Kubernetes Adoption Statistics 2026: CNCF Survey, Cost and GPU Data

This is a reference page. Every figure below comes from a named, publicly available source, with the publisher and year stated next to it. None of it is our own data. Where a report’s detail sat behind a download gate and we could not read it directly, we left it out rather than repeat a number second-hand. Where a source is more than a year old, we say so next to the figure.

Figures last verified: 6 September 2026.

The short version: adoption stopped being the interesting question some time ago, and efficiency became the interesting question. The 2026 datasets show a platform that has won on adoption while getting measurably worse at using the hardware it runs on.

Kubernetes adoption statistics

• 82% of container users run Kubernetes in production, per The CNCF Annual Cloud Native Survey: The Infrastructure of AI’s Future (Cloud Native Computing Foundation, published 20 January 2026).
• Nearly 20 million developers are adopting and scaling cloud native technologies, per State of Cloud Native Development Q1 2026 (CNCF, published 24 March 2026).
• 88% of backend developers now work in standardised DevOps and platform environments (CNCF, 2026). Platform engineering stopped being a trend and became the default working environment for backend work.
• More than 230 projects and over 300,000 contributors hosted by CNCF, per the CNCF Annual Report 2025 (CNCF, published 24 February 2026, covering 2025).

Cluster utilisation and overprovisioning

CAST AI’s benchmark is the most useful public dataset here because it measures telemetry rather than opinions. The 2026 edition covers tens of thousands of Kubernetes clusters across AWS, Azure and GCP for the period 1 January to 31 December 2025, with GPU data extended through April 2026.

• Average CPU utilisation fell to 8% in 2025, down from 10% the prior year, per the 2026 State of Kubernetes Optimization Report (CAST AI, 2026).
• Memory utilisation declined from 23% to 20% over the same period (CAST AI, 2026).
• CPU overprovisioning rose from 40% to 69% year over year, meaning the gap between what is provisioned and what is even requested widened sharply (CAST AI, 2026).
• Memory overprovisioning stands at 79% (CAST AI, 2026).
• Approximately 50% reduction in provisioned CPU footprint for organisations that apply automated rightsizing (CAST AI, 2026). This is the largest single-lever improvement in any of the datasets on this page.
• 9% of the total CPU fleet now runs on ARM nodes, up from a near-negligible share a year earlier, with ARM growing roughly 3.5x faster than x86 since Q2 2024 (CAST AI, 2026).
• Around 7 to 8% of deployments are stateful workloads, roughly flat year over year (CAST AI, 2026). The “nobody runs databases on Kubernetes” claim and the “everybody does now” claim are both wrong.

Two things are worth saying about the utilisation numbers. First, low average utilisation is not automatically waste: burst capacity, HA headroom and node-level fragmentation all consume real capacity for real reasons. Second, a drop from 10% to 8% alongside overprovisioning rising from 40% to 69% is not headroom. That is drift.

GPU and AI workloads on Kubernetes

GPU efficiency is where the money is in 2026, and it is the worst-measured part of most clusters.

• 5% average GPU utilisation across the analysed clusters (CAST AI, 2026).
• Only 2% of GPUs run on Spot Instances, with spot adoption for GPUs described as negligible through most of 2025 (CAST AI, 2026). On the most expensive compute in the fleet, the cheapest purchasing option is almost entirely unused.
• 98% of FinOps practitioners now manage AI spend, up from 31% two years ago, per the State of FinOps 2026 (FinOps Foundation, 2026, 1,192 respondents representing over USD 83 billion in annual cloud spend, the sixth annual survey since 2020).
• AI cost management is identified as the number one skillset teams need to develop, and “FinOps for AI” as the top forward-looking priority (FinOps Foundation, 2026).
• CAST AI’s 2025 Kubernetes GPU Trends and Cost Report analysed A100 and H100 instances across 66 cloud regions from January 2024 through September 2025 (CAST AI, 2025), which is the underlying dataset for regional GPU price variation claims.

The CNCF’s 2026 annual survey is subtitled The Infrastructure of AI’s Future and frames Kubernetes as the common operating layer for both cloud native applications and AI workloads. The specific percentage of clusters running AI or ML workloads sits inside the downloadable report rather than on the public page, so it is not quoted here.

Kubernetes cost statistics

• 83% of container costs are associated with idle resources, split 54% cluster idle and 29% workload idle, per the State of Cloud Costs analysis (Datadog; note the published findings are based on data collected between May 2023 and April 2024).
• About 35% of EC2 compute spend goes to running containers, up from 30% a year earlier, and about one quarter of organisations allocate more than 75% of their EC2 spend to containers (Datadog, data through April 2024).
• 14% of compute costs are spent on GPU instances, up from 10% previously (Datadog, data through April 2024).

We are flagging the Datadog data period deliberately. It is the most detailed public breakdown of container idle cost that we could read directly, and it is also more than a year old. Treat the split between cluster idle and workload idle as structurally informative and the absolute percentages as dated.

Kubernetes security posture

The most recent readable survey on Kubernetes security posture is Red Hat’s, based on 600 DevOps, engineering and security professionals. It is a 2024 edition and we are labelling it as such.

• 67% of respondents have delayed or slowed down deployment of container-based applications due to security concerns, per the State of Kubernetes Security report (Red Hat, 2024).
• 46% experienced revenue or customer loss as a result of a security incident (Red Hat, 2024).
• 40% detected misconfigurations in their container or Kubernetes environments (Red Hat, 2024).
• 30% were fined as a result of an incident (Red Hat, 2024).
• 60% worry about vulnerabilities, misconfigurations and exposures in their container and Kubernetes environments, with vulnerabilities cited by 33% and misconfigurations by 27% (Red Hat, 2024).
• 90% have a DevSecOps initiative underway, comprising 48% in early stages and 42% in advanced stages (Red Hat, 2024).

The pattern that survives the report’s age is the one worth acting on: the dominant Kubernetes security cost is not breach cleanup, it is deployment friction. Two thirds of teams slowing releases because security posture is unclear is a platform engineering problem with a security label on it.

What we could not verify

Several figures that circulate widely could not be read from a primary source this session, so they are not on this page.

The CNCF annual survey’s detailed cross-tabs, including the percentage of respondents running AI or ML workloads on Kubernetes, the sample size and the fielding dates, sit inside the gated PDF. Only the published headline figure appears above.

Sysdig’s 2026 Cloud-Native Security and Usage Report, Fairwinds’ Kubernetes Benchmark Report and Spectro Cloud’s State of Production Kubernetes were all either gated or unreachable, so none of their numbers appear here despite being commonly quoted.

There is no current public figure we could verify for the average number of clusters per organisation, average cluster node count, or the share of Kubernetes estates running managed services versus self-managed. Numbers for all three circulate freely. We could not trace any of them to a readable primary source.

How to cite this page

kubernetes.ae, “Kubernetes Adoption Statistics 2026: CNCF Survey, Cost and GPU Data”, published 6 September 2026, https://kubernetes.ae/kubernetes-adoption-statistics-2026/. Figures compiled from the primary sources listed below and last verified 6 September 2026. Where you are citing an individual statistic, please cite the original publisher named alongside it rather than this page.

Methodology and sources

Every statistic on this page was read directly from the source listed below on 6 September 2026. Figures available only in search snippets, secondary summaries, or gated downloads were excluded rather than repeated. No figure is ours, estimated, or extrapolated. Where a source’s data period is materially older than its publication date, that period is stated inline.

If a source has since published a newer edition, or a figure here has been superseded, tell us and we will update the page and move the verification date.