Build Your Own Brain? Vertical Models and the Post-Training Moat

For years, the prevailing view among software companies was that competitive advantage came from the product, data and customer experience built around AI, rather than the underlying model itself. That assumption is now being challenged as AI-native software providers explore post-training open-weight models on proprietary interaction data to deliver better performance, lower latency and reduced costs than frontier alternatives.

This roundtable will examine whether developing a specialised model can create a genuine competitive moat for independent software vendors (ISVs), or whether the investment introduces unnecessary complexity. Attendees will explore the different levels of model ownership, from prompt engineering and distillation to fine-tuning, reinforcement learning and continued pre-training, considering where the returns justify the effort for organisations of different sizes.

The discussion will also address the practical challenges of building proprietary models, including customer data rights, contractual restrictions, data protection requirements and the hidden costs of evaluation infrastructure, model serving and ongoing maintenance. Participants will consider how to establish clear ownership and accountability for models throughout their lifecycle.

Finally, we will explore hybrid approaches that combine specialised models for high-volume tasks with frontier models for more complex requests. Through AWS capabilities spanning Amazon Bedrock, SageMaker AI, HyperPod and model evaluation, attendees will discuss how to validate performance, manage costs and determine when owning a model becomes a genuine business advantage.