As large-language models become more widespread, who captures most of the value from these products? A brief look at some possibilities.
Incumbents that can layer in generative AI as a feature to existing application distribution.
The most obvious winners are incumbents that can leverage their existing distribution to enhance their products. Notion's new Notion AI. Microsoft Word's writing assistant super-powered by AI suggestions. Facets of Google Search. The incumbents are not only aware of the advances in AI, but are driving them and can afford the most R&D spend.
- Will the best use cases of LLMs fit cleanly into existing products or require new experiences? While not counterpositioned, if applications need to significantly change their product to accomodate generative AI solutions, they might falter. For example, building image generation into Photoshop vs. a new, specialized product.
- Does generative AI challenge the existing business model, e.g., stock photography?
The first layer of "selling shovels". Model training and inference require significant amounts of specialized hardware – large machines and cutting-edge GPUs.
- The more that a startup spends on hardware, the more they are likely to figure out ways to repatriate the most expensive parts of their infrastructure. You can view this two ways – one, that hardware providers have a marginal advantage because they can use subsidized infrastructure from their own cloud, or two, that startups can selectively choose cheap cloud services to scale and then repatriate later on or selectively.
- It's actually not that expensive. LLMs are getting cheaper to train and cheaper to serve. In the Commoditization of LLMs (part 1 and part 2), I showed just how cheap and easy it is to train these models from scratch. If the majority of startups use off-the-shelf open-source models or fine-tune an existing model, the cost will be even cheaper.
OpenAI, StabilityAI, Midjourney, and the numerous inference-as-an-API companies that will spring up are well positioned to capture value. Usage is already skyrocketing, and usage-based APIs are a well-understood business model.
- There's already been pricing pressure at the infrastructure level. Revenue lags behind usage. When the model is open-source and fairly easy to run behind an API, margins shrink.
- Infrastructure companies don't have a relationship with the end user. Distribution often matters more. Companies at scale can eventually switch out the underlying model – defecting to a cheaper solution or building their own in-house.
Like many of the developer-focused companies building platforms on AWS, there's value in the user experience. Platform companies can combine raw inference APIs into more useful building blocks.
- How will they differentiate from the API layer?
- Can margins be high enough between the cloud tax (30%) and model tax (OpenAI, etc.)?
While generative AI can solve a variety of problems, many industries will need tailored solutions for their workflows.
- Will the markets be big enough?
- If they are big enough, can they differentiate enough from generalized solutions?