The Three Scaling Laws

NVIDIA frames Blackwell around three “scaling laws” — three distinct ways that spending compute makes a model better. Understanding them is the key to understanding every design choice in the hardware. Here’s each one, with an analogy.

#1 · Pre-training scaling — going to school

Pre-training is the original law of AI development. Increase three things together — the training dataset size, the model’s parameter count, and the compute you spend — and you get predictable improvements in intelligence and accuracy. This law drove the rise of billion- and trillion-parameter transformer models.

#2 · Post-training scaling — learning the job

Pre-training teaches general knowledge; post-training teaches the model how to think for a specific purpose, improving its relevance for an organization’s use case. If pre-training is school, post-training is on-the-job training. Developers use synthetic data to augment their fine-tuning sets — and because synthetic data has no upper limit, the appetite for compute is large.

#3 · Test-time scaling — thinking before answering

An ordinary LLM fires back a quick response in a single pass — fine for simple questions, weak on hard ones. Test-time scaling (also called long thinking or reasoning) spends extra compute during inference, letting the model reason through multiple possible answers before committing to the best one. This is the capability behind agentic AI.