- The gigawatt-scale Prometheus cluster in Ohio trained Muse Wonder; Hyperion in Louisiana will scale to five gigawatts.
- Infrastructure costs run $50–60 billion per cluster; a gigawatt-year costs roughly $15 billion to operate.
- Fewer than 200 researchers staff Meta Superintelligence Labs under Alexandr Wang, concentrated on one training ladder.
- Internal evaluations now favor internal agent-reliability suites over public benchmarks.
- Zuckerberg argues the compute holds value regardless: other labs would pay a premium for access.
- Subscriptions, commerce take-rates, and API revenue are expected to fund continued CapEx.
Useful?