Meta details its Prometheus supercomputer in Ohio, featuring 1 million GPUs and 1 gigawatt of power

On Monday, the 19th, Meta unveiled its artificial intelligence training facility in New Albany, Ohio. The first phase is already operational and has been used to develop models from the Muse family, including the recently launched Muse Wonder.

The center currently houses hundreds of thousands of GPUs—chips designed to process large volumes of calculations simultaneously. The plan is to expand to approximately 1 million units, distributed across multiple buildings that operate in a coordinated manner.

How multiple buildings function as a single computer

The primary challenge is communication. During AI training, chips must constantly exchange results; any delay can halt the entire process.

To connect the buildings, Meta uses a network layer called Backend Aggregation (BAG). It links networks from different locations and provides alternative data routes, minimizing the impact of failures.

This expansion builds on previous growth. In late 2023, the company assembled two clusters of approximately 24,000 GPUs each. It then consolidated 129,000 H100 chips from five data centers into a single system after decommissioning the previous facilities.

Cooling is also critical

The project was announced with an electrical capacity of 1 gigawatt. Beyond purchasing chips, the expansion requires a solution to dissipate the heat generated by increasingly concentrated machines.

At Prometheus, liquid cooling is used in both adapted facilities and buildings specifically designed for AI. The principle is straightforward: a fluid circulates through the equipment, absorbs heat, and passes through exchangers that cool it before returning it to the circuit.

In buildings lacking this infrastructure, Meta installs modules with pumps and heat exchangers near server racks, adapting the environment without relying solely on ventilation.

Why this matters for AI agents

The goal is to support models capable of operating for longer periods and maintaining context for complex tasks. Muse Wonder is already available to developers, and its integration into the Muse personal assistant will launch later, starting with subscribers.

Prometheus is part of Meta’s strategy to regain ground in the AI race. Funding for the expansion comes from advertising revenue generated by Facebook and Instagram.

With the first phase complete and Wonder launched, the company now aims to transform the infrastructure into agents capable of performing longer tasks, not just answering questions.

Frequently Asked Questions

• What is Meta’s Prometheus? It’s a computing complex dedicated to AI training, located in New Albany, Ohio, USA.
• Is Prometheus operational yet? Yes. The first phase is complete and has already been used to train models from the Muse family.
• Does Prometheus already have 1 million GPUs? No. Currently, there are hundreds of thousands, with plans to reach approximately 1 million during the expansion.
• Why do chips need a special network? Training requires constant exchange of results between machines. The network connects buildings and provides alternative paths to reduce failures.
• What is the relationship between Prometheus and Muse Wonder? Prometheus is the computing infrastructure; Muse Wonder and Wonder Full are AI models trained on this system.
• How does Prometheus’ cooling system work? A fluid absorbs heat from the equipment, passes through exchangers that cool it, and returns to the circuit for reuse.