What Is a Foundation Model?

ChatGPT, Copilot, Gemini, you’ve probably heard the names. But underneath all of them sits something you’ve likely never seen mentioned: a foundation model. It’s the thing that actually makes the AI work, and understanding it will help you make better decisions about which tools to trust and why some cost more than others.

Start with the basics

A foundation model is a large AI system trained on a massive amount of text, images, code, or some combination of all three. The training happens once, or at least once per major version, and it’s expensive. We’re talking hundreds of millions of dollars and months of compute time. Once that training is done, the model becomes a kind of general-purpose base that other products are built on top of.

Think of it like a commercial kitchen that a chef rents to run a catering business. The kitchen exists, it works, someone else paid to fit it out, and the chef just uses it to make their food. The foundation model is the kitchen. ChatGPT, Claude, Gemini, they’re the meals.

What makes it “foundational”

The word “foundation” isn’t marketing fluff. These models are trained broadly, not narrowly. They don’t start with one job in mind. A foundation model trained on a large chunk of the internet learns grammar, facts, reasoning patterns, tone, how to follow instructions, and hundreds of other things, all at once, all from the same training run.

That breadth is what makes them useful as a starting point. You can take a foundation model and adapt it for customer service, legal research, medical triage, or writing marketing copy, without retraining the whole thing from scratch. That adaptation process is called fine-tuning, and it’s far cheaper than training from zero.

Why it matters for your business

Most of the AI tools you’re considering right now are built on someone else’s foundation model. That means a few things worth keeping in mind.

First, the quality ceiling. If the underlying foundation model is weak, no amount of clever product design will make the tool smart. When you’re evaluating AI software, asking “which model does this run on?” is a reasonable question, not a technical one.

Second, pricing structure. Foundation models are typically rented per use. Every time you or your staff prompt an AI tool, a tiny cost is incurred at the infrastructure level. That cost gets passed on, usually wrapped inside a subscription or per-message fee. Understanding this helps explain why AI tools that feel similar can have very different price points.

Third, data handling. Your input goes somewhere. Foundation model providers have their own terms about how training data is handled. Many enterprise tiers now let businesses opt out of having their data used for future training. If you handle sensitive customer information, that’s worth checking before you sign up for anything.

The main players right now

As of late 2025, a handful of organisations run the foundation models that power most commercial AI: OpenAI (the GPT series), Anthropic (Claude), Google DeepMind (Gemini), and Meta (Llama, which is open-source). Each has different strengths, pricing structures, and safety approaches.

There’s also a growing number of smaller, more specialised foundation models, trained on legal text, medical records, or scientific papers, designed for narrower but deeper expertise.

Open-source vs closed

Some foundation models are open-source, meaning the weights, the actual learned parameters, are publicly available. Meta’s Llama models are the most prominent example. This matters for businesses that want to run AI locally, avoid sending data to external servers, or customise deeply without licensing restrictions.

Closed models, like the GPT series, are only accessible via API. You send in a request, get back a response, but never touch the model itself. Most small business tools use closed models because they’re easier to integrate and often more capable at the top end.

The honest summary

Foundation models are the engines behind almost every AI product being marketed to small businesses right now. You don’t need to know how to build one. But knowing they exist, and that they vary in quality, cost, and data handling, puts you in a better position to ask the right questions when a salesperson is pitching you their “AI-powered” solution.

More plain English AI explainers: AI Plain English Hub →

Sources

• Stanford HAI, Reflections on Foundation Models
• Anthropic Research Overview
• OpenAI Research
• Meta AI, Llama Models
• Google DeepMind, Gemini

Related: AI Can’t Save Australian Tradies — It Can Only Help the Ones Already Running Good Businesses | PHP2Shell: Two Critical Bugs Just Hit Every PHP Website. Here’s What WordPress Site Owners Actually Need to Do.

📬 Subscribe to our weekly newsletter | 💼 Free resources & AI prompt packs