In my last post, I mentioned how the past few months feel like a frantic race to grab any piece of the pie before AI turns all software into a commodity. If you’re reading this, you’re probably building with the same AI dev tools I am, so I won’t belabor how astonishing they’ve become.
Tools like Cursor, Copilot, Replit, and Windsurf are getting so powerful, so quickly, that it’s almost surreal. They can one-shot entire MVPs and crank out new features in a single pass. Sure, code quality can be hit-or-miss, but context windows keep growing so fast that we may soon never have to worry about technical debt. In fact, Cursor’s “agent” feature is so shockingly good that I’ve started feeling downright superstitious using it. Just yesterday, I caught myself praying to my MacBook Pro.[^2]
The evidence is everywhere. Anthropic recently published a paper showing that the vast majority of Claude conversations revolve around programming or math. This makes sense, partly because code is formal and structured — a perfect training ground for LLMs — and partly because developers are overrepresented among early adopters. Plus, Claude is just ridiculously good at programming (in my opinion, still better than any other model).
But as the low-hanging fruit in code-related markets get picked and the developer-tooling space saturates, we can expect the next wave of AI focus to shift toward other industries. Programming is relatively easy to automate because it’s formal and quickly verifiable. Automating more chaotic domains — ones with real-world constraints, messy data, physical labor, or unpredictable human behavior — will require a different architecture and more training resources. But it’s naïve to think breakthroughs there aren’t coming. Someday, fields outside STEM will see the same breathtaking results we’re now watching in code and math.
Which brings me to the real reason I’m writing this: finding SaaS alpha[^4] in a world where software is becoming a commodity, foundation models are steamrolling everyone else, and the future gets harder to predict by the day. That sentence is half a punchline, but it’s the core question: where are the new software moats when code has become so cheap to create?
Recent AI breakthroughs have been exciting[^5] mostly because they promise to transform every industry. They also hint that, eventually, most human labor will be replaced. That’s an uncharted scenario. Current economic models aren’t built to handle labor costs approaching zero.
What happens when the cost of labor goes to 0?
Traditional economics depends on scarcity. Labor scarcity is a big part of that equation. If labor becomes free, that upends almost every economic model. We’d expect:
- Companies to produce more at lower costs — human labor expense disappears.
- Prices to plummet, since production becomes nearly free.
- Wealth distribution to skew even more: owners of AI and hardware get very rich; everyone else loses their main source of income.
- Big companies, with huge capital reserves, scale faster than ever. Smaller ones without capital struggle to keep pace.
- Growth stops being about hiring constraints or salary budgets and becomes about raw materials, AI system capacity, and actual consumer demand.
Given all this, it’s safe to assume that if something can be automated, it soon will be. It’s also safe to assume that traditional moats — long product cycles, big engineering teams, problem complexity, resources, etc. — are dissolving. AI dev tooling lowers the cost of software labor so dramatically that Silicon Valley is hunting desperately for new moats. If nobody finds any, software becomes a commodity, and fewer fortunes get made.[^6]
If I had to guess, I’d bet many successful future software companies will anchor themselves in industries so specific and messy that they resist automation by general-purpose models. Construction is a good example. It’s nothing like code, which is formal and consistent. Real-world building projects are chaotic: schedules slip, people make mistakes, weather can derail everything. It’s not “plug in algorithm, get building.” That unpredictability makes it less likely a single foundational model could just bulldoze the entire sector.
Because so many real-world tasks are inherently non-deterministic — and because construction (along with industries like agriculture and logistics) lies mostly outside the bubble of traditional tech — I think there’s huge opportunity there. The fragmented nature of these domains and their massive variation in processes mean you need deep domain knowledge to build effective tools. Generic LLMs won’t cut it without serious customization.
A small personal anecdote
This is why I left my SWE job to work in construction. After years in software, I realized I couldn’t think of any new ideas outside typical dev tools or security products. Everything I came up with was already saturated by people like me. I realized I needed domain expertise somewhere else. Reading a book wasn’t enough; I had to do the work, live in that world, and learn the nuances. So I quit my SWE job to become a project estimator for a commercial subcontractor. Ironically, the final push came on March 14, 2023, when GPT-4 launched. I saw the writing on the wall and resigned the next day.
Construction is just one example of a domain with fundamentally human, unpredictable processes that require true on-the-ground know-how. The same is likely true of agriculture, manufacturing, retail, and all sorts of real-world businesses. Because these fields aren’t easily “one-shotted” by a general-purpose model, they’ll require specialized, end-to-end AI tools tailored to the oddities of each vertical.
Of course, I could be completely wrong. Maybe AI bulldozes everything, maybe it bulldozes nothing, or maybe some third scenario emerges. But I’m convinced there’s plenty of demand in every industry for tooling as powerful as what developers enjoy today. That’s why I’ve been heads-down building construction tech the last few months. I’ve also been keeping an eye on adjacent fields like manufacturing, real estate, and retail to see if new AI products pop up. For now, it’s quiet. I don’t expect it to stay that way.
[^1]: Just kidding, not yet.
[^2]: Source: I made it up.
[^3]: The longer version of my personal journey from SWE to construction can be found in earlier posts.
[^4]: I’m really stretching the term “alpha,” but it sounds more exciting than “business ideas.”
[^5]: Exciting doesn’t always mean fun. Being chased by a grizzly is exciting, but I wouldn’t be smiling.
[^6]: Distribution alone isn’t a moat; it just means you have a commodity with good marketing.