Building Castles in Code

Why software moats will endure in the AI era, and what it means for startups

Welcome to Infinite Runway, a newsletter I’ve recently launched that will cover all things tech, startups, and investing.

Earlier this month was Tech Week in NYC, and I was asked to sit on a panel to discuss software’s defensibility in the age of AI. This topic is so often debated (and often misinformed) in today’s startup zeitgeist that I enjoyed chiming in with perspective I’ve developed over the last decade working at the intersection of software and investing, both in public and private markets.

In this post, I’ll argue that while AI will transform software, it won’t fundamentally change the core moats that create competitive advantages for software companies: distribution, brand, product architecture, and culture/operational processes.

A caveat before we dive in: this post is about software companies. Other industries like biotech and robotics have different competitive dynamics, and need their own assessment of what makes a moat.

Defining a moat

The GOAT of investing, Warren Buffett, is widely credited with popularizing the term moat as it relates to describing a business’s enduring advantages over its competitors. In Buffett’s metaphor of the moat, the business is the castle, and the business’s competitive advantage is the moat.

“Both Coke and Gillette have actually increased their worldwide shares of market in recent years. The might of their brand names, the attributes of their products, and the strength of their distribution systems give them an enormous competitive advantage, setting up a protective moat around their economic castles. The average company, in contrast, does battle daily without any such means of protection” - Warren Buffett in his 1993 annual letter to Berkshire Hathaway shareholders

Buffett used this metaphor frequently over the decades he ran Berkshire to talk about the concept of durable competitive advantage¹.

Classic examples are Apple’s integrated hardware & software, Walmart’s scale and low prices, Visa’s card network, and Proctor & Gamble’s retail distribution. These moats have historically helped protect these iconic companies from losing market share to competitors.

Why are people in the software industry talking about how AI affects moats?

The software industry is grappling with an important question: If AI enables people to build software simply by typing natural language prompts (without code), won’t everyone be able to build their own software, eroding the value provided by software vendors?

This debate began for good reasons, starting when OpenAI first released its GPT models via an API that enabled developers to integrate GPT models into software applications. Many startups initially claimed competitive advantages such as early access to the GPT API, or proprietary relationships with OpenAI, or interfaces that were slightly easier to use than ChatGPT.

Looking back, those claims were flimsy at best. They provided a temporary differentiation that created a window of opportunity for growth, but were not durable sources of competitive advantage.

Still, in the immediate aftermath of the “ChatGPT moment” at the end of 2022, users were clamoring to adopt new AI tools, and many startups grew really fast in 2023. Some even reached millions (or tens of millions) of dollars in monthly revenue.

The excitement quickly faded as OpenAI released improvements in both ChatGPT’s underlying GPT models and its UI (things like memory, personalization, canvas text editor).

Users of the early batch of generative AI startups found fewer reasons to adopt their products, and their growth plateaued as a result. Some had to lay off employees to cut burn due to slowing growth, while others pivoted. They had not yet developed the distribution or product depth necessary to retain users in the face of competition from a larger incumbent.

That was a jarring experience for founders, employees, investors, and anyone else in the industry. Now, here we are in 2025, with moats and defensibility being debated constantly on Twitter, in podcasts, and probably inside of the investment committee meetings at your favorite VC firm.

Software Moats vs. MBA Moats

One of the most famous books on moats is Hamilton Helmer’s 7 Powers. The book provides a framework with seven “powers” that are common moats seen in modern businesses, including: economies of scale, network effects, counter-positioning, switching costs, branding, cornered resources, and operational processes. These are common topics discussed in business strategy classes.

However, enterprise software moats are more nuanced. It doesn’t mean that the 7 Powers framework doesn’t apply to software, but we can be more specific about it.

There are four common moats in enterprise software:

1. Distribution leverage. Enterprise software is sticky, and depending on whether companies sell to small businesses or large enterprises, 70-100% of customers typically stick around each year. This means companies that achieve distribution and build a customer base have inherent advantages including lower customer acquisition costs and significant opportunities to cross-sell their customers new products. Over time, successful software companies can leverage distribution to dominate multiple product categories.

   1. Example - Salesforce is widely known for their CRM software. But their largest product by revenue is Service Cloud (customer service software), which is a product line built through a combination of internal development and multiple acquisitions. Salesforce was able to leverage its base of CRM customers to cross-sell Service Cloud, and along the way captured roughly 50% of the customer service software market (by far the largest player in the category).

2. Brand. Software companies that develop market leadership often have brands that become the ‘default’ or safest option within their category. Customers of business software tend to be risk-averse and have procurement processes that tend to favor established providers, even when better products exist (more on why that is the case in a future post).

   1. Example - the old saying, “nobody got fired for buying IBM” captures this idea nicely. A more modern example is Stripe, which has quickly become the default choice for software developers integrating payments APIs into their applications.

3. Product architecture. This moat is a catch-all for product-derived moats, which vary in enterprise software. Some examples are platform lock-in, integration ecosystems, and learning loops where the user’s product experience improves over time.

   1. Example - ServiceNow’s core IT service management product has deep integrations into various IT workflows and systems. Once these workflows are set up, they can be automated to reduce the need of manual human intervention, which improves the product experience and creates lock-in (ServiceNow has a 98%+ customer retention rate). CrowdStrike’s security platform is another example, where security vulnerabilities are found in one customer’s environment, they’re applied to the rest of the network, and that enhances security for the entire customer base .

4. Culture & operational processes. This one is harder to pin down, but you know it when you see it. Strong internal cultures, which often manifest as unique operational processes that a company has, can amplify a company’s ability to out-execute competitors.

   1. Example - Palantir has a mission-driven intensity that attracts high-quality engineering talent, which, when paired with their forward-deployed engineering model that embeds engineers into customer environments, results in deep customer relationships (over 30 companies pay Palantir $10M+ in annual contract value).

But what about technical moats? A fifth moat for software companies are technical advantages. This applies particularly to infrastructure software, such as databases or systems software, as well as cyber security. These can be inference speed in an AI inference platform, or compression algorithms in network infrastructure. Examples include Snowflake’s data architecture that separated compute from storage, or Palantir’s ontology layer for modeling complex data relationships.

The main challenge with technical moats in enterprise software is they often erode. Companies pioneer new software architectures that then become the standard among engineers, and competitors emerge boasting similar technical capabilities.

That doesn’t mean the company gets disrupted, it just means the original pioneer of a technical moat will see its moats revert back to the original four I wrote about above: distribution, brand, product architecture, and culture/operations. Often times these companies end up having the best brand in a category with strong distribution. In 2007, Palo Alto Networks created the next-gen firewall, and while competitors released their own version of the technology, Palo Alto Networks built a large customer base, strong brand, and diversified product suite. Today it is the largest cybersecurity company in the world.

Note: This dynamic does not apply to all areas of technology (Google’s search algorithm, Nvidia’s GPU architecture, etc.), but technical moats tend to be rare in enterprise software.

How do Software Moats Change in the Age of AI?

They won’t. Enterprise software will dramatically change in the era of AI (more on that in a future blog post), but the core moats will remain unchanged - distribution leverage, brand strength, product architecture, and culture/operational processes.

Generative AI has made software creation dramatically easier. Software engineers can use Cursor and Windsurf to improve their productivity in writing code, and anyone can use Lovable to vibe code their way to a simple application solely through natural language prompts. It’s clear this trend will continue, especially as AI agents become capable of continuously updating and managing software without the need for human intervention.

However, AI is only the latest innovation in a long line of innovations that have made software progressively easier to build over the last several decades. From mainframes, to the cloud, to no-code tools, and serverless developer platforms, software has become progressively simpler and more accessible to develop.

Yet, this pattern of software becoming cheaper to build has only resulted in increased demand for specialized software solutions that actually deliver more value than previous generations of software.

Despite software becoming cheaper to develop, building an enduring software company actually takes more capital today than at any time in history because buyers don’t purchase software as a standalone product. They expect exceptional service, deep integrations, reliability, and a trusted relationship with their software vendor. They want solutions that they don’t have to maintain themselves, so they can focus on their core business. AI will commoditize software creation, but it won’t commoditize the business relationships, distribution networks, and product ecosystems that exist between software vendors and their customers.

Of course, that does not mean there won’t be disruption to incumbent vendors. On the contrary, as a venture capitalist focusing on software & AI, I am literally betting my career on there being quite a large amount of disruption to incumbent software companies. But this will result from startups taking advantage of the market transition that AI represents to deliver new types of value to customers - not because AI renders competitive advantage obsolete.

New AI-native software startups like Glean, Rilla, Harvey, and Synthesia will create new categories, grow quickly, and build castle-worthy businesses. But their durability will rely on the same software moats that have existed for the last 20+ years: distribution, brand, product architecture, and culture/processes.

Until of course, the same way that AI-native startups are creating competition for cloud-native incumbents, new market shifts will disrupt the AI-native generation of software startups. My guess is we’ll be waiting until the late 2030s or even 2040s for that. Until then, AI is going to be an exciting and rewarding wave for founders and investors to ride in the decade to come.

Bottom Line

Moats protect castles. Startups, especially at the seed stage, aren’t castles yet. They’re more like open fields. Early-stage founders should prioritize the one reliable competitive advantage they have over incumbents: speed. Build the castle as quickly as possible.

Yet even the sturdiest castle can crumble if the ground shifts. No matter how wide, a moat can erode if customer needs change, new distribution channels open up, or innovation changes what’s possible. IBM’s mainframe dominance didn’t stop Microsoft from capturing the PC wave, Oracle’s database dominance didn’t stop Snowflake’s cloud-native data warehouse, and many companies that rose to dominance during the cloud era over the last 15 years will cede market share to AI-native startups.

AI represents that next major market shift that will create an opportunity for startups to reshape entire industries. The field has never been more open for building castles.

1. If you really want to nerd out on defining competitive advantage, check out Michael Mauboussin’s Measuring the Moat. He has been one of the best research analysts on Wall Street for over 20 years. Mauboussin’s definition of a moat, or a durable competitive advantage, is a business’s ability to consistently earn a return on capital (ROIC) that is above its cost of capital (WACC), with the positive spread between ROIC-WACC being persistently larger than its competitors.