DevOps Roadmap 2026: Skills, Certifications and What AI Changed

I get asked for a DevOps roadmap about once a week, usually by someone who has just watched an agent write a Terraform module in forty seconds and wants to know whether the job they are training for will still exist. I have spent fourteen years doing this work across AWS, GCP and Azure, and I now run PandaStack (my company, a Firecracker microVM cloud for AI agents) mostly on my own, so I have a fairly direct view of which skills an AI assistant replaces and which ones it leans on harder. This post is my 2026 DevOps roadmap: what the job market data actually says, what AI changed in the day-to-day work, which certifications are worth the money, and a learning path from Linux to platform engineering that I would follow if I were starting again this year.

What does the DevOps job market look like in 2026?

The headline numbers are better than the mood. Indeed Hiring Lab reported in July 2026 that US software development postings have grown by almost 15% since late February 2025, while overall postings fell 7% over the same period. That is the good half. The other half is that software postings are still about 27.5% below their pre-pandemic level, and the rebound is lopsided: 71% of the increase between May 2025 and May 2026 came from senior roles, and 37% came from jobs with "AI" in the title (the two groups overlap).

Indeed does not split out DevOps or SRE, so treat that as the weather for the whole category. Stanford's Digital Economy Lab found that employment for the youngest workers in the most AI-exposed occupations, software development included, fell about 13% relative to less exposed peers in its original paper, and its February 2026 follow-up put the gap at about 16% by October 2025. The authors are careful to say AI is not the sole cause. Still, the pattern is consistent: companies are hiring people who can already run things, and they are hiring fewer people to learn on the job.

Pay for infrastructure work has held up. The Stack Overflow 2025 survey put the global median for DevOps engineers at $87,011 and for cloud infrastructure engineers at $103,113, with US medians of $165,000 and $189,000. The DevOps median rose 16.2% on the year. The 2026 survey, published on 6 October with just over 30,000 responses, has not broken out salaries by role on its public pages yet, so 2025 is the latest figure I can cite.

Two more signals point the same way. The Linux Foundation's 2025 State of Tech Talent report, based on more than 500 hiring and training leaders, found 61% short of FinOps and cost optimisation skills and 65% short on cybersecurity, and 72% now prioritise upskilling existing staff over hiring, up from 48% in 2024. And the CNCF's 2025 annual survey found 82% of container users running Kubernetes in production, with 66% of organisations that host generative AI models running some or all of their inference on it. The AI boom is running on exactly the infrastructure DevOps people already operate.

What AI actually changes in a DevOps engineer's job

AI changed the cost of typing. It did not change the cost of being wrong in production. The 2026 Stack Overflow results say 73% of people who use AI coding assistants or agents use them daily, but only 20% use AI for deploying or operating production systems, the lowest of the uses listed. Developers will happily let a model write code in areas they know, and they get noticeably more cautious when the blast radius is a live system.

Trust tells the same story. In 2025, 46% of respondents actively distrusted the accuracy of AI tools against 33% who trusted it, and only 3% trusted it highly. The 2026 survey asked the question differently and found trust is conditional rather than absent: 48% trust AI output when they can verify it easily, and only 6.6% trust it for important decisions.

That is the job description for the AI era in one chart. The model drafts; someone has to verify. DORA's 2025 report found AI adoption, for the first time, associated with higher delivery throughput but still with higher instability, which I wrote about at length in DevOps Still Matters in 2026. More change flowing through the same pipeline means the pipeline, the tests, the rollback path and the observability matter more, not less.

In my own work, first drafts of Terraform, Helm values and CI workflows got much cheaper. Reading them got more important, because a plausible wrong IAM policy looks exactly like a correct one. Debugging did not get much faster: an assistant can suggest causes for a 502, but someone still has to know that the load balancer's idle timeout is longer than the app's keepalive and go and prove it (I walk through that exact failure in How to Fix 502 Bad Gateway and 504 Gateway Timeout). The roadmap below is weighted towards the skills that let you judge an answer, not produce one.

The DevOps learning path for 2026: from Linux to platform engineering

This is not a list of forty tools. The hard part is depth in a handful of layers, in an order where each one explains the next.

Start with Linux and networking, and stay longer than feels comfortable: processes, systemd, permissions, ss, dig, curl -v, tcpdump, what happens between DNS lookup and TLS handshake. Nearly every incident I have worked bottoms out here, including the ones inside Kubernetes. Bash and Python are the two languages to pair with it: the 2026 Stack Overflow survey has Python used by 58% of respondents and shells by 51.3%. Add Go later if you want to read the source of the tools you run.

Then Git and CI/CD: branching, pull requests, a pipeline that builds, tests, scans and deploys, and the discipline of small changes. Then containers: images, layers, registries, why a container is a process with namespaces and cgroups rather than a small VM. Then one cloud provider properly: IAM, networking (VPCs, subnets, routing, security groups), load balancers, managed databases, and the bill.

Infrastructure as code comes next. Terraform or OpenTofu, state, modules, plan review, drift. This is the layer where AI assistance is most useful and most dangerous, because a generated plan that destroys a database looks tidy. Kubernetes follows: scheduling, Services, probes, requests and limits, RBAC, and above all troubleshooting. Then observability (metrics, logs, traces, OpenTelemetry, SLOs) and security and supply chain (secrets, workload identity, image scanning, signed artefacts, policy as code). Platform engineering sits on top: turning all of the above into paved roads other teams use without filing tickets.

Stage	You can do this without help	Proof that convinces a hiring manager	Primary resource
Linux and networking	Debug a slow or failing service from a shell	Write-up of a real outage you traced	roadmap.sh DevOps, man pages
Git and CI/CD	Build a pipeline with tests, scan and deploy	Public repo with a working pipeline	Your CI vendor's docs
Containers	Build small, non-root images and explain layers	Image under 100 MB with a scan report	Docker and OCI docs
Cloud (one provider)	Design a VPC, IAM and load balancer from scratch	A diagram plus the Terraform for it	Provider's free tier and docs
Infrastructure as code	Plan, review and recover state safely	Module with tests and a CI plan step	Terraform or OpenTofu docs
Kubernetes	Fix a broken cluster under time pressure	CKA (troubleshooting is 30%)	kubernetes.io docs
Observability and SRE	Define an SLO and alert on burn rate	Dashboard and alert rules for your own app	OpenTelemetry and Prometheus docs
Security and supply chain	Remove static credentials, sign and scan images	OIDC-based deploy with no long-lived keys	SLSA, Sigstore, provider docs
Platform engineering	Ship a self-service template others use	CNPA, an internal platform demo	CNCF platforms white paper

The best way to practise the Kubernetes stage costs nothing. Spin up a local cluster, break it on purpose, and fix it without an AI assistant first. This is the lab I would give anyone preparing for the CKA in 2026:

# Local CKA-style practice lab. Requires Docker.
# kind v0.33.0 (August 2026) defaults to Kubernetes v1.37.0, the version the CKA currently tests.
go install sigs.k8s.io/kind@v0.33.0        # or download the release binary

cat > kind-lab.yaml <<'EOF'
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
nodes:
  - role: control-plane
    image: kindest/node:v1.37.0@sha256:a1ed56cfb0e7b93589bdf97c8cd566405a265939e3620fc4f5de89adff580ae5
  - role: worker
    image: kindest/node:v1.37.0@sha256:a1ed56cfb0e7b93589bdf97c8cd566405a265939e3620fc4f5de89adff580ae5
EOF
kind create cluster --name lab --config kind-lab.yaml

# Break something: a Deployment whose readiness probe points at the wrong port.
kubectl create deployment web --image=nginx:1.30 --replicas=2
kubectl patch deployment web --type=json -p='[{"op":"add","path":"/spec/template/spec/containers/0/readinessProbe","value":{"httpGet":{"path":"/","port":8080},"periodSeconds":5}}]'
kubectl expose deployment web --port=80

# Now diagnose it the way the exam (and production) expects:
kubectl get pods -l app=web                    # READY 0/1
kubectl describe pod -l app=web | tail -n 15   # look for "Readiness probe failed"
kubectl get endpointslices -l kubernetes.io/service-name=web   # no ready endpoints
# Fix the probe port to 80, confirm endpoints appear, then delete the cluster:
kind delete cluster --name lab

Ten rounds of that, each with a different fault (bad image tag, missing ConfigMap, NetworkPolicy blocking DNS), teach more than a month of videos.

Which DevOps certifications are worth it in 2026?

Certifications do not get you hired on their own, but a hands-on one is a reasonable signal when you have no production history. Multiple-choice exams test vocabulary; hands-on exams test whether you can do the work with a clock running.

Certification	Format	Price	Valid for	Notes	Source
CKA (Certified Kubernetes Administrator)	Hands-on, 2 hours	$445, one retake	2 years	Tests Kubernetes v1.37; troubleshooting 30%, cluster architecture 25%, networking 20%	Linux Foundation
CNPA (Platform Engineering Associate)	Multiple choice, online	$250, one retake	not stated on page	Core fundamentals 36%; launched June 2025	CNCF
CNPE (Platform Engineer)	Hands-on, 2 hours	check current page	check current page	Announced November 2025 as the advanced step above CNPA	CNCF
Terraform Associate (004)	Multiple choice, 1 hour	check current page	check current page	Tests Terraform 1.12; adds ephemeral values and custom conditions	HashiCorp
AWS DevOps Engineer Professional	75 questions, 180 minutes	$300	3 years	Pearson VUE, in person or online	AWS

If you buy one, buy the CKA: hands-on, vendor-neutral, and its biggest domain is troubleshooting, the skill AI is worst at replacing. A cloud professional certification is worth it if you already work in that cloud and your employer pays. The CNPA is cheap and a decent reading list, but I would not expect it to move a hiring decision yet.

Is DevOps still a good entry point for juniors?

Honestly, it is harder than it was in 2021, and the data above explains why. DevOps was never really an entry-level job: it is the job you grow into after you have shipped software or kept servers alive. With postings tilted 71% senior, the realistic path in 2026 runs through an adjacent role (support engineer, junior backend developer, sysadmin, NOC, QA automation) and moves sideways once you have production stories to tell.

What gets someone through that door is evidence: a public repo with a real pipeline, a cluster you built and broke, a postmortem for your own homelab outage, a merged pull request to an infrastructure project. And when 72% of organisations prefer to upskill existing staff, volunteering for on-call and the deploy pipeline inside your current company is often faster than applying cold. Use AI as a tutor that explains why, then close it and do the exercise yourself.

Where the roadmap ends: platform engineering

Gartner has long predicted that 80% of large software engineering organisations would have platform teams by 2026; I found no published check on whether that happened, but the job titles certainly did. The difference from classic DevOps is the customer: a platform team builds internal products such as golden-path templates, self-service environments and guardrails expressed as policy. I go through the role boundaries in Platform Engineering vs DevOps vs SRE and the Kubernetes side in Kubernetes as the Agent Control Plane.

The AI angle is concrete here. Agents are now consumers of the platform: they open pull requests, request environments and run code. Somebody has to give them scoped identities, sandboxes, budgets and audit trails, and that somebody is a platform engineer. The CNCF survey's top adoption challenge was not technology at all: 47% cited cultural change within development teams. That is the work platform engineering exists to do.

FAQ

Is DevOps still a good career in 2026?

Yes, with a caveat. Software postings on Indeed are up almost 15% since February 2025 while overall postings fell, and DevOps and cloud infrastructure salaries remain among the highest in the Stack Overflow survey. The caveat is that the growth is concentrated in senior roles, so the entry route now runs through adjacent jobs.

Will AI replace DevOps engineers?

Not on current evidence. Only 20% of developers in the 2026 Stack Overflow survey use AI to deploy or operate production systems, and only 6.6% trust it for important decisions. AI is replacing the first draft of infrastructure code, which makes review, testing and incident skills more valuable.

Is the CKA worth it in 2026?

It is the one certification I would pay for myself. It costs $445 with one retake, runs two hours hands-on against Kubernetes v1.37, and 30% of it is troubleshooting. It proves you can work a broken cluster, which a multiple-choice exam cannot.

How long does it take to learn DevOps from scratch?

If you can already program, expect roughly a year of steady practice to reach a working level across Linux, CI/CD, containers, one cloud, Terraform and Kubernetes; longer without that background. Documented projects at each stage matter more than the timeline.

This quarter, if I were building my own skills from scratch, I would spend the first month only on Linux and networking with no AI assistant open, book the CKA for the end of the quarter, and build one end-to-end project: a small app with a CI pipeline, Terraform for the cloud pieces, a Kubernetes deployment with proper probes, an SLO and an alert. Then I would let an agent rewrite parts of it and practise catching what it gets wrong, because in 2026 that review is the job.

Related: DevOps Still Matters in 2026, Infrastructure as Code in 2026 and Kubernetes as the Agent Control Plane in 2026.