Two completely different phases of a model's life, with different costs and hardware needs.
Training — learn the weights
The expensive, one-time(-ish) phase: feed huge data through the network and adjust weights until it performs. Frontier models take thousands of GPUs for weeks and millions of dollars.
Inference — use the model
Running the trained model to get an answer. Far cheaper per run, but it happens constantly — every chat message is an inference. At scale, total inference cost dwarfs training.
The comparison
| Training | Inference | |
|---|---|---|
| When | Once (then fine-tunes) | Every request |
| Cost shape | Huge upfront | Small × billions |
| Hardware | Many GPUs, fast interconnect | One GPU (or CPU) can serve |
| You do it? | Rarely (mostly download) | Always |
For local AI you almost always download someone's trained weights and only do inference — which is why VRAM and serving matter more than training rigs. Customize with the cheaper fine-tuning rather than training from scratch.
Related: How Neural Networks Learn · Choosing a Local LLM · Fine-tuning vs RAG vs Prompting