AI Applications and Vertical Integration

AI application companies are gradually evolving into a full-stack form. AI products largely consist of three layers: the Model layer, the Application/Agent layer, and the Human/Service layer that completes the final output.

While traditional companies stayed in the middle Application layer, recent companies are attempting Vertical Integration in two directions.

The first is the 'Full stack down' approach, moving down into the model layer. This is prominent in fields like coding and customer service, and occurs when a company secures sufficient usage and unique Traces data.

• Cursor: Recently introduced Composer 2, applying fine-tuning and reinforcement learning specialized for coding tasks based on Kimi K2.5.
• Intercom: Through Fin Apex, handles most English-language chat and email customer support with its own model.

The key driver of this downward integration is the Flywheel effect. This is because data such as agent prompts, outputs, and corrections can be used to continuously improve the model. Additionally, using small models optimized for specific purposes reduces cost (COGS), increases speed, and helps achieve differentiation from other services.

The second is the 'Full stack up' approach, moving up into the service layer.