From Hyperion to Prometheus, 2,000 km distributed training, splitting pretraining and reinforcement learning. This engineering architecture idea is quite interesting. 3,000 people dedicated to RL tasks. Meta's AI ambitions can no longer be hidden.

CoinWorld news: Meta is simultaneously building five large-scale AI clusters, each with a power capacity exceeding 1GW. According to estimates from well-known semiconductor analysis firm Semianalysis, Meta's total computing power may surpass that of OpenAI and Anthropic by the end of this year. Among them, the Hyperion facility in Louisiana is under construction with approximately 1GW of capacity, with future scalability up to 5GW. The Prometheus facility in Ohio is already partially operational, with a planned scale exceeding 3GW, consisting of six campuses and 27 data centers. Meta connects these facilities into a single cluster using high-speed networks, and in the future plans to connect campuses that are more than 2,000 kilometers apart. Due to high network latency over long distances, Meta will place pre-training in a single region and distribute reinforcement learning tasks across different campuses for execution. In addition, Meta has assembled a team of about 3,000 people dedicated to creating reinforcement learning tasks and training environments.