The rise of big companies has been, in part at least, attributed to the network effects that their products have built over time — where a product becomes incrementally useful as more users join the network. At its core, the theory suggests that every marginal user creates value for every other user and provides enough data to improve the product. This idea of network effects questions the company’s knowledge about its value propositions and how the dynamic could shift as the company evolves. 

A company will have positive network effects if it manages to incent new users and disincentivise contaminants in the network.

In this context, people, and many times, founders, tend to conflate network effects with viral effects. Viral effects are about increasing the user base, where one user brings in more users (preferably for free), while network effects are about building defensibility. It is about creating value to the end-user to build loyalty and be insulated from insurgent competition. 

Source: stories.platformdesigntoolkit.com

Traditionally, items like preferred trade agreements, access to unique raw materials, geopolitical elements, and captively generated assets were counted as highly defensible elements for any organisation. With the world going more digital and duplication becoming easier, and the replication cost becoming lower than that of the incumbents, traditional moats nearly cease to be defensible elements. 

In software, for example, geopolitical and demographic barriers scarcely matter. It is easy to reach new markets because the same infrastructure required that the government or any third party would build can be used for myriad other ‘development and welfare’ activities. 

Effectively, we’ll be left with just a small number of competitive advantages that truly provide unique defensibilities to an organisation. It is important not to confuse competitive advantages with defensibilities. As NFX rightly quoted, “Competitive advantages make your business successful. Defensibilities help you stay there.”

So what are those defensibilities relevant to today’s ecosystem? NFX and A16Z both say the same 

• Economies of Scale
• Branding
• Embedding
• Network Effects

We will explore them more a little later.

Hardware Doesn’t Capture the Majority of Value. The Applications Built on Top of Them Do

Asymmetric and Asymptotic Effects

Laws like the works by Sarnoff, Metcalfe, and Reed help us understand how networks behave. It would be a fallacy to assume that these laws are immutable because, in real life, networks are far more complex than when the laws were proposed. Networks today, in the digital world, are far-reaching, all-pervading, and intensely complex and messy. Finding symmetric relationships between the networks is rarely easy and comfortable. More often than not, there are multiple laws present across a network. 

These asymmetric networks are what create economies of scale, thus creating value defensibility across the entire network. This improved perception of value is key because it reduces customer onboarding cost over time for all products that rely on network effects. Take, for example, Uber’s product. The more users on the platform, the greater will be the amount of data generated. This will help in calculating better routes and decreased wait time for all the users. 

Zhu, Vondrick, Fowlkes, and Ramanan’s paper titled, “Do We Need More Training Data” is a good case study in that regard. 

But it is important to realise that data is asymptotic. Save for a very few companies and business models, data accretion beyond a point contributes to a marginal increase in value production. Take the case of a hospital that uses AI to improve its diagnosis and detection algorithms. Having the 20000th case might not contribute to the system as much as the 2499th case might. Essentially, the size of the firm doesn’t matter here. This means, the competitors could claim the same value generation, and they wouldn’t be wrong.

This is an accepted hazard of work in today’s digital age. Today, it is hard to develop products that are so revolutionary that the company that came up with that can milk it for a considerable period. With increased competition, customer acquisition costs increase once the business reaches an asymptote, and technology gets replicated easier and better by those that learnt from the incumbents’ mistakes.

Where the data starts being an asymptote varies on the industry. Take search engines and applications that rely on real-time data like Uber, for example. The more data they have, the better their service becomes. This acts as defensibility because the amount of data that needs to be collected by the competitors to offer competing products is just too voluminous to be feasible. 

But data collection network effects, by themselves, are asymptotic too. It is important to understand that the app or the website only channels virality. The value creation and precipitation depend on the users. More often than not, it is the users who generate the network effects that run a platform or a product.

Once a system starts to grow itself, it is hard to predict the exact value generation and its proportion to the incremental data collected. Sometimes, the 5th review on TripAdvisor could add more value than the 50th. Besides, product usage and new and useful data generated are hardly synchronous with the perceived value addition through the process’s systemic growth. 

Consider that only a handful of people who used a service leave a review on Yelp or TripAdvisor. Only a few people let the downloaded file seed on torrents. Yelp, TripAdvisor, Torrent platforms are left to generate data just from the available lot — vastly under-indicating the platform’s total usage or the product.

Sometimes Increasing Sophistication Contributes to Better NFX Than More Data

The Cold Start Problem

It is worth noting that articles by websites like NFX, A16Z, and Platform Design Toolkit distinguish between data network effects and network effects. This article chooses to treat that interchangeably because this article only tries to understand the impact of network effects. 

To make it clear: take Uber’s case, for example. More users create more value for drivers and customers. That’s how network effects precipitate. More users contribute more data for bettering the product. That’s how data network effects work. It is thus possible to have standard network effects and data network effects working in the same company.

There really is no measure that states the quantum of data required to create network effects. Data network effects work equally well across industries for both businesses and customers. They help in facilitating learning across those industries for both the customers and the businesses. But for those network effects to happen, founders and companies need to commit to building the right infrastructure actively — build data feedback loops that allow the system to learn.

This process has to be done manually first because there isn’t enough dataset to start with. The scalability is low even if companies were to venture out and ask people for their data for amazing things to happen later with the said data. Without an upfront value proposition, that idea just wouldn’t work. Once the right data flows in, the process can be done automatically. In fact, the more the automation, the likelier the network effects. 

But here is where the glitch happens. Once you scale a certain level, finding unique data that contributes value will be more expensive. It becomes more difficult to find useful information from amidst the noise. This is exactly what we were talking about when we discussed the value creation beyond a point taking the example of cases in a hospital. 

The crux of the previous paragraph is just that companies are going to run into a rut beyond a point unless they change some element of their value proposition. This is one reason why Facebook removed the feature of chronological order in showing the newsfeed and started curating content based on engagement statistics. 

An interesting problem here is called the ‘Cold Start Trap’ or the ‘Chicken-Egg Problem.’ Do you build infrastructure first and wait for it to collect data or will you accumulate data first and build the infrastructure once there’s enough traction? There isn’t a blanket answer to that question. It really depends on the kind of industry that the product works in

An e-commerce business offering personalised recommendations based on historical preferences can afford to accumulate data first and build the required infrastructure later. But a company that scours the dark web (legally, of course) for finding datasets, or companies that work a lot of with power-hungry AI/ML models need to build the right infrastructure to create defensible moats*. 

*A moat is a competitive advantage for a company. 

This is where it pays to revisit the basics — Network Effects occur because founders build something and people join because ‘that’s where the customers are and the data exists.’

One way to maintain the defensibility is by making it difficult for the customers to shift either by integrating SDKs into third-party apps (called embedding) or cashing in on the brand value that a company has generated. Coca-Cola’s ‘Secret Recipe’ myths are a great testament to this. 

Some moats that can help a business make its products defensible are

• Deep Tech/IP/Trade Secrets
• High Switching Costs
• Branding/Customer Loyalty. 

Conclusion

The point of every business is to create insulation from replication and duplication. Massive returns only happen if the business is highly defensible, and manages to stay so. Network effects are fast replacing the traditional moats that the pre-internet and early-internet companies have capitalised on.