP endpoint. The blog post describes cost as part of the design. It says teams can set a baseline sampling rate “to balance visibility, data volume, and cost,” gives 1% of requests as an example baseline, and describes Trace Rules that capture 100% of traffic for a specific hostname, source IP or header while leaving other traffic at the baseline.

Cloudflare’s post also states that, instead of charging by the number of spans or events, pricing “is based on how much observability data you ingest and how long you retain it,” effective December 1, 2026 across Cloudflare Tracing and Workers Tracing. The free plan includes 0.5 GB of ingestion per day with 7-day retention. Paid and Enterprise plans include 50 GB of ingestion and 10 GB-month of storage per billing cycle, with additional usage at $0.25 per GB ingested and $0.10 per GB-month stored; retention of up to one year is listed as coming soon. The Herald’s earlier article covers the launch itself.

InfoQ adds context that the vendor post does not. It notes that the change is the second pricing change for Workers Tracing in two months and that its August coverage had reported every span would become a billable event from October 1, 2026. InfoQ calls charging by volume rather than by span “the more consequential half” and writes that “Sampling decisions now map directly to a bill.” It also notes that coverage is partial by design, with broader automatic instrumentation still planned.

Implications

The following points are what follows from comparing the sources, not findings reported by any of them as a whole.

Cost control now operates at several layers. Atlassian’s savings come from the collection and aggregation tiers, where data is reduced before storage. Cloudflare’s design puts the lever at the sampling decision, and its billing unit is ingested volume. InfoQ’s reading connects the two: once the bill tracks volume, sampling and ingest-time reduction are the means by which a team controls it. For operators running their own pipelines, the Atlassian account suggests the savings available depend on where in the pipeline reduction happens, though it is one company’s experience at one scale.

Standardization has not meant switching. The survey figures reported by The New Stack show a hybrid estate rather than a migration from Prometheus to OTel: nearly half of respondents mix the two styles for infrastructure metrics. For teams, the practical consequence is running and reconciling two instrumentation conventions, with the remaining complaints being about data-model and naming differences.

The sources partly support the “competition moves up a layer” claim, and partly leave it open. Cloudflare exporting standard OTLP and accepting W3C trace context fits the portability story Aniszczyk describes. Cloudflare’s move to a pricing unit based on data volume is a change in how one vendor charges, and none of the cited sources compares it with other vendors’ pricing. The New Stack’s own caveat that OpenTelemetry does not on its own solve observability economics remains the more cautious reading.

Evidence quality is uneven. Of the sources above, the survey on interoperability was run by project contributors, the New Relic report by a vendor, the Atlassian post by the migrating company and the pricing details by Cloudflare. The independent reporting comes from The New Stack and InfoQ, and The New Stack’s graduation article is from May 2026, several months older than the other material. The New Stack column also discloses that HPE is a presenting sponsor of the series, though the observability items cited here are not about HPE.

What the sources do not establish

None of the cited material reports how many Cloudflare customers will see a higher or lower bill after December 1, 2026, and the one-year retention option is listed only as coming soon. The sources do not report whether other vendors are changing their billing units, and they do not independently verify Atlassian’s measurements.