What is Foundation Model?

What Is a Foundation Model?

A foundation model is a large-scale AI model trained on vast, diverse datasets that serves as a general-purpose base for a wide range of applications. The term was coined by researchers at Stanford University in 2021 to describe AI models that are trained once at enormous scale and then adapted to many different downstream tasks.

Think of a foundation model like a highly educated generalist who has read millions of documents across every field. This generalist can then be given specific instructions or additional training to become effective at particular tasks -- writing legal contracts, analyzing medical images, translating documents, or coding software -- without needing to be trained from scratch for each new application.

Why Foundation Models Changed the AI Landscape

Before foundation models, building an AI system for a specific business task required collecting task-specific data, training a model from scratch, and maintaining it separately for each application. This approach was expensive, time-consuming, and accessible only to organizations with significant AI expertise and resources.

Foundation models fundamentally changed this equation by creating a "train once, use many times" paradigm:

• Pre-training: The model is trained on massive datasets by the model provider (OpenAI, Google, Anthropic, Meta) at enormous cost, often tens or hundreds of millions of dollars
• Adaptation: Businesses take the pre-trained model and adapt it to their specific needs through fine-tuning, prompt engineering, or retrieval-augmented generation (RAG) at a fraction of the original training cost
• Deployment: The adapted model is deployed for specific business applications

This means an SMB in Jakarta or Ho Chi Minh City can access AI capabilities that previously required the resources of a tech giant, paying only for the adaptation and usage rather than the billions of dollars in training costs.

Types of Foundation Models

Foundation models exist across multiple domains:

Language Models Models like GPT-4, Claude, Gemini, and Llama that understand and generate text. These are the most widely used foundation models in business today, powering everything from chatbots to document analysis.

Image Models Models like DALL-E, Stable Diffusion, and Midjourney that generate images from text descriptions. Businesses use these for marketing visuals, product mockups, and design prototyping.

Multimodal Models Models like GPT-4o and Gemini that can process and generate multiple types of content including text, images, audio, and video. These are increasingly important for businesses that need AI to work across different media types.

Code Models Models like Codex (behind GitHub Copilot) and Code Llama that specialize in understanding and generating programming code. These accelerate software development and are particularly valuable in Southeast Asia's growing tech sector.

Domain-Specific Models Models trained on specialized data for specific industries, such as Bloomberg's financial model or Google's medical AI models. These are emerging for various ASEAN-relevant industries.

The Business Model of Foundation Models

Understanding the economics of foundation models helps business leaders make better AI investment decisions:

For Model Providers Building a foundation model requires massive investment in compute, data, and talent. OpenAI, Google, Anthropic, and others spend hundreds of millions to train these models, then monetize through API access, enterprise subscriptions, and platform fees.

For Businesses Using Foundation Models The cost structure is fundamentally different. Instead of investing in model training, businesses pay for:

• API usage (per-token or per-request pricing)
• Fine-tuning costs (adapting the model to specific tasks)
• Infrastructure (if self-hosting open-source models)
• Integration development (connecting the model to business systems)

This shifts AI from a capital-intensive investment to an operational expense that scales with usage, making it far more accessible to SMBs.

Foundation Models in Southeast Asian Context

The foundation model paradigm is particularly advantageous for ASEAN businesses for several reasons:

Reduced Infrastructure Requirements Companies do not need massive GPU clusters or data centers to use state-of-the-art AI. Cloud-based access to foundation models means a small team in Manila or Kuala Lumpur can leverage the same AI capabilities as a Fortune 500 company.

Multilingual Advantage Foundation models trained on global datasets include substantial representation of ASEAN languages, meaning businesses can build multilingual applications without training separate models for each language.

Leapfrog Opportunity Southeast Asian businesses can skip the expensive process of building AI capabilities from scratch and jump directly to using and adapting the most advanced models available globally.

Choosing a Foundation Model Strategy

Business leaders should consider several factors when developing their foundation model strategy:

• Build vs. buy: Most SMBs should use existing foundation models rather than attempting to train their own
• Open vs. closed: Evaluate whether open-source models (more control, self-hosting) or proprietary APIs (easier to use, managed by provider) better fit your needs
• Single vs. multi-model: Consider using different foundation models for different tasks rather than trying to use one model for everything
• Today vs. tomorrow: The foundation model landscape is evolving rapidly, so design systems that can adapt to new models as they become available