Why network effects, taste, and rails are the new software moats
In an era of AI coding agents, the true moats for companies now lie in four key areas: deterministic rails for critical systems, non-replicable network effects, genuine taste and strong branding, and core software infrastructure that is too integral to replace, says Lewis Liu
I sit on advisory committees for several investment firms, and over the past few months, one question keeps coming up: how do you think about moat for software companies now?
Recently I had breakfast with a partner at a large-cap PE fund who put it this way: “At least it’s not boring. Better than just doing vanilla leveraged buyouts”. He was joking – sort of. But the underlying anxiety was real.
SaaStr, the well-known SaaS community platform, recently published a chart showing roughly an almost 80-percentage-point gap in share price performance between traditional SaaS workflow companies and infrastructure and cybersecurity firms. Earnings per share at many of these software companies are still going up. And yet the market is already pricing in disruption. That bifurcation tells you everything.
We’ve seen this before according to Goldman Sachs. Around the dot-com boom, 25 years ago, newspaper companies were printing money, earnings had never been higher. But the share price decline had already started. It took five years for the fundamentals to catch up with what the market had already priced in. Those two lines eventually converged. I think the same thing is happening to software right now.
So when investors and founders ask me how to think about moat in a world of AI coding agents, I put it into four buckets.
1. Deterministic rails: the stuff that can’t afford to be wrong
There’s a real and still-unresolved question about whether AI coding agents can produce genuinely deterministic software. Not “usually works” (not good enough), but actually 100 per cent deterministically works, every time, no exceptions.
Amazon Web Services had a 13-hour outage caused by its own internal AI coding tool, Kiro, which decided the best fix was to delete and recreate the environment it was working in. In my network, a fintech lost several million dollars because an AI-written pricing model passed every automated test and then failed in production because it got a pricing formula wrong. Every day, I hear of AI coding agents wiping out production databases or deleting entire repositories. These aren’t edge cases, they’re the predictable result of removing the human from the loop.
Think about payment rails; medical workflows; legal contracts; the settlement infrastructure of financial markets. I genuinely cannot imagine a serious bank vibe-coding its payment stack. The regulatory exposure alone would (or at least should) kill it, before you even get to what happens when it fails.
My view: software that underpins these critical systems retains its moat, even if AI compresses pricing on the edges and automates lower-risk workflows. The contested question is how much territory that actually covers, and whether it’s shrinking faster than most people think.
2. Network effects: nothing to do with the code
Here’s the uncomfortable truth about Whatsapp, Visa and Linkedin: none of them are technically complicated in the grand scheme of things. You could, with current AI tools, rebuild the feature set of any of them fairly quickly. The point is completely irrelevant.
The value isn’t the codebase. It’s the network itself – the accumulated relationships, the flows of value between users, the simple fact that everyone’s already on it. You can’t vibe-code your way to 3bn WhatsApp users. You can’t replicate a payment network by writing better software.
As AI accelerates the commoditization of software production, network effects become a more interesting moat, not a less interesting one. Precisely because they’re immune to the dynamic disrupting everything else.
3. Taste: the most underrated moat in tech
“Taste” is a word you hear a lot in the Valley right now, and I think it deserves more analytical credit than it usually gets.
Think about the original iPhone. Nearly every component Apple used was highly commoditized; that’s exactly why they could manufacture it in China. What wasn’t commoditized was Steve Jobs’s obsessive, almost irrational understanding of what consumers actually wanted before they did. His taste. His ability to package known components into something that felt, to the person holding it, like inevitability.
Take Lovable, the AI development tool. Frankly, it’s probably not as technically capable as Claude Code or Codex. But I still use it regularly for product and design work because it has a particular feel: easy, intuitive. It is designed for the way I think as a founder/CEO and product person rather than as an engineer. That’s a taste advantage, not a functional one. And taste, combined with strong branding, is genuinely hard to replicate even if the underlying code isn’t.
4. Software infrastructure: veins vs organs
There’s an analogy doing the rounds in investment circles right now that I find useful. Replacing a workflow SaaS system, a Salesforce, say, is like an organ transplant. It’s painful, expensive, and takes months, but you can do it.
Replacing your core infrastructure, your security architecture, your database layer, your system of record, is more like trying to remove someone’s veins. Almost impossible to do without killing the patient.
This is partly what explains the SaaS bifurcation. The market isn’t just distinguishing between good and bad software companies. It’s distinguishing between software that can be replaced and software that structurally cannot. These systems sit in a meaningfully different position than workflow tools, which are increasingly vibe-codeable by anyone with a prompt and an afternoon – perhaps not because they are hard to build, but because they have a “software network effect” both within and outside the organization.
The uncomfortable answer
The honest answer to “what has a moat” in this world of existing software is: less than before, but definitely not nothing.
The newspaper analogy is instructive again here. Print advertising did eventually collapse, the market was right. But the New York Times still exists, still earns revenue, still matters. The winners were the ones who got brutally honest about which parts of their business were genuinely defensible and doubled down on those, rather than pretending the disruption wasn’t coming.
For software companies and their investors, the question now is the same one. Which category are you actually in? Are you building deterministic rails that can’t afford to fail? Do you own a network that can’t be replicated by better code? Is your product built around genuine taste with a depth of customer understanding that goes beyond features? Or are you a workflow tool sitting in the middle, waiting to be disaggregated?
The market is already asking this question. Earnings haven’t caught up yet. But if the newspaper analogy holds, they will.