Note: This is not a random rant against AI, after all I am a Founder of an AI startup, Overlord Systems.
Much has been written recently about how software no longer has a moat because AI gives you both engineering ability and a high speed of producing software, therefore making any piece of software copyable in a short time, and thus software is no longer a moat. Only, this is isn’t true in the way it’s being advertised.
Most of the articles claiming no moats are handwavy making it hard to have a discussion, so I’ll try to make this piece as concrete as possible.
When discussing moats I’ll focus on the technical aspect, which is important because the moat of a company is made up of many different orthogonal things: brand is a moat, market share is a moat, vendor lockin is a moat, and so on. A moat is anything that makes it harder for competitors or gives you an advantage, and your ’total’ moat is the combination of all these factors. Software is just one of those factors.
Those saying software no longer matters and that it’s now all about brand/marketing/etc. are mistaken in both the pre- and post-AI worlds.
Now when discussing software moats, we have to understand the reality of both software and AI. AI’s ability to code has lowered the barrier to entry to software, meaning that the most common type software (webdev) and software on the lower-end of the complexity specturm, now produce a much smaller moat.
However, the crucial thing to understand here is that this only applies to common and/or low-complexity software. The usual argument goes that as AI advances even advanced software will be doable with AI, but that’s just not how (current) AI works!
The ability of AI in coding comes from it’s training data, which is mostly open source webdev/saas/personal-projects/etc. Empirically, most of that code is low quality and I would even go as far to say that most software is low quality. If it weren’t so, we wouldn’t need a Gigabyte of RAM to run a chat app (e.g. Discord, MS Teams), have latency when typing (hello JetBrains!), nor would we have the rotating door of JS frameworks, each with many, many versions, and of course the crown jewel of our industry, left-pad.
As long as the training data remains what it is (and we see no indications of this changing any time soon), AI will remain stuck at being only decent for general web development and other simple things (e.g. small GUI apps). That or a fundamental change in AI architecture allowing it to understand deeply and greatly generalize beyond its training data.
This means that while webdev software on the simpler side is now less of a moat, AI has almost no effect on moats of software that is ’non-standard’, complex, or that is trying to operate at the higher end of quality and performance.
AI only reduces the moat of simple software
To make this more concrete I’ll give a few examples of places where software is still a moat and places it’s not.
Software is less of a moat for simple CRUD apps, blogs, websites, translation, basic image editing, simpler use cases of no/low code tools. Many (especially smaller/more focused) SaaS companies are here.
Things I see being essentially untouched by AI is more complex software which includes things like: Figma, Notion, Tailscale, Asana, Stripe, Google/MS office suites, FFmpeg, databases, creative software (e.g. Blender, Photoshop), simulation and scientific software (e.g. MATLAB, SolidWorks), games/game engines, and many more.
Essentially, any software of sufficient size and/or complexity remains difficult for both engineers and AIs, and in those cases software remains a strong moat (please raise your hand if you are brave, and skilled, enough to write a custom OpenGL renderer and collaborative editor to rival Figma).
In fact, anyone with the skills and drive to create high quality and high performance replacements (not even more features!) for classic software is at a huge advantage. This hasn’t been explored much because software still has many areas it hasn’t touched, but we are slowly starting to see this with performance focused terminals/editors/file-explorers/languages/etc., like File Pilot and Zed.
Perhaps a recent (and honestly funny) example of this is DeepSeek, winning in big part thanks to doing hardcore software engineering that few others can/would do. When you take a step back, you see that many successful tech companies have significant software moats, either because of breadth, depth, or both.
Small and simple SaaS never had a true software moat, and with AI it has even less, and this is where a lot of the ‘software moats are dead’ talk comes from, but simple software is not all software!
Paradoxically, the more your product relies on underlying AI model capabilities, the less software moat you have. However, a complex software integrating deeply with AI will provide a good moat.
What’s true before and after AI is: solve a software problem outside the norm, dare to tackle difficult engineering challenges, or push your standards for quality up, and you will have a significant software moat.