AI companies are growing faster than we’ve ever seen. But rapid growth doesn’t guarantee lasting success.

So how do we separate the companies building toward category leadership from those headed for a wall?

It’s going to come down to long-term defensibility. Along with scale, embedding, and brand, network effects have been a dominant method for building defensibility for startups in the last ten years.

But AI has changed the game. It’s never been easier to start a company and build a product. As a result, some of the “typical” defensibility strategies are too slow to compete in this era of rapid AI scaling.

The answer: you need multiple short-term and long-term defensibility strategies. Speed to scale is your primary lever, but you have to be constantly building for the future. Network effects will play a key role in what companies survive this AI boom and become dominant, but you need to deploy them when the time is right. As part of a long-term battle strategy.

One way to think about it: Your startup should be like a motte-and-bailey castle. In medieval warfare, a motte-and-bailey castle had two distinct defensive positions: the bailey – a large, easily accessible courtyard where daily business happened and initial battles were fought – and the motte – a heavily fortified tower on a hill where defenders could retreat when the bailey was overrun. The bailey was designed to be abandoned when necessary; the motte was built to be impregnable.

For AI startups, your “bailey” consists of the fast-deploying defensibilities that establish your market position: superior distribution, rapid scaling, and brand momentum. These get you in the game quickly, but won’t hold forever against determined competitors. Your “motte” is where you retreat as competition intensifies: true network effects, deep workflow embedding, and systematic lock-in that become nearly impossible to dislodge.

The key is knowing when to fight in the bailey and when to build the motte.

The Core Defensibilities in the AI Era

The core defensibilities we see in the AI era today are:

1. Network Effects: each additional user makes the product better for existing users.

Today’s network effects are just as powerful, although less obvious than in years past. For example, ChatGPT appears to be a single-player game (it’s just you and the AI, after all), but a deeper look reveals a litany of multiplayer functionalities. For example, every user who converses with the AI improves the service for the additional user. Or, the more memory your AI co-pilot builds alongside you, the more personal utility you derive from that co-pilot – an example of a personal utility network effect.

2. To some extent, data moats: Access to large scale or proprietary data will give you an early edge, as you can use this to build higher performing AI models. That said, there are clear limitations to data network effects as we’ve mentioned before. (A key differentiator is real-time data.)

3. Distribution: Early stage companies with potential for long-term growth are experts at today’s distribution game. Cursor, Lovable, Clay – these companies have mastered modern distribution and can use that momentum to catapult into higher layers of the defensibility map.

4. Brand: Once considered a weaker defensibility, brand has become paramount. Because many products have similar functionality and concerns around hallucinations and data privacy, creating a brand (which is inevitably linked to excellent distribution and product quality) helps set companies apart. There is a reason some people prefer ChatGPT to Claude or Claude to Grok.

5. Scale: In today’s terms, scale is often synonymous with systems for computing, and collecting more data. AI companies that are able to scale up compute are likely to increase their flywheel in terms of product, performance, etc.

6. Embedding: Historically, embedding has been the purview of large incumbents. This is still true: incumbents are now embedding AI into their products to leverage distribution advantages.

Interestingly, we’ve seen startups borrow this mindset and embed their own functionality into existing workflows and become incredibly sticky that way.

For example, NFX backed legal company Evenup has embedded their demand letter writing services into the workflow for personal injury lawyers. Embedding into a workflow is becoming a strong tactic for many AI startups.

The truth is that startups at very early stage have very few defensibilities at all – rather the defensibilities are layered over time. The question is when to deploy each one. (Some notable exceptions are in areas like techbio, or deeptech where IP moats create immediate defensibility.)

First, build your bailey, then your motte. This will maximize your speed early on, allowing you to gain the necessary resources to move toward deeper defensibilities.

For example, distribution has become a major source of rapid, early growth for companies scaling from seed to series A. (We’ve written about that here.) Rapid growth itself becomes a defensibility at early stages as you attract better investors, better talent, and you build a flywheel around your company.

But as you graduate from Series A into Series B or C, the underlying mechanics of your growth become important. Now, you’re competing against companies that have all successfully grown to this stage.

Then, we see tools like network effects come into play. It’s part of a larger defensibility cascade. Just when your competitor thinks they’re catching up, you deploy the next round of defense.

The Motte-and-Bailey Strategy: Sequential Defensibility in Periods of High Competition

A layered defensibility strategy has worked before. In fact, if you look back at similar periods where there is high competition and a technology platform shift, this is how most defensibility strategies are built.

Let’s take the birth of search for example. Google’s journey perfectly illustrates how defensibilities should be deployed sequentially.

In 1996, when Larry Page and Sergey Brin developed the PageRank algorithm, Google’s initial advantage was pure velocity – building a superior data and algorithmic foundation faster than competitors could replicate. Search as a concept already existed – Yahoo manually categorized websites and other search engines used simple keyword matching. But the PageRank algorithm was vastly superior, and their minimalistic design approach helped them differentiate from the cluttered portals of the time.

The approach paid off: Google grew from processing 10,000 searches per day in 1998 to over 42 million searches per day by 2000. That rapid scaling deepened their core data advantage, which Google later deployed into distribution and network effects strategies.

By 2000, they launched Google AdWords, which became crucial in funding further innovations while creating a distribution flywheel.

True network effects emerged later – Google’s search engine service improves as more searches are conducted by users and as websites optimize themselves to figure prominently in Google search results. Furthermore, their enormous consumer demand and technical capability enabled them to incorporate an auction model into their ad products and build a highly valuable and defensible online advertising marketplace.

Finally came systematic embedding across the internet ecosystem through distribution partnerships and then product innovations such as AdSense, Gmail, Maps, and eventually Android and Chrome.

This disciplined sequencing – data foundation first, then distribution, then network effects, then embedding – created nearly insurmountable advantages that allowed Google to consistently retain 90%+ of the global search engine market share.

Counterexamples are also abundant. Companies that fail to transition to motte building tend to die. Groupon is an example of one of those companies.

Back when Groupon launched, they were a tech darling. They reached a billion dollar valuation in ~16 months, and famously rejected a $6B acquisition offer from Google.

Their model seemed brilliant: aggregate demand for local businesses, offer steep, viral discounts once enough people “bought in,” and take a cut. They moved fast out of the gate. Well-executed distribution can get you very far.

But they never had true network effects and had weak retention. Groupon felt like a network effect—every additional user made deals more likely to happen. But Groupon’s users had no affinity for one another. They had little loyalty to Groupon itself. And they had only fleeting interest in the businesses they visited through deals. Merchants complained that Groupon customers rarely returned at full price.

What looked like a network effect was actually just virality disguised — cash incentives that attracted users fast, but never gave them reasons to stay.

The difference between Google and Groupon is that they failed to nail the transition points. The short-term strategy eclipsed the long-term.

In your early days, focus on the areas of our defensibility that allow for maximum speed. Once you start to see signs of traction or early PMF, or people begin to copy you, begin to optimize for long-term defensibility channels.

Frameworks for Evaluating AI Network Effects

As many of the AI-native network effects are still developing, these frameworks may help identify emergent strategies.

The Switching Cost Test

Question: “What happens if I stop using this product?”

• Weak answer: “I’ll use another similar tool”
• Strong answer: “I’ll lose months of team context, shared workflows, and collaborative relationships”

The Collaborative Value Test

Question: “Is this product more valuable when others use it too?”

• Weak answer: “It works the same regardless”
• Strong answer: “It’s significantly better when my team/community uses it”

The Hub-and-Spoke Test

Question: “Do users interact with each other through the product?”

• Weak answer: “Users can share outputs if they want”
• Strong answer: “Users regularly collaborate, build on each other’s work, form relationships or develop followings on-platform.”

Emergent Network Effects Strategies for the AI Era

Many AI startups at the application layer are still at early stages – so many have yet to pivot to network-based strategies and are plowing along at full speed.

That said, we are beginning to see some companies turn the corner toward lasting defensibilities. A few themes are emerging:

1. Collaborative Context + Memory = Personal Utility Network

Memory within AI systems – whether that’s from a user’s interaction with AI in single-player mode or from embedding within a company’s workflow – has become a common form of defensibility.

Cursor exemplifies this pattern. While it started as an AI code completion tool, Cursor’s real defensibility comes from learning team patterns and shared context. When your entire engineering team uses Cursor, the AI understands your team’s codebase architecture, coding conventions, and project structures.

A developer trying to switch away doesn’t just lose a coding assistant; they lose months of accumulated team knowledge embedded in the AI. The context compounds across team members, making the tool exponentially more valuable as adoption spreads through the engineering organization. This creates switching costs that multiply with team adoption.

2. AI-Native Distribution = Hub-and-Spoke Network Effects

In 2022, we defined a network effect called: Hub-and-Spoke. These occur when nodes submit content or goods to a central hub, and that hub then pushes a chosen few pieces out to all – or nearly all – of the nodes.

That on-platform distribution drive tremendous attention and value to those few lucky nodes, asymmetrically benefiting them relative to others in the network, creating an internal power law. Winners gain followers, income, status, love…the stuff of great networks.

This is the network tactic behind TikTok, or early Medium. Creators could build loyal followings on the platform, and every user has that sense that they might ascend to the top of the power law.

In the AI era, these effects have evolved beyond traditional social media platforms into sophisticated AI-native experiences.

Character.ai seems to be a powerful early example. Anyone visiting the Character.AI website, or app, is greeted by a large number of chatbots, but a curated few have an extremely high number of conversations – Harry Potter, for example, has had 34 million chats to Albert Einstein’s respectable, but far lower 570,000.

Character.AI perfects the hub-and-spoke model for AI. Users can create “characters”, craft their “personalities,” set specific parameters, and then publish them to the community for others to chat with. The magic happens in the promotion: while thousands of creators submit characters daily, only a select few get elevated to the platform’s featured sections, trending lists, and recommendation algorithms.

When your character gets “pushed” by Character.AI’s hub, the rewards are dramatic. A featured character can jump from hundreds to millions of conversations almost overnight.

3. On the Horizon: AI-Agent Networks

We have been documenting the rise of AI Agents for the past two years. Eventually, a great deal of the value in AI Agents themselves will be in how well they network with one another.

AI Agents will become infinitely more powerful when connected than they are in isolation. For example, your scheduling agent communicates with your travel agent to automatically book flights around confirmed meetings, or your financial agent coordinates with your procurement agent to automatically optimize company spending based on budget constraints.

These agents will develop some known network effects like:

1. Cross-Agent Communication Networks The more agents that can “speak” the same protocols, the more valuable each individual agent becomes. Early examples are emerging in frameworks like LangChain and AutoGPT, where agents can pass context and tasks between each other. This might be similar to how

2. Shared Action Repertoires Agents that can leverage a common library of “actions” (MCP, API calls, tool integrations, workflow templates) become exponentially more useful as that library grows. Each new integration benefits all agents in the ecosystem.

But many more will emerge as we begin to understand agents more fully.

The AI Network Effects Era is Imminent

In 2017, we estimated that 70% of the value in the tech ecosystem could be explained by network effects. But most of the companies we were evaluating, at the time, were the newly established digital incumbents. They had layered network effects into their strategies.

Today, there’s been a lot of talk about whether network effects are dead in the age of AI. A more plausible explanation is that we don’t have the new set of AI-native application layer incumbents yet.

The AI application ecosystem is still in the rapid speed and rapid scale era. We are only just beginning to turn the corner into more durable defensibilities, like network effects.

In the next few years, the network effects of the AI era will crystalize. They will define which companies are this decade’s Googles and which are this decade’s Groupons.

If you want to be the former, start building your roadmap now. Don’t sacrifice speed, but be prepared to move up the defensibility matrix when the time comes.