Definition
Pretraining is the first and costliest stage of building an AI model, where it reads enormous amounts of text to learn general language and facts before any specialization.
At a glance
- The model learns by guessing the next word in real text billions of times, absorbing grammar, facts, and reasoning[1].
- It uses huge, unlabeled datasets (books, websites) and produces a general foundation, not a finished tool.
- It is hugely expensive: GPT-4 cost an estimated 78 million dollars, Gemini Ultra around 191 million.
- You almost never pay for it; you adapt a shared, pre-built foundation instead.
How it works
The model reads ordinary text and plays a guessing game: predict the next word, check the answer, adjust. Repeated billions of times, this builds grammar, world facts, and basic reasoning, with no human labeling required.
Why it is so expensive
Pretraining runs for weeks on thousands of specialized chips, dominating the cost of modern AI[3]. GPT-4’s compute is estimated near 78 million dollars and Gemini Ultra around 191 million[4]. That is why most companies never pretrain their own model.
What it means for your business
You use a model someone already pretrained, then prompt it or fine-tune it on a little of your own data. Fine-tuning often costs a few hundred to a few thousand dollars, because the expensive learning already happened[2].
Bottom line
Pretraining is the costly, one-time education behind every AI model; you stand on a shared foundation and adapt it for a tiny fraction of the original price.