If you’re building an AI-business in 2024, your options to protect it are fairly limited:
- play the short game or the long term, but not the mid-game
- fall in love with the problem vs. the AI solution
- choose a hard / complex / ugly problem
- if you’re in the application layer, build a data feedback loop
Every AI investor is asking today: Is your business defensible? Can it be easily copied? What’s your moat?
The truth is that many AI businesses can be copied in the short term – especially software businesses.
- Your AI MVP can be copied in weeks.
- Even the AI developer tools are not safe. The recent “AI programmer” Devin has an open-source alternative that’s keeping up.
- On the foundation model front, GPT-4 has many open-source alternatives, like Llama and Mistral.
- There are plenty of Cloud Infra alternatives which offer access to GPU machines. Btw, I particularly like the server-less GPU ones. Still wondering when AWS will offer this…
- Even AI hardware like Nvidia’s isn’t safe. It’s being competed by AMD, Apple, Groq and I suspect AWS. The Nvidia profits are a hardware startup’s opportunity.
All this dynamic is great for the end customers, but not so much for investors or entrepreneurs. If what you are building can be copied, why waste your money and time?
I’ve been thinking about this over the past days and I don’t have perfect answers. So, this article is a “think in public” experiment – not investment advice. I will update it over the next weeks.
I think we all know that great distribution, capital and IP are great moats. But let’s assume you don’t have them. What’s left?
The good news is that strategic considerations for AI businesses don’t differ significantly from other tech-based ventures. The bad news is that strategic considerations for AI businesses don’t differ significantly from other tech-based ventures…
I keep coming back to 4 generally applicable answers. Obviously, any business will have specific insights for the market / niche they are operating in, but what I’m focusing on is the ones applicable to the majority.
The first one is long-term commitment (and execution)
- If you play the long game, attrition takes care of itself Many people have ideas. Only a fraction will do something about them. Of the ones that start implementing them, even fewer go beyond a demo / MVP. Many people and companies will give up sooner than you think. Once the AI novelty fades away, people jump onto the next shiny thing. Opportunists wanting a quick exit will either succeed and exit the arena, or give up and move on. There’s also natural attrition when things get tough. And things do get tough. Customer support is tough. Reacting to new tech changes is hard and frustrating.
- A higher likelihood of finding gems The longer you stay in the game, the more likely you’ll find more clues on how to build the moat. The answers will show up (e.g. a unique feature, a strategic channel, a surprising insight, a genius colleague).
- Consistency is what builds a brand If you execute exceptionally, keep the eye on the ball, and show up every day, you should have built a solid brand and market recognition to have customers choosing with you. This is more important for B2B than B2C. Enterprise customers want assurance you will be there to support them when things get tough vs. a unique feature you’ve developed. Some plans are measured in centuries 🙂
- If you don’t want to play the long game, you have speed to market as an advantage and then a fast exit. You’d be surprised how much of an advantage this is. While you ship 2 key features in a week, other companies are trying to find a slot to discuss their Q3 roadmap.
The 2nd one is loving the problem vs. obsessing over the tech / solution.
- Customers don’t really care if you are using an LLM or a regex to solve their problem. If you care more about the problem, you will use any tool at your disposal to solve it, vs. build the most complex AI agent chain.
- Marinating yourself in the problem you will help develop an X-ray vision for details generic “AI experts” will miss. This has staying power and your users will empathize more with you vs. your competitors. That’s because you understand them; you think of them more than you think about quantizing the next Llama model. And this in itself leads to… a stronger brand and increases stickiness.
The 3rd one is choosing a hard problem to solve. If not that, then choose a simple problem, but aim to build a ridiculously great solution for it. 10x better on at least 2 dimensions.
- You will have many competitors for easy problems. But way fewer if you set a ridiculous ambitious objective.
- In other words, complexity is a natural moat. If you love solving complex puzzles and are ready for the challenge, you’ll be a rarefied field.
For these 3 strategies, my go-to example here is Superhuman – an email client. I consider Superhuman’s founder – Rahul Vohra an outstanding entrepreneur and a great thinker. Everything he speaks of, publishes, or ships is high quality.
Superhuman should not exist, yet I am a customer!
- Superhuman wanted to build a better email experience, where users will get to see Inbox Zero for the first time in years. Every interaction is supposed to be under 100ms. And for a price is … $30 / month. And they started with Gmail!
- No investors should have put money into this… It has no intrinsic moat at all. It started as Chrome app over a free email tool like Gmail! The product took 2-3 years to be released. It was expensive. All cardinal mistakes. Or were they?
After taking their time, executing flawlessly, Superhuman built an amazing moat and likely a >$1B in the process through:
- long-term thinking Rahul Vohra didn’t launch an MVP immediately. He took his time for years and years perfecting a great solution. The first layer of the moat was exceptional design. The product was a premium one, not a rushed demo. Many would have given up and investors would have pressured them to ship fast. Yet he didn’t.
- loving the problem more than the tech In being obsessed with the problem, Rahul’s focus was somewhat surprising. He understood that speed was key. As such Cmd+k shortcuts were introduced. Emails were downloaded in the browser so the search was lightning fast. And the list goes on.
- choosing a hard / complex / ugly problem solving email has always been a schlep… It was even flagged by Paul Graham for years, and no-one touched the problem.
While these first 3 strategies are generic, this fourth one is AI specific. And mostly applicable to AI apps – not AI models and AI chips. It is data feedback loops.
Data is the key ingredient to building AI models and apps. If you don’t have a unique (or defensible) data set, you have to start building one immediately. And the most elegant way to build one is through your AI app.
The customer interactions should make your product smarter, and more valuable to them. And as such, get users to use your product more, creating a virtuous circle.
Some examples are:
- TikTok’s “For You” page recommender (but maybe less aggressive).
- Nest’s thermostat learns how to set the temperature through repeated use
- Chat GPTs interactions are used to improve the model answers.
- A translation tool learns from the corrections made by professional linguists and ideally avoids making the same mistakes in the future
- Tesla auto-pilot learns from the millions of kilometers driven by the drivers
Unfortunately, building a data feedback loop is very hard to implement, though not impossible.
- You need a product / app that a lot of people want to use, of a few people want to use a lot. This means you need to understand the problem space really well.
- You might have to delay monetization and raise too many barriers to adoption. That means playing the long game.
- There has to be an incentive for users to engage with your product and provide you the data. A great UX is absolutely necessary, but not sufficient.
- Data privacy or data sensitivity is key. In some industries this will be very hard to get right (e.g. life sciences, health apps).
- The tech needs to be designed in such a way that the improvements are visible to the user. Ideally you should not wait months to retrain the model. By that time the users might lose interest.
So all in all:
- long term thinking
- problem focus
- choose a complex / ugly problem
- data feedback loops
Let me know what you think. How are you thinking about building your moat?
