Re-cutting the Same Budget

A lab has drafted a training run for a new language model:

Two rules of thumb from scaling-law studies:

Cost. Training a model with $N$ parameters on $D$ tokens takes about

$$C \approx 6ND$$

floating-point operations (FLOPs). This $C$ is the compute budget.

Best split. For the lowest loss a fixed budget can buy, use roughly 20 training tokens per parameter:

$$D \approx 20N$$

The lab decides to keep the draft's compute budget exactly as it is, but divide it between parameters and tokens using the 20-tokens-per-parameter rule.

How many training tokens, in billions, should the re-planned run use? Round to the nearest whole billion.