Generative AI is having a moment just as the year wraps up. With a recent technology and crypto crash, it is a surprise to finish the year with any industry optimism. However, Generative AI has combated market pessimism with a surge of eye-catching model releases by various tech companies. With as simple as a few words entered into a text prompt, users are able to generate professional-quality images, video, and audio. This type of production quality was not possible only a year ago. It may come as no surprise that this technological magic has caught the eye of many entrepreneurs who have seized the opportunity to launch businesses on top of these models. From generative copywriting to interior design to illustrated profile pictures, companies tackling common creative problems have sprung up and raised funding in just this past year. It is an exciting time to watch generative AI take the main stage, but I wonder how most of these companies will sustain growth without having a moat.
For those reading who are not familiar with the economic definition of a “moat”, it is used when evaluating the prospect of a business and metaphorically represents the business’ key competitive advantage over its competitors. In generative AI, a competitive moat would be the businesses that are providing their models for users to build on top of (e.g. DALL-E, Stability.ai). These models are extremely complex and resource-intensive, costing businesses millions of dollars and years to train. Given the high cost of entry and easy accessibility for developers to build on top of, it is difficult for new competitors to spring up and reach parity unless they have the money and resources to do so. This moat does not carry over to the businesses that have built their product on top of these platforms.
These models are proprietary software and their access can be taken away as easily as they are available to end users. In order to use these models you have to abide by the terms of service and regulations that are in place. Any violation could lead to suspension or termination of their access to the model and ultimately failure of the business. It also means that most of your business value is tied up in software that you don’t own. Ultimately when you look to scale your business, most investors will view this point as a red flag and turn down the opportunity to invest.
Another limitation is the lack of domain knowledge for these models. I’m sure there are many applications built on top of the models that have talented data scientists and AI thought leaders that understand how the models generate the work they are selling, but for most, the products are run by small engineering and product teams that depend on model magic, rather than understanding the secret sauce. Without understanding how these models work, there is a heavy dependence on the tool working as anticipated. In reality, that is never the case and when something relies on statistical modeling, all it takes is an undetected anomaly to make it fail. I can understand using these models as an MVP to secure funding, but ultimately you will need to develop your own proprietary system if the model is your application's primary feature which is not easy.
Come One, Come All
These models have a finite amount of functionality and customization at the moment. This means that anyone building on top of the models will have an easy ability to recreate popular features on other applications. Unless you tweak hyperparameters and feed the model with better data, then the application is just a shell surrounding the model. It is a rat race to be the market leader with the likely winner the one with the biggest pockets and best UX.
Who Owns the Training Model Data?
Finally, there is a grey area hanging over the whole Generative AI space at the moment around copyright and intellectual property. A lot of these models have been built on top of copyrighted material scraped from the internet. Without permission from the owners of the content, there are a ton of copyright infringement cases that are lurking in the shadows. It is too early to tell when and how governments will address this, but the threat will always be present until there is more clarity around the legality of the data training the models.
This is not to say that these businesses are bound to fail without a moat, but they all certainly face an uphill battle. If the model is the main feature, competitors will always be around the corner building to reach feature parity or overtake the incumbent with a better product. A moat creates the separation between your business and competitors and allows you to control your own destiny without fear of what the competition could be building. At the moment I feel like this is lost with most generative AI startups.