I decided to publish this blogpost i’ve been writing in April 2026. I’ve felt increasingly wortheless due to LLMs beeing able to - “generate code”.
The Crisis Is Real. So Is the Way Out.
I maintain nixpkgs, NixOS, infrastructure and build on clan.lol. I watch the AI wave and formulate my own thesis: the layer where things either run correctly or they don’t. From that vantage point, the current crisis looks both more serious and more navigable than most people realize.
Diagnosis
Something structural is breaking.
The value chain itself.
Developers are asking whether their skills still matter. Founders are rebuilding pitch decks around AI features they haven’t thought through. VCs are concentrating capital into AI at a rate that makes everything else feel like a second-class citizen. Everyone is moving, almost nobody has a clear direction.
The worst part is that all the standard reassurances are getting recycled: Developers won’t be replaced, just augmented. AI can’t replace creativity. The fundamentals still matter. maybe true, maybe not.
Why This Moment Is Different
Every decade has a forcing function. Mobile reorganized distribution. Cloud reorganized infrastructure. Those shifts were large, but they mostly expanded the playing field — more apps, more services, more surface area for developers to build on.
This one is different in kind.
AI is a substitute for the act of building itself. When the tool can do part of the work, the value of doing the work changes.
Three things are changing:
The unit of software is changing For years, software was functions, services, APIs. Now the unit is agents, context, and intent.
The criteria for “good enough” seems to have collapsed A vibe-coded prototype is often sufficient for a proof of concept. From what I see in the NixOS and infra community, this is landing hard on developers whose identity was tied to being the person who could write the code. Vibe-coded projects hit a wall fast, but many don’t survive long enough to hit that wall. Meanwhile, software developer job postings in the US are down more than 33% from their 2020 peak1. The market has changed before the conversation has even started. My thesis: FOMO.
The VC market is skewed In 2025, AI startups captured more than half of all global venture capital for the first time. ($192 billion out of $366 billion total2). The number of venture funds actively raising capital collapsed from 4,430 in 2022 to 823 in 2025, an 81% decline3. PitchBook’s director of research summarized it plainly: “You’re in AI, or you’re not.”4 Non-AI companies can still raise; But they’re competing for a dramatically smaller share of available capital, and investors now expect a clear path to profitability that wasn’t required two years ago.
Here is what I think is actually true: there is no single correct answer to this moment:
Substrate Wins
The value chain collapses toward the foundation.
Every time a new layer gets commoditized, value migrates down. This has already happened to thin AI wrappers: startups that built “AI for PDF summarizing” or “AI for copywriting” and charged subscriptions for it got wiped out when foundation models absorbed those features natively. They had no proprietary data, no unique architecture. (The middlemen got cut out5).
What remains valuable is the thing AI runs on and runs through: compute, correctness guarantees, reproducible systems, trust infrastructure, deployment primitives. This is why infrastructure companies like CoreWeave and Databricks dominated the 2025 landscape while application-layer generalists struggled6.
This is the future I’m personally betting on, at least in part. Clan is built on the premise that as more production systems are generated rather than written, the correctness and reproducibility of the substrate becomes more valuable, not less. The more AI output you get, the more a principled foundation layer is worth.
To the Infrastructure engineers, security architects, anyone working on systems that have to be correct by construction. Formal verification. Cryptographic infrastructure. Reproducible builds. Not glamorous, but increasingly non-negotiable.
Let’s build the thing AI needs to run safely in the world. That’s the unglamorous responsible work.
Domain Beats Syntax
Generic software approaches zero marginal cost. The new moat is knowing what’s worth building.
When anyone can generate code, code is no longer the constraint. The constraint becomes: do you understand the problem space well enough to know whether the output is right? Can you evaluate a generated system correctness against real-world requirements that are impossible to fully specify? Can you see what’s missing? Can you stress test your system, where it falls apart unnoticed?
A cardiologist who uses AI to read ECGs is made more powerful by it. A cardiologist who only reads ECGs and has no model of cardiac physiology might be in trouble.
The same applies in software. The generalist who can write code but doesn’t deeply understand his domain is exposed. The specialist who understands the domain better than the AI can model it — and uses AI to accelerate execution — compounds their advantage. Late-stage AI companies with vertical focus commanded a 100% valuation premium over non-AI peers at Series C in 20257, and investors explicitly shifted toward vertical tools with domain specificity over horizontal generalists8.
To the developers with deep vertical expertise. Healthcare, legal, manufacturing, agriculture, infrastructure — any domain where correctness has real-world consequences and requirements are messy and contested: Your expertise is the scarce input. Code is the output. Stop pricing yourself as a coder; start pricing yourself as the person who guarantees whether the system is right.
Orchestrator Emerges
Software development doesn’t disappear. The craft transforms.
Developers become the people who define intent, specify correctness, design failure modes, and set the constraints within which systems of agents operate. Code becomes a byproduct. The job looks less like programming and more like architecture; Closer to systems thinking combined with editorial judgment.
This is a harder discipline. Writing a function is easier than specifying precisely what you want a system to do, under what conditions, with what guarantees. Most people who call themselves engineers today have never had to do that rigorously. I’ve watched this in the NixOS ecosystem: the people who thrive are the ones who can reason about what a system should do and communicate that precisely enough.
To the architects, technical leads, anyone who thinks about systems rather than components. People who can write precise, unambiguous requirements. People who can reason about edge cases and failure modes at the system level. The human role in software shifts from author to director. The developers who make that shift deliberately will define what engineering looks like in five years.
Conclusion
Start by asking what would be true if your most important skills were worth more in two years than they are today — and work backward from there.
The crisis is real. But it is a crisis of “code” no longer being sufficient as a full value proposition.
The developers who get through this are the ones who can articulate clearly what they do that isn’t just writing code.
Can you?
April 2026. Written from somewhere in the middle of the transition, when everything felt weird
Footnotes
-
Indeed/Federal Reserve data via Visual Capitalist, March 2025: Charted: The Decline of U.S. Software Developer Jobs ↩
-
PitchBook data reported by Bloomberg, via TechBuzz, 2025: AI Startups Capture Over Half of All VC Money for First Time. Cross-referenced with Crunchbase end-of-year analysis: 6 Charts That Show The Big AI Funding Trends Of 2025 ↩
-
PitchBook data, ibid. The fund count drop (4,430 → 823) reflects the collapse in smaller and emerging managers’ ability to raise new vehicles. ↩
-
Kyle Sanford, Director of Research, PitchBook, quoted in Bloomberg, ibid. ↩
-
Editorial GE analysis of the 2025 AI market: AI Winners and Losers 2025 ↩
-
CoreWeave reached $1.9B in 2024 revenue (737% YoY growth); Databricks reached $4B ARR as of Q2 2025. Source: Top Most Profitable AI Startups in 2025 ↩
-
SVB State of the Markets data, cited in: State of VC: It’s all about AI now ↩
-
Wellington Management VC Outlook 2026, via Harvard Law School Forum on Corporate Governance: Venture Capital Outlook for 2026: 5 Key Trends ↩