For years, CUDA engineers were the rock stars nobody outside Silicon Valley had heard of. These were the people who could look at Nvidia’s chips and squeeze performance out of them that seemed almost impossible – shaving milliseconds here, optimizing memory there, making AI models run faster and cheaper. Companies paid them ridiculous salaries because when you’re burning through millions on compute costs, someone who can cut that bill by 20% is worth their weight in gold. Nvidia's chips have gotten valuable enough that one tech CEO is now facing two decades in prison over a $300 million chip-smuggling scheme.

Now? They’re teaching AI to do their jobs. And they’re weirdly okay with it.

The Job That’s Eating Itself

Here’s what a CUDA engineer actually does – or did, I guess. They write these things called kernels, which are basically tiny, hyper-optimized pieces of code that tell a GPU how to perform a specific task as fast as physically possible. It’s tedious work. You write a kernel, test it, tweak it, test it again, maybe write five more versions, test all of those, and eventually figure out which one is fastest.

It’s the kind of work that requires deep expertise but also… it’s repetitive. Really repetitive. Which, if you’ve been paying attention to where AI is actually useful (not where the hype says it’s useful), makes it a pretty obvious target for automation.

So now we’ve got AI that can generate hundreds of kernel variations, test them all, and pick the winner. Faster than any human could. And the CUDA engineers who used to spend their days doing exactly that are now spending their time managing the AI instead.

Wait, Isn’t This Supposed to Be Scary?

Look, every story about AI and jobs follows the same script: AI comes for your job, workers panic, future uncertain, etc. But the engineers I’m hearing about aren’t panicking. They’re… adapting? Jeremy Nixon, who founded a startup called Infinity that builds exactly this kind of optimization AI, says engineers are shifting to “setting goals, checking results, and stepping in when the AI gets stuck.”

Which honestly sounds less like replacement and more like promotion. Instead of being the person writing the hundredth kernel variation at 2am, you’re the architect deciding what needs to be optimized and whether the AI’s solution actually works.

The Part Nobody Talks About

But here’s where it gets interesting, and this is the thing that should probably worry us more than the “robots taking jobs” angle: AI writes weird bugs.

Anne Ouyang from Standard Kernel (another AI infrastructure startup – they’re everywhere now) says the bugs AI introduces are “bizarre” – the kind that humans just wouldn’t write. And I mean, of course they are. AI isn’t thinking through the code the way a person does. It’s pattern-matching its way to solutions that technically work but might be structurally insane.

So now these engineers aren’t just managing AI output. They’re becoming bug archaeologists, digging through AI-generated code to find problems that don’t make intuitive sense. Ouyang says code review has become “more intense,” which feels like an understatement. You’re not just checking if the code works – you’re checking if it works for reasons that won’t blow up later in some completely unpredictable way.

“The job of reviewing AI code has become more intense. AI can introduce bizarre bugs that humans wouldn’t have written.”

This Is Happening Everywhere, By the Way

CUDA engineering is just the canary in the coal mine here. This same pattern – developers moving from writing code to managing AI that writes code – is spreading across the entire software industry. And it makes sense, kind of. If AI can handle the repetitive stuff, the boilerplate, the thousand tiny variations of the same problem, then yeah, let it. That same AI-driven demand is reshaping hardware economics too, as seen in Micron's 88% margin secret on memory chips.

But there’s something unsettling about the speed of this shift. These aren’t junior tasks being automated. CUDA engineers are specialists. They’re the people with the deep, hard-won knowledge. And even they’re stepping back to let the AI take the wheel (while they watch nervously from the passenger seat). That same unease about moving fast shows up elsewhere in the industry, like how OpenAI's safety team vanished overnight.

The optimistic take is that this frees up human engineers to do more creative, strategic work. The pessimistic take is that we’re training a generation of developers who know how to prompt AI but not how to actually solve the problem themselves when the AI inevitably fails.

What This Actually Means

I don’t think CUDA engineers are going extinct. Not yet, anyway. But the job is fundamentally changing, and faster than anyone expected. What used to be hands-on expertise is becoming supervisory expertise. You still need to know how GPUs work, how memory optimization works, how CUDA itself works – but increasingly, that knowledge is used to validate what AI produces rather than to produce it yourself.

And honestly? Maybe that’s fine. Maybe this is just how technical work evolves. Developers used to write assembly by hand, then higher-level languages automated that away, and the world kept spinning.

But I keep thinking about those bizarre bugs. The ones that humans wouldn’t write. Because the thing about automation is that it works great until it doesn’t, and when highly optimized AI-generated code fails in bizarre ways at scale… well, you better hope you’ve still got engineers who actually understand what’s happening under the hood. Not just how to ask the AI to fix it.