Meta Invests Billions to Build World's Largest AI Data Center, Hyperion

Meta is embarking on an unprecedented strategic transformation to overcome its lag in the foundational model competition. The company is investing thousands of billions of dollars to build several large data centers, with the first, Prometheus, expected to be operational by next year. This initiative is part of a broader effort to enhance Meta's computational capabilities and attract top talent to compete with leading AI labs like OpenAI.

Meta's aggressive transformation is a response to the failure of its Llama 4 Behemoth large model. Despite Llama 3's success in leading the open-source model wave, the failure of Llama 4 has damaged Meta's reputation. The technical roots of this failure include architectural missteps, data quality bottlenecks, and shortcomings in scaling and evaluation. Meta aims to address these issues by investing heavily in computational infrastructure and strategic acquisitions.

To rapidly acquire massive computational power, Meta has abandoned its decade-long data center construction blueprint. Instead, it is adopting a new design inspired by xAI, featuring a "tent-like" structure with prefabricated power and cooling modules, and ultra-lightweight construction. This design sacrifices some redundancy, such as backup diesel generators, to expedite the deployment of GPU clusters.

Meta is advancing two major infrastructure projects to achieve its goals. The Prometheus cluster, located in Ohio, is a 1-gigawatt AI training cluster. Meta is employing a comprehensive strategy, integrating self-built campuses, third-party leasing, and on-site natural gas power generation. The Hyperion cluster, situated in Louisiana, is even more ambitious, aiming to surpass OpenAI's highly anticipated Stargate project. Hyperion's first phase will exceed 1.5 gigawatts of IT power and is expected to become the world's largest single AI data center by the end of 2027.

These efforts aim to transform Meta from a "GPU-poor" to a "GPU-rich" entity, enabling its training capabilities to match those of leading AI labs. Meta's strategic focus has shifted to another critical element: talent. Recognizing the talent gap between Meta and top AI labs, the company is personally overseeing the recruitment of a new "superintelligence" team. Meta offers top researchers compensation packages typically amounting to 2 billion dollars over four years, with some key positions receiving rejected offers of 10 billion dollars.

This strategy not only attracts talent but also increases the cost of hiring for competitors. Recent notable additions include former GitHub CEO Nat Friedman and Daniel Gross, co-founder of SSI with Ilya Sutskever. In addition to talent acquisition, strategic acquisitions are another pillar. The investment in Scale AI is seen as a crucial step, directly addressing the data and evaluation shortcomings exposed by Llama 4. Scale AI's founder, Alex Wang, and the SEAL lab, which specializes in model evaluation, will bring urgently needed capabilities to Meta, particularly the HLE (Humanity’s Last Exam) benchmark for reasoning model evaluation, which will significantly address Meta's deficiencies.