# Foundation Models: The Next Platform Shift in Artificial Intelligence

Author: Norbert Biedrzycki
Published: 2022-12-15
Language: en
URL: https://norbertbiedrzycki.pl/en/foundation-models-the-next-platform-shift-in-artificial-intelligence/
Summary: A few days ago OpenAI released a chatbot to the public. More than a million people signed up in under a week. Something structural is happening, and investors need a framework for it.

**A few days ago OpenAI released a chatbot to the public. More than a million people signed up in under a week. Something structural is happening, and investors need a framework for it.**

Last year, researchers at Stanford published a long paper introducing a term that is now spreading quickly: foundation models. The idea is simple but powerful. Instead of training a separate machine learning system for each task, you train one very large model on enormous amounts of general data, and then adapt it to many tasks with relatively little additional effort.

GPT-3, Google's PaLM, DeepMind's Chinchilla, Stable Diffusion and now ChatGPT are all examples of this approach. ChatGPT, launched on 30 November, is in many ways a user interface on top of an existing model. Yet its reception shows that the interface was the missing piece. For the first time, millions of non-technical people are interacting directly with a general purpose AI system, asking it to write emails, explain concepts, draft code and summarise documents.

**Why this is a platform shift**

In my career I have lived through several platform shifts: client server, the internet, mobile, cloud. Each one followed a similar pattern. A new layer appeared that made it dramatically cheaper to build certain kinds of applications. Value was then distributed between those who owned the platform, those who controlled distribution to customers, and the applications built on top.

Foundation models have the characteristics of such a layer. A single model can support customer service, content generation, software development, legal research, education and enterprise search. The marginal cost of building a new AI application has dropped sharply, because developers no longer need to collect training data and build models from scratch for every use case.

**The capital question**

But platform shifts are not free. Training a frontier model requires thousands of specialised chips running for weeks, a rare kind of research talent and access to vast amounts of data. Estimates of the cost of training a model like GPT-3 run into millions of dollars for a single training run, and each new generation is considerably larger. This means the platform layer itself may be accessible only to a handful of organisations with access to very deep pockets.

That raises the strategic question I find most interesting. Who will own the value?

One scenario is that independent model developers become the new platform companies, the equivalent of operating system vendors. Another is that value flows primarily to cloud providers, because every model needs compute and the cloud companies own it. A third is that models become commoditised over time, much like databases did, and the real value sits with application companies that own customer relationships and workflow.

Today I cannot tell you which scenario will dominate. I suspect all three will coexist in different parts of the market. But I am confident that the relationships between model developers and cloud providers will be among the most important strategic alliances of the coming years.

**What investors should do now**

For early stage investors, this is a moment of opportunity and danger. Opportunity, because a whole generation of new applications becomes possible. Danger, because many of those applications will be thin wrappers around a model they do not control, with no defensible position once the model provider decides to move up the stack.

For private equity and corporate boards, the more pressing question is how foundation models will change the cost structure and competitive position of existing businesses. If a large share of knowledge work can be assisted by such systems, the implications for professional services, customer support and software development are significant.

*Investment question: Who will own the next generation of AI platforms: model developers, cloud providers or application companies?*
