When we hear “network effect” our minds jump to social networks. Instagram. Facebook. Twitter. LinkedIn. Snapchat.
But before we get into that, let’s take a step back and define “network effect”.
If you visit the Wikipedia page for network effect you’re going to be confronted with this: “In economics, a network effect is the phenomenon by which the value or utility a user derives from a good or service depends on the number of users of compatible products.”. We prefer Naval Ravikant’s more lucid definition:
What is a network effect? Let’s just define it precisely. A network effect is when each additional user adds value to the existing user base. Your users themselves are creating some value for the existing users.Naval Ravikant
It’s easy to see how Twitter fits this definition. While Jack Dorsey was sleeping, we summarized this blog as a Twitter thread and you out followers extracted some value from it. Well we hope they did.
It’s a cliché to say, but network effects are strong. They are so strong that movements like #DeleteFacebook had absolutely no effect on Facebook’s daily active users metric (DAU). It’s just “too hard” to leave Facebook.
In our opinion, a far more interesting network effect is being built by many artificial intelligence companies right now. When data collected from one customer is used to train a machine learning model that benefits another customer that is a network effect. And when we consider that the most important factor when building effective machine learning models is the quantity of data you have, we get a golden feedback loop.
More customers means more data which means better machine learning models which means better products and, therefore, more new customers. This is not a novel insight, many VC’s have pointed this out. However, most of them stop here. But how should one go about critically analyzing an AI company’s network effect?
When data from one customer is used to train a model that benefits another you have a network effect. More customers → more data → better models → better products → more customers.Stuart Reid Tweet
There are two important concepts that matter when evaluating an artificial intelligence network effect.
The first concept speaks to how easy it will be for any competitor to catch up whilst the second concept speaks to how fast that feedback loop we described earlier will iterate.
A data moat is the AI equivalent of an “economic moat”. Economic moats refer to the ability of a company to maintain their competitive advantages over long periods of time. Network effects are economic moats as are massive economies of scale. Good economic moats create high barriers to entry for competitors.
A data moat is the ability of an AI company to maintain an edge through their most valuable intangible asset, their data and their AI models. Let’s go through some questions to help you evaluate a data moat.
First off you want to ask how homogeneous their customer data is. If all customers produce exactly the same data at the same time then a company with just 1 customer would be able to compete effectively against another company with 10,000 or a even million customers. Ideally different customers should generate data that is differentiated from each other and that collectively adds value. The sum should be more than the parts.
Next you want to ask how dynamic the data is. Machine learning models solve problems by finding patterns in data. They can only learn how to play chess because patterns emerge from the board, pieces, and rules staying the same game after game. Unfortunately not all games are so well behaved. In some cases an externality “changes the game” and nullifies the value of the old data and any machine learning models trained on it. In such games, every time the externality occurs the data moat is reset and everybody ends up neck and neck again. Sustainable moats can only be built on data that is relatively static. Without that compounding can’t happen.
Next you need to ask yourself how expensive it is to acquire the data. If a competitor can buy tons of it online for thousands of dollars a month, there is no data moat. The more expensive it is to acquire the better because fewer competitors will have enough capital to compete. Similarly, if realistic data can be synthesized using cheap simulators, generative machine learning models, or simple statistical techniques there is no data moat!
All data has a distribution. Narrow distributions have many high-probability events whereas wide distributions have many low-probability events. After the point at which there is enough data to build the product (Minimum Viable Data aka MVD), companies are better off competing in an industry with wide distributions because it takes a much longer time to see diminishing returns on new data. You need to ask how long those tails are.
In industries with wide distributions of data whoever has the most data will dominate because their dataset will be the most representative and, therefore, their models will most probably be the best.
In very extreme industries very wide distributions can, however, backfire. This happens when the minimum amount of data required to build a Minimum Viable Product (MVP) is out of reach. If you can’t get to MVD you will never get to MVP. That is the problem that most self-driving car initiatives are facing today. The biggest problem with AI right now is that nobody knows where MVD is. If we could tell you how to work that out we would be rich.
Lastly you want to ask yourself how imbalanced the data is. Imbalanced datasets are ones with many classes of events with uneven probabilities. Fraud is an incredibly imbalanced dataset because for every 1 million legitimate transactions you may only get 10 fraudulent ones. It is better to compete in an industry with high data imbalances because it takes significantly longer for your competitors to accumulate representative datasets for each class.
Note that in order to build your MVP you need to achieve MVD for every class!
The most defensible data moats consist of heterogeneous data that is imbalanced, relatively static over time, has wide distributions with not-too-long tails, and is expensive as hell to accumulate.Stuart Reid Tweet
The second concept that matters when evaluating an AI network effect is the quality of the company’s distribution channels. It doesn’t help if the data is heterogeneous, imbalanced, relatively static, wide, and expensive if you don’t have any way to actually get any of it. After an AI company has built a product they must distribute it as widely as possible and as fast as possible in order to rapidly expand the width and depth of their data moat.
It should also go without saying that they also need to be able to use that data. If they can’t, then they have bigger problems that should probably have raised red flags earlier in your analysis.
Let’s take a step back and ask how social network effects were created.
How did Facebook beat Myspace and Friendster? Virality. They created a free product that incentivized users to invite more users. That is a great distribution channel. Most AI companies can’t rely on virality so they need to think extra hard about their distribution channels relative to their competitors.
The thinly-veiled purpose of this blog is obviously to demonstrate the value of Nosible to you. To do that we have decided to become investors ourselves. Warren Buffett tells you to stick within your circle of competence. Since ours is tech and AI we will be using this blog to, among other things, build up a portfolio of tech stocks with a focus on AI that excite us. If you don’t want Nosible, that’s fine, stick around anyway if only for the blogs :-).
We seeded our portfolio with Google, Facebook, Nvidia, and Microsoft (the kings of AI) and Nosible recommended 30 other stocks. One of which was Palo Alto Networks. We read up on them and we really liked what we saw.
Palo Alto Networks is a cybersecurity company located in the United States. They’re not a young company, they were founded 16 years ago. Back when you couldn’t just spin up as many servers as you want on Amazon Web Services. The massive public clouds we take for granted today didn’t exist so everybody needed to roll their own infrastructure. Which means they also needed to manage their own networks and security.
In 2007 Palo Alto Networks shipped their first product. The world’s first next generation firewall!
Next generation firewalls don’t block traffic based on port numbers, they inspect the traffic on every open port and block traffic that does not originate from approved software. Over the course of the next decade Palo Alto Networks’ next generation firewalls went on to dominate the enterprise firewall market. Their strongest competitors in the enterprise firewall market currently are Fortinet, Cisco, and Check Point.
The shift from private clouds to public clouds picked up in earnest for large corporates in 2015. The Right Scale 2015 State of the Cloud Report pegged just 26% of respondents as being “cloud-focused heavy users” and a further 26% being “cloud explorers”. As of the 2020 report those numbers sit at 53% and 30% respectively. The same reports show that the % of enterprise respondents using AWS has grown from 50% to 76% and a much higher percentage of those users are “heavy users” running all or most of their company in the cloud.
This was not good news for Palo Alto Networks. Whilst their enterprise firewall business continued to grow they were losing market share. It took them until 2017 to start responding to the threat of cloud-native solutions such as Secure Access Service Edge (SASE pronounced Sassy – how funny is that pronunciation?) products.
In their most recent earnings call Palo Alto Networks disclosed that over the past two years they have split themselves into two businesses. The first business is their traditional enterprise firewall business. The second is their cloud and AI business. The traditional enterprise firewall business did revenues of $3bn in FY 2020 and $3.5bn in FY 2021 (a 14% YoY growth). The cloud and AI business, on the other hand, did revenues of $318mn in FY 2020 and $605mn in FY 2021 (a 90% YoY growth). That’s phenomenal for an incumbent.
This shift has been driven by focused internal efforts to move to containerized, cloud-native, one-click deployable SaaS products as well as by acquisition. Palo Alto Networks has a strong balance sheet to support further growth by acquisition and it’s clear why startups would want to be acquired by Palo Alto Networks: distribution.
This article is about distribution. The cheapest distribution channels are direct to consumer. And what direct to consumer distribution channel is better than your existing customer base? This is the position Palo Alto Network finds itself in. They boast >70,000 customers in their enterprise firewall business many of whom will be wanting to make the shift to public clouds. This channel combined with the AI network effect is formidable.
We also believe that the data moat collected by Palo Alto Networks would be defensible in the long term because it consists of heterogeneous data (all customer networks are different) that is imbalanced (legitimate network activity far outweighs malicious network activity), relatively static (attack vectors do evolve but not so rapidly that it would nullify historical data), has wide distributions (security is all about edge cases), and is expensive.
Other incumbents including Cisco, Fortinet, and Check Point do not have clear strategies to compete with newer players like Zscaler and CyberArk. This is reassuring for both Palo Alto Networks and the new players because all three of those incumbents have their own massive customer bases that they could choose to target.
Lastly, when we zoom out to the macro level and look at the world we believe that cybersecurity and, in particular, cloud-based cybersecurity, is a massive growth industry for two reasons.
Cyber-attacks are on the rise. Nation state sponsored hacking groups are a reality. Businesses and governments have to respond by upping their game in the years to come or we will see many more devastating attacks.
The changes to the way we work during the COVID pandemic are likely to stay in place afterwards. Remote work affords many benefits to businesses including the ability to hire talent globally, distribute risk across geographic regions, and reduce fixed assets and costs. That said, remote work massively increases the attack surface for hackers. So Secure Access Service Edge products are almost certainly going to see a lot of growth.
In conclusion, when data collected from one customer is used to train a machine learning model that benefits another customer you have a network effect. More customers = more data = better models = better products = more customers. The data moat accumulated by cybersecurity AI companies looks like it is defensible and Palo Alto Networks has a very strong distribution channel to acquire this data. That distribution channel is their existing customer base. This leads us to believe that they can create a robust network effect.
That having been said, history has not been kind to on-prem incumbents who failed to move to the cloud fast enough. Salesforce crushed PeopleSoft. So the only question that remains is whether it is all far too little too late, or if Palo Alto Networks can move fast enough to secure itself another decade as an industry leader?