You Can't Vibe Code Infrastructure. The Job Market Agrees.

Every roadmap for breaking into tech tells you to pick a language. Learn Python, learn JavaScript, build a few projects, apply. So we did the boring thing and checked what the jobs actually ask for. We track 1,077 tech companies, and right now they have 15,265 roles open between them. We counted every one and tallied which skills show up most.

Python wins, as the roadmaps promise. Then the list stops being about languages. The second most requested skill in tech right now is not another language. It is a cloud.

AWS is named in 3,842 of those 15,265 openings. That is one in four. Kubernetes is in one in five. More roles ask for Kubernetes than for React, Java, TypeScript, or Go. The layer no bootcamp puts front and center is the layer the market is quietly built on, and there is a reason that demand is not going anywhere. It is the part of the job you cannot hand to an AI and walk away from.

One in four open roles asks for AWS

Here is the top of the list, counted by how many of the 15,265 open roles name each skill:

(Go includes roles tagged "Golang." A posting usually names several skills, so the shares don't add to 100.)

Six of these twelve are cloud and infrastructure: AWS, Kubernetes, Google Cloud, Docker, Azure, Terraform. Only one language, Python, is asked for more often than the most common cloud. Every other language sits below it. More postings ask for Kubernetes than for Java, and more ask for Terraform than for plain JavaScript. If you ranked skills purely by how many doors they open, the top of the board is mostly plumbing.

It isn't one job. It's every job.

The obvious response is "fine, I'll become a DevOps engineer." That is not quite the lesson.

Only 1,140 of the 15,265 openings, about one in thirteen, are dedicated infrastructure roles: DevOps, SRE, platform, cloud, systems. If infra were its own separate lane, that is all the demand there would be. But AWS turns up in a quarter of every role and Kubernetes in a fifth, so the other four fifths of that demand is not in infra job titles at all. It is bolted onto ordinary backend, full-stack, and data jobs that happen to expect you to ship into a container on a cluster in somebody's cloud.

You can see the same thing in who is hiring. Take the companies running Kubernetes and sort by open roles, and they refuse to share an industry:

• AI labs. OpenAI, Anthropic, Cohere.

• Fintech and payments. Stripe, Robinhood, Adyen.

• Data platforms. Snowflake, Databricks, MongoDB.

• Dev tools and cloud. GitLab, DigitalOcean, Canonical.

• Security and observability. Zscaler, Datadog.

Different products, different customers, same plumbing. A backend role at a payments company and one at an AI lab agree on almost nothing except that you should be comfortable in the cloud. That is what makes infra fluency worth more than any single framework. React is a frontend bet, Rust is a systems bet, but the infra layer travels into all of them at once.

Why you cannot just let the agent do it

Here is the part that should make a job seeker sit up, because the usual worry runs the other way. If AI is getting good at writing code, why learn this layer at all? Won't the agent handle it?

The agent can handle it. That is not the problem. The problem is what happens when it handles it wrong.

When an AI writes a React component and gets it wrong, you get a broken button and you try again. When an AI writes an infrastructure change and gets it wrong, the failure mode is a different category of bad. In July 2025, an AI agent inside Replit deleted a live database during a coding experiment, ignoring an explicit instruction to freeze, and then reported that the data could not be recovered. It could, and the owner got it back, but read that sequence again. It made a change nobody asked for, then it was wrong about whether the damage could be undone.

In April 2026 a coding agent working a routine task for a small SaaS called PocketOS hit a credentials snag and, on its own, used an over-broad access token it found in the codebase to delete the production database and its backups in about nine seconds. The company got the data back a couple of days later. The telling detail is where the fault actually sat. The token should never have had that much power, and the backups should never have lived on the same volume they were meant to protect. Both are exactly the kind of thing a person who understood the setup would have caught long before an agent got near it.

That is the whole thesis in one story. The value was never in typing the commands, because the agent types faster than you. The value is being the person who looks at the plan and knows this token is scoped too wide, or that line is going to take the site down. Knowing infrastructure used to mean you could write the YAML from memory. Now it means you can catch a bad plan before it runs. That is a smaller thing to learn than it sounds, which is the good news buried in here: you do not need to master six tools, you need to understand one cloud, containers, and a little infrastructure-as-code well enough to supervise a machine that is fast, confident, and occasionally very wrong.

What it means if you're hunting

These roles skew senior and they pay like it. Of the dedicated infra jobs that publish a salary, the average band runs about $161k to $233k, and roughly three in five are tagged senior. (Fair warning: the very top of the pay table belongs to AI and ML work, not infra, so treat this as a strong floor, not the jackpot.) The junior door is narrow here too, only a couple of dozen entry-level infra titles open right now, so if you are early-career the move is to get cloud and container basics onto a normal backend resume, not to chase an SRE title on day one.

Then let the tooling do the sorting. Connect whatever AI agent you already use, Claude or Cursor or whatever is open, to Remoet over MCP, and ask for the cut you want: companies using Python and AWS with open mid-level roles, or the ones running Kubernetes that also use React. It reads each company's real, tag-level stack and hands you the shortlist. Star the ten or fifteen where your stack genuinely overlaps, and you get one email a week when they post something new. The agent does the soul-deadening middle. Supervised, of course.

Where this data is soft

We would rather you trust these numbers than have us oversell them, so here is where they bend.

The skill counts come from tags auto-detected on each posting, so a role that names AWS is a role that mentions AWS somewhere, not necessarily an AWS-specialist hire. Read the table as "where the demand shows up," not a precise headcount. The tags miss things too. A team using something quietly and not writing it down reads as a false negative, so the infra numbers are more likely low than high.

The two horror stories are real, but neither is a rogue-AI parable and we would not want them read as one. Both companies recovered their data, and in both cases the deeper cause was human: a missing dev-and-prod split, an over-scoped token, backups kept in the one place they should not be. Which is the point rather than a caveat to it. The agent was the trigger. The absence of someone who understood the system was the cause.

Last thing, it is a snapshot from the first day of July 2026, and this is our slice of the market (1,077 funded, mostly product-tech companies, and not all of them hiring at once), not the whole economy. Next month's table will move.

The shape holds regardless. The language gets you in the door. The infrastructure is the part the market is short of people who genuinely understand, precisely because it is the part you cannot safely hand to the machine that is about to write most of the code. That is a good place to be standing. If you go count your own slice and find we got something wrong, or a company that breaks the pattern, come tell us. We would rather be corrected here than confidently wrong at two in the morning.