What Is a Foundation Model? The Architecture Behind ChatGPT, Claude, and Copilot

A foundation model is a large AI model trained on enormous quantities of data at enormous cost, which serves as the base for many different AI applications. The training cost for a top-tier foundation model runs to hundreds of millions of dollars, which is why only a small number of organisations (OpenAI, Anthropic, Google, Meta, Mistral) have the resources to build them.

The 'foundation' metaphor is apt: the model provides a base capability that many different applications can build on. GPT-4 is the foundation model that powers ChatGPT, but it also powers many third-party applications that have built AI features using OpenAI's API. Claude is the foundation model that powers Anthropic's Claude.ai product but is also available through AWS Bedrock and used in many other applications.

Microsoft Copilot is built on top of foundation models from OpenAI (GPT-4 and related models). Microsoft does not have its own proprietary foundation model; it has invested in OpenAI and built the Copilot product layer on top of OpenAI's models, combined with Microsoft's enterprise infrastructure, security, and integration capability.

Google Gemini is both a foundation model and a product: Google builds its own foundation models and uses them in its Workspace AI features and Google Cloud AI services.

Anthropic builds its own foundation models (the Claude family) and makes them available through its own products and through partner platforms.

This architecture matters because when you buy an AI product, you are often buying access to a foundation model through a product interface. The underlying model's capabilities and limitations shape what the product can do.

The foundation model market is consolidating around a small number of major providers. The competitive dynamics between OpenAI, Anthropic, Google, and Meta shape what capabilities will be available to enterprises over the next two to five years.

For strategic planning, the relevant questions are: which foundation model provider is most aligned with your use cases (some models are stronger in code, others in reasoning, others in long document processing)? What are the data governance terms when you use each provider's models? How does your primary AI vendor's relationship with its foundation model provider affect your exposure to changes in that relationship?

Microsoft's significant investment in OpenAI is a specific example: if that relationship were to change, it could affect the capabilities and terms of Microsoft Copilot. Boards responsible for AI strategy should understand these dependencies.