Sorry if this is slightly off-topic, but given that you’ve been asking quite a few similar questions over the past weeks - have you considered running your models on random samples?
Looking e.g. at the NHS England resource allocation formula, which uses data on 55 million people over 3-5 years, and fits similar models (albeit often with lots more covariates) they seem to get by perfectly fine on a small random sample - you might want to think about just sampling as much as your hardware can handle in a reasonable amount of time?