The AI Business Stack

Most observers probably expect that artificial intelligence will produce some combination of new features for existing functions (image processing, speech to text, language translation, summarization, search, content creation, editing, shopping, research) as well some new use cases that could well produce entirely-new firms, industries and revenue streams. 

It is logical that AI will be used to improve existing products such as Apple’s Siri or Google search, by making existing functions “smarter and faster.”

But entirely-new firms and industries are more likely to be built, one might argue, in a different way, using AI agents, which arguably are better at tasks where lots of unstructured data and relationships between data are involved. 

source: Ark Invest 

To use an analogy, think of the difference between random access memory and sequential access memory. In the early days of the personal computer, that difference was between disk drives and tape drives. 

Or think of programmed processors such as application-specific integrated circuits or field-programmable gate arrays, versus general-purpose central processing units. 

In the pre-recorded music use case, think of tape drives versus compact discs or streaming delivery. While it is possible to pick a single song off a tape, there is substantial winding or rewinding time. With CD or streaming access, there is little to no navigation required, to say nothing of the “discovery of content” function. 

Other similar analogies are possible. Think of live or pre-recorded linear television (broadcast or streamed) versus on-demand video, where one experience is scheduled and linear; the other built on-demand access allowing users to skip around a catalog to make choices. 

The point is that AI agents might resemble general-purpose CPUS instead of ASICs; random access memory; on-demand audio or video: allowing unstructured operations using unstructured and structured data for unexpected tasks. 

And since investors are so focused on AI opportunities, it might be helpful to liken various AI functions to the standard software “stack.”

OSI Layer	OSI Model Functions	AI Model Functions
Physical Layer	Physical transmission medium (cables, etc.)	Hardware Layer: Physical components like CPUs, GPUs, TPUs that perform AI computations.
Data Link Layer	Packet transmission and error detection	Data Layer: Management of data used to train and operate AI models. This includes data pre-processing, cleaning, and formatting.
Network Layer	Routing data packets across networks	System Software Layer: AI frameworks like TensorFlow or PyTorch that provide the structure and tools for building and training models.
Transport Layer	Reliable data transfer between applications	Machine Learning Layer: Machine learning algorithms that power AI models. This includes algorithms for tasks like classification, regression, and natural language processing.
Session Layer	Establishes, manages, and terminates sessions between applications	Model Training and Optimization Layer: The process of training and optimizing the AI model using the chosen algorithms and data.
Presentation Layer	Data presentation and encryption	Model Deployment and Inference Layer: Deploying the trained model and using it to make predictions or decisions on new data.
Application Layer	Provides user interface and network services	AI Application Layer: The actual application where users interact with the AI, such as a virtual assistant, recommendation system, or self-driving car.

Broadly speaking, investment opportunities will occur at every level of the AI stack.