Meta has officially launched an artificial intelligence supercomputer that it claims already ranks among the top ten most powerful systems on the planet. The machine, built entirely from the ground up by Meta’s own engineers, is designed to train the next generation of large-scale AI models and bring the company closer to its long term goal of achieving artificial general intelligence, or AGI.
Custom hardware, massive scale
The supercomputer, which Meta has not given a flashy consumer name, relies on a cluster of thousands of Nvidia A100 Tensor Core GPUs linked together with a proprietary networking fabric developed in house. Meta says the system can handle over 5 exaflops of mixed precision compute, placing it in the same category as the world’s fastest research clusters. To put that in perspective, 5 exaflops is roughly the combined peak performance of several of the top publicly ranked supercomputers from just a few years ago.
Meta designed the system with a two phase rollout. The first phase, already operational, uses 16,000 GPUs. The second phase, expected to be completed later this year, will more than double that number, pushing the total GPU count past 32,000. At full capacity, the company says its supercomputer will be capable of training models with trillions of parameters, far beyond the scale of today’s largest language models.
Training the future of AI at Meta
Meta’s immediate plan for the supercomputer is to train the company’s core AI research models, including those that power content moderation, recommendation algorithms, and natural language understanding across Facebook, Instagram, and WhatsApp. But the company is also clear that this machine is a stepping stone toward AGI, a type of AI that can learn any intellectual task a human can.
The system has already been used to train a new model called OPT 175B, a 175 billion parameter language model that Meta made publicly available to researchers. That model, while large, is dwarfed by what the company says will be possible once the full cluster is online. Meta’s AI chief, Yann LeCun, has long argued that scaling up compute and data is a necessary but not sufficient path to AGI. This supercomputer gives his team the raw horsepower to test those theories at a scale previously reserved for a handful of government labs and cloud hyperscalers.
One challenge Meta had to overcome was cooling. The system generates enormous amounts of heat, enough to require a custom liquid cooling loop that sends chilled water directly to each GPU rack. Meta engineers had to design the facility’s power distribution and cooling infrastructure from scratch, since no off the shelf solution could handle the density of compute they planned to install.
Why this matters for the AI race
Meta enters a fiercely competitive arena. Microsoft, Google, Amazon, and several well funded startups are all racing to build ever larger AI supercomputers. Microsoft recently revealed it has been working with OpenAI on a multi billion dollar supercomputer hosted in Azure data centers. Google has its own TPU powered clusters that train models like PaLM and Gemini. Meta’s entry signals that it intends to remain a first tier player in AI research, not a company that simply licenses models from others.
The company also emphasized that it will share some of the architectural lessons learned with the broader open source community. Meta has published several papers detailing the networking and scheduling software that keeps the cluster running efficiently. That openness stands in contrast to the more closed approaches taken by some competitors, and it could accelerate progress across the entire field.
For now, Meta’s supercomputer is a tool for research, not a commercial product. But as the company moves toward AGI, the line between research and product is likely to blur. By building its own infrastructure rather than relying on cloud providers, Meta keeps control of its own destiny. Whether that bet pays off may depend on how quickly the team can scale the system and how effectively they can use it to crack some of the hardest problems in AI. For more analysis on how major tech companies are reshaping AI infrastructure, check out our deep look at {$link_text}.
