According to their S-1:

In H1 2026, cost of revenue is $2.3B. That is primarily inference compute, but also includes cost of training models (but not research).

In the same period, R&D was $9.4B. That is primarily spent on compute. That is what they are "holding back" already, in full public view.

This is not a company that is optimizing for OpEx or end user revenue. The only thing relevant to their plans is achieving AGI/ASI first, and controlling enough compute to leverage it to earn the majority of the current economy.