Foundation Models

Foundation models are large AI models trained on vast, diverse datasets — text, images, code, audio, and more — that develop broad general capabilities as a result. They serve as a starting point that can be adapted for specific tasks through prompting, retrieval, or fine-tuning, rather than training a model from scratch for each use case. Large language models are the most prominent type, but foundation models also power image generation, speech recognition, and multimodal applications. Their scale requires significant compute to train, which is why a handful of providers — OpenAI, Anthropic, Google, Meta — produce the base models that most organizations then build on.

Foundation models have shifted the build-versus-buy calculus: for most organizations, building from scratch is no longer a realistic option for general-purpose AI capabilities. But "buying" comes with consequential choices. Which model, deployed how, with what data protections, under whose terms — these are governance decisions with real implications for data exposure, vendor dependency, intellectual property, and operational continuity. Model choice can't be delegated entirely to the technical team; it involves legal, procurement, and risk.