How Artificial Intelligence Business Model Differs From Traditional Software Companies

Although it seems like Artificial Intelligence ("AI") companies operate under the same economics as traditional software companies, a closer look into how the AI business model works will yield a different conclusion.

The first difference is an economic one - lower gross margin of ~50-60% vs. 60-80% of SaaS companies. Lower gross margins stem from various factors, including the costs associated with enormous compute resources required, higher storage space, and complex cloud operations that demand routine transfer of trained models. AI models need to be trained and re-trained in order to keep up to date with data, which counterintuitively goes stale as time passes. Due to the constant update required in AI model, companies need to spend significant amount of money on compute resources. The raw material that goes into AI model - images, audio, video - are often richer in nature and thus higher storage space. In order to maintain quality reliability, latency, and compliance, AI models are often transferred regularly across cloud regions, which again incurs significant costs.

Unlike traditional software companies that "develop once and deploy forever", AI models are bound to be restricted by "edge cases". AI models are often referred to as a black box, so users often lack intuition where AI models can best be used. It is thus difficult for developers to accurately pinpoint what the intended use of AI models should be, if they do not establish that early on with the customers. Even if they do so, users can easily sidetrack from the original intention of the model to solve issues that reside in long tails, which applies only to limited number of cases.

Combined with lower gross margin and difficulty in scaling, AI companies often lack defensive moat to fend off competitive pressures. AI algorithms are often open source, developed within the academia that cherishes the public nature of AI research, so there is no such thing as proprietary code that software companies leverage to win competition. Not only the algorithms, but the data itself is easily available to companies who would like to obtain it. Even if the data is not directly from customers themselves, the bootstrapping or resampling of public data sets can be used to build an AI model. Also, many software companies benefit from network effect, through which the value of engaging in such software and ensuing stickiness increase over time. However, the value of data often declines, because new data collected will only apply to the small subsets / special use cases. Such commoditization of data and algorithms behind it leads to absence or lesser degree of defensive moats that AI players can have.

Source: https://a16z.com/2020/02/16/the-new-business-of-ai-and-how-its-different-from-traditional-software/