Infrastructure as code after AI is the question I keep getting from junior engineers in Colombo: if a model can write my Terraform, is there still a job here? A recent Stack Overflow podcast episode, What's left for infrastructure as code after AI moves in?, put that question to Rosemary Wang, a Developer Advocate at IBM. I want to add my own read on it, aimed at people learning cloud on a Sri Lankan salary and a free-tier budget.
My short answer: the typing got automated, the judgment did not. And judgment is where you should spend your learning hours.
π What AI actually automates in IaC
The part of infrastructure as code that a model handles well is the part that was always the most searchable: syntax. Provider blocks, resource arguments, the exact name of an argument you forgot. That is real time saved, but it was never the hard part.
Here is the honest split of where the work sits:
| Task | AI does it well | Still needs you |
|---|---|---|
| Writing a resource block from a description | β | |
| Remembering exact argument names | β | |
| Deciding what the architecture should be | β | |
| Understanding blast radius of a change | β | |
Reading a plan diff before you apply
|
β | |
| Owning what breaks in production at 2am | β |
Key takeaway: AI compresses the authoring of infrastructure as code. It does not compress the responsibility for it. Those are different skills, and only one of them was ever the moat.
If your whole value was memorizing HCL, that value just dropped. If your value is knowing why a network is laid out a certain way, it went up.
β‘ The dangerous part: a plan you didn't write
The episode's framing that stuck with me is about AI not just writing infrastructure but deploying it. That is where I get cautious. Generating code is low-stakes. Applying it against a live cloud account is not.
A model can produce a Terraform plan that looks clean and still quietly does something you did not intend:
- Recreate a database instead of updating it (data gone).
- Open a security group wider than you meant.
- Rename a resource so Terraform destroys and recreates it.
The skill that matters here is reading a diff and stopping. When you review a generated plan, treat it like a pull request from a stranger who is fast, confident, and occasionally wrong. If you want to eyeball two versions of a config side by side before you commit, even a plain text diff checker beats trusting your memory of what changed.
A generated plan is a suggestion, not a decision. The
applyis the decision, and it's yours.
π Why this is good news for Sri Lankan engineers
I think this shift favors people starting out here, and I don't say that to be cheerful. The old barrier to cloud infrastructure was partly knowledge and partly cost. Both are lower now.
- The knowledge cost dropped. You no longer need to buy a long course to get past the syntax wall. You can ask a model to scaffold a config and then spend your energy understanding what it built and why.
- The practice cost is near zero. AWS, GCP, Azure, Cloudflare, and Oracle all have free tiers. You can write real infrastructure as code, apply it, break it, and destroy it for the price of your electricity.
- The remaining skill is portable. Systems thinking, security defaults, cost awareness. None of that is Sri Lanka specific, which means it pays in USD-remote rates.
Here is the practice loop I'd hand a student today:
- Pick a tiny real goal (a static site bucket, a single VM, a scheduled job).
- Let AI draft the Terraform. Read every line before you run it.
- Run
terraform plan. Predict the output before you look at it. - Apply, verify it works, then
terraform destroyso you pay nothing. - Repeat with something one step harder.
Step 3 is the whole game. If you can't predict the plan, you don't understand your own infrastructure yet, and that's exactly the gap to close.
π οΈ The skills that get more valuable, not less
If authoring is commoditized, where do you aim? My list, roughly in order of how much I'd invest:
| Skill | Why it survives AI |
|---|---|
Reading a plan/diff critically |
The last human checkpoint before real damage |
| State management (remote state, locking) | Subtle, high-consequence, easy to corrupt |
| Security and least-privilege defaults | A model will happily generate an over-permissive config |
| Cost modeling | Nobody wants a bill they can't explain |
| Debugging drift and failed applies | Requires understanding the system, not the syntax |
Notice none of those are "write the resource block faster." They're all judgment under consequences. That's the part of the job that was always underpriced because it's hard to teach and hard to fake.
For the mechanical glue around this work, keep small tools handy. Scheduling a nightly infra job? A cron expression builder saves you from second-guessing whether 0 2 * * * means what you think. Reviewing machine-generated config? Formatters and diff tools keep you honest.
π‘ What this means for you
If you're a student or a small-team builder in Sri Lanka watching AI write Terraform and wondering if the door is closing, I'd flip it. The door that's closing is the one where memorizing syntax was enough to get paid. That door was never a great career anyway.
The door that's opening is bigger:
- Learn the concepts, not the incantations. Let AI type; you decide.
- Practice on free tiers. Build, break, destroy, repeat, at zero cost.
-
Own the
apply. Be the person who reads the plan and says "no" when it's wrong. - Sell judgment, not keystrokes. That's what remote clients still can't get from a model.
Bottom line: AI moving into infrastructure as code doesn't remove the engineer. It removes the excuse that the hard part was syntax. What's left is the part that was always the actual job, and it's more learnable now than it has ever been.
The Stack Overflow episode is worth a listen if you want the full conversation with Rosemary Wang. My take is just this: don't compete with the model on speed. Compete on being the person who understands what it built.
Top comments (1)
I particularly appreciate the distinction made between AI handling syntax and the human judgment required for deciding architecture, understanding blast radius, and reading plan diffs. The table highlighting tasks that AI can handle well versus those that still require human expertise is especially insightful. The point about AI compressing the authoring of infrastructure as code but not the responsibility for it resonates deeply, as it underscores the importance of understanding the why behind infrastructure decisions. What strategies do you think junior engineers can use to develop this critical judgment and systems thinking, especially when working with AI-generated infrastructure as code?