According to Dynamic Beating monitoring, a well-known semiconductor analysis firm SemiAnalysis stated that Meta is simultaneously building 5 AI clusters, each with a power capacity exceeding 1GW. By their estimation, Meta's total computing power may surpass that of OpenAI and Anthropic by the end of this year.
Among them, the Louisiana-based Hyperion is constructing a facility of around 1GW, with future expansion potential up to 5GW. The Prometheus facility in Ohio has already partially commenced operations, with a planned scale exceeding 3GW.
Prometheus consists of 6 campuses and 27 data centers. Meta interconnects them into a cluster with high-speed networking and is planning to link campuses that are over 2000 kilometers apart in the future.
Due to high network latency over long distances, Meta will centralize pre-training in a single region and distribute reinforcement learning tasks to run in different campuses.
Meta has also assembled a team of about 3000 people dedicated to creating reinforcement learning tasks and training environments. The company had previously begun recording some employees' screen, keyboard, and mouse operations to train an AI Agent using real work processes, but related plans were temporarily paused due to privacy and data management issues.
SemiAnalysis concludes that Meta has successfully secured talent, data, and computing power, putting it in a strong position to catch up with OpenAI and Anthropic. However, Muse Spark has not yet entered the top tier, and whether the substantial investment can translate into leading models remains to be validated.
