Meta says the first phase in New Albany is complete and was used to train Muse models; the site now has hundreds of thousands of GPUs, with a stated target of roughly one million. The company reports more than 85% goodput for Muse Wonder pre-training and more than 90% for reinforcement-learning runs.
Thesis
Large-scale model training depends on the reliability and coordination of the full compute system, not only accelerator peak throughput.
Observed facts
Meta says Prometheus’s first phase is complete and was used to train Muse models; the cluster currently comprises hundreds of thousands of GPUs across interconnected buildings, with a target of roughly one million.
1Meta reports more than 85% goodput for Muse Wonder pre-training and more than 90% for reinforcement-learning runs, with interruption intervals later rising above 54,000 GPU hours after fleet changes and firmware fixes.
1
Mechanism
A layered system combines interconnected buildings, a dedicated Ethernet back-end for GPU-to-GPU communication, successive NVIDIA GPU platforms, and cooling designed for high-density deployment; reliability improvements increase usable training time.
Why now
Meta ties the completed first phase to training its Muse family and frames Prometheus as infrastructure for future frontier models, while the published figures make reliability measurable rather than merely asserting scale.
Implications
The first implication is that cluster-level reliability and network design can determine how much nominal compute becomes usable training time.
As capacity scales, operators will need to track operational performance across buildings and hardware generations, not just accelerator counts.
Evidence boundary
These are Meta-reported engineering metrics; the post does not independently validate them or provide a complete cost, energy, or comparative-performance analysis.
Watch conditions
The first watch condition is whether Meta publishes comparable goodput and interruption metrics as Prometheus scales toward its stated million-GPU target.
Whether reported reliability gains persist across additional buildings and hardware generations.