There are 47 million software developers in the world. More than half are under 35.
That second number is the one that should keep you up at night. Young developers are the most exposed to a shift that's already happening. The median developer salary in the US sits around $150,000. Good money. Life-changing money in most of the world. And the question everyone keeps asking is whether AI will make that number go to zero.
I don't think AI replaces developers. I think it already changed what "developer" means. Most people just haven't noticed yet.
The Chart Everyone Misreads
Kale Writes Code put together a sharp breakdown of the Indeed hiring data that's worth understanding properly, because most people get it wrong.
The chart shows an index of software developer job postings over time, anchored to February 2020 as the baseline (100). In early 2022, the index hit 234. That's not 234 jobs. It means hiring was 134% above pre-pandemic levels relative to seasonality. Companies were hiring like the world was ending. Then the layoffs started in late 2022, and the index cratered.
Today, the US sits at 69. The UK at 63. Germany at 58. Canada at 78. Australia is the outlier at 122, actually above pre-pandemic levels.
Most people misread this: a value of 69 doesn't mean the developer job market is dead. It means we're still recovering toward the pre-pandemic baseline. The question is whether we'll ever get back to 100, or whether AI has permanently shifted where the baseline should be.
When you overlay AI milestones on this chart, the picture gets interesting. GPT-3.5 launched in November 2022, right as the layoffs accelerated. GPT-4 arrived in March 2023. The first commercial reasoning model (o1) dropped in September 2024. The DeepSeek moment hit in January 2025. And throughout all of this, coding agents like Claude Code and OpenAI's Codex started reshaping how actual work gets done.
The tug-of-war is real. Is the market recovering slowly toward normal? Or has "normal" moved?
Look at Australia and Canada. Both are nearing or exceeding pre-pandemic levels. The US and Europe are still well below. One interpretation: those markets simply recovered faster. Another: the US tech industry, where AI coding tools were adopted earliest and most aggressively, is the first to feel the structural shift. The countries that trail in AI adoption look healthier on the hiring chart. For now.
Dario Amodei, Anthropic's CEO, said publicly that within a few years, 50% of entry-level white-collar positions will be impacted by AI. Debate the timeline if you want. The direction is clear.
The K-Shaped Split
Orchestrators vs. Operators
There's a divergence happening among developers that looks like a K on a chart. One line goes up, one goes down, and they started from the same point.
On one side: developers who've restructured their entire workflow around AI. They don't write most of their code anymore. They write specs, architect systems, and orchestrate agents that handle implementation. The code is output, not input.
On the other side: developers who use AI the way they'd use Stack Overflow. Copy a snippet here, autocomplete a function there. The workflow is the same. AI is just a faster search engine.
Robert C. Martin (Uncle Bob, if you've read Clean Code) tweeted something that caught my attention. He described how going from two AI agents to three made him the bottleneck. He couldn't review and coordinate fast enough.
Andrej Karpathy shared a similar feeling: that he felt behind as a programmer when using AI coding agents. Karpathy. The guy who built Tesla's neural network architecture.
When two of the sharpest minds in software say "I can't keep up," pay attention. The role is changing from writing code to managing a system that writes code. A different skill entirely.
Look at OpenAI's Codex app. It doesn't even show you the code in the main interface. Diffs are reviewed separately. The entire design philosophy assumes you're orchestrating, not implementing. The shift from coding to what some call spec-driven development isn't a prediction. It's already the product design.
The Trust Wall
The DORA report shows that while early AI coding adoption traces back to 2022, the majority of developers didn't start using AI tools until 2024. According to the latest data, 52% of developers still don't use AI agents at all.
The obvious explanation is trust. Developers don't trust AI output. And honestly, that's reasonable. I've seen AI generate beautiful code that was completely wrong. I've watched it confidently introduce bugs that took hours to find.
But the trust problem is the surface. The deeper issue is workflow. Using AI agents properly isn't adding a tool to how you already work. It's changing how you work entirely. That's why Uncle Bob hits a wall at three agents. That's why Karpathy feels behind. The mental model shifts from "I write code and AI helps" to "AI writes code and I architect the problem."
That's a big ask. Especially for experienced developers who've spent years building expertise in implementation. You're basically telling them: the thing you're best at is now the thing the machine does. Your new job is the thing you used to skip over (specs, architecture, coordination).
No wonder there's resistance.
What I See From Both Sides
I use Claude Code with Opus every day. Multi-file refactors, testing, debugging. Last month I stopped writing code on my personal projects entirely. I design the scaffolding, write detailed specs, and let the agent handle implementation. The results are better than what I produced manually, and I'm faster by a factor I wouldn't have believed a year ago.
But I also teach executives about AI adoption. And the question I get most often isn't "how does this work?" It's "should we hire fewer developers?"
My answer is always more nuanced than they want. The demand data tells a clear story: machine learning engineers saw a 40% jump in demand going from 2024 to 2025. Data engineers saw 10%. Meanwhile, front-end and mobile engineers shrank by more than 5%. The market isn't shrinking uniformly. It's rotating.
The Semi Analysis report found that more than 4% of public repositories on GitHub now use Claude Code. Four percent sounds small until you realize how fast that number is growing. The adoption curve for coding agents looks less like gradual integration and more like a phase transition.
The developers thriving right now aren't necessarily the most experienced. They're the ones who adapted their workflow fastest. I've seen junior developers outperform seniors with AI, not because they're smarter, but because they don't have 15 years of implementation habits to unlearn. They approach AI as the primary builder and themselves as the architect. Seniors often do the reverse, reviewing AI output line by line, applying deep expertise at the wrong level of abstraction.
It's like checking assembly code when you should be reviewing the system design.
I catch myself doing this too. Twenty years of building things teaches you to care about implementation details. Letting go of that instinct takes effort. But every time I force myself to stay at the architecture level and trust the agent with the implementation, the output improves. The agent isn't better than me. I'm just finally working at the right layer.
The Wrong Question
"Will AI replace software developers?" is the question everyone asks. It's also the wrong one.
AI isn't replacing developers. It's splitting the profession in two. One group will design systems, write specs, and orchestrate AI agents that handle implementation. Their value goes up because they can do in a day what used to take a team a week. The other group will keep writing code the traditional way, competing with tools that get better every quarter.
The $150,000 median salary isn't disappearing. But it's moving. It's flowing toward the people who figured out how to work with AI agents, not just alongside them. Toward the ML engineers and the system architects and the people comfortable working at a higher level of abstraction.
If you're a developer reading this, here's what I'd actually do. Spend a week building something real with an AI coding agent. Not autocomplete. Not chat. A proper agent that handles implementation while you handle architecture. Claude Code, Codex, whatever fits your stack. The workflow change will feel uncomfortable. That discomfort is the signal that you're learning something important.
If you lead a team, stop asking "how many developers do we need?" Start asking "what should our developers be doing?" The answer increasingly looks like architecture, system design, and quality oversight rather than line-by-line implementation. The teams that figure this out first won't need fewer people. They'll ship more with the same people.
AI won't take your job. But someone who works with AI agents the way architects work with contractors might make your current workflow obsolete.
Which side of the K are you on?