A training run is the single, sustained computational process by which a neural network learns from data. Billions of text examples are fed through the model, mistakes are measured, and the model's parameters are nudged in the right direction — then the cycle repeats, trillions of times, across tens of thousands of specialized chips running for months.

What makes this expensive isn't just hardware rental — it's that the process is largely irreversible. You can't pause halfway, fix a data problem, and cheaply resume. Mistakes in data curation or hyperparameter choices can waste months of compute. Frontier training runs are estimated to cost tens of millions to hundreds of millions of dollars.

This is why labs invest heavily in inference efficiency afterward: once training ends, the weights are fixed. Every improvement in serving that model — quantization, smarter batching, caching — is an attempt to extract more value from a sunk cost that can't be undone.