Elon Musk recently shared a video on social media, showcasing the internal construction of Tesla's Cortex AI supercomputer cluster being built at their Austin headquarters. This newly named supercomputer cluster is part of Tesla's "Giga Texas" factory and is expected to require 130 megawatts of cooling and power upon launch, growing to 500 megawatts by 2026.
The scale of the Cortex supercomputer cluster is impressive, with an estimated 70,000 AI servers. Musk's video displays numerous server racks being assembled, with each row containing 16 computer racks and each rack housing 8 servers. The video shows approximately 16-20 rows of racks, estimating around 2,000 GPU servers, which is less than 3% of the expected full scale.
During Tesla's July earnings call, Musk revealed that the Cortex supercomputer cluster would include "50,000 Nvidia H100 GPUs, plus 20,000 of our own hardware." This is a reduction from his previous estimates, where he had anticipated 50,000 Tesla Dojo AI hardware units. Tesla's own hardware is expected to come online later, with Cortex relying entirely on Nvidia's hardware at launch.
The purpose of building the Cortex supercomputer cluster is to "solve real-world AI problems." This includes training Tesla's Full Self-Driving (FSD) system, which will power consumer Tesla cars and the upcoming "Cybertaxi" product, as well as training the AI for the Optimus robot. Optimus, an autonomous humanoid robot slated for limited production in 2025, will be used in Tesla's manufacturing processes.
The cooling system for the Cortex supercomputer cluster is also noteworthy. Musk showcased the massive fan system under construction in June, designed to cool the entire supercomputer cluster. This fan stack provides cooling for Supermicro's liquid cooling solution, which is built to handle 500 megawatts of cooling and power.
The Cortex supercomputer cluster joins the ranks of other supercomputer clusters being developed by Musk. Currently, the first operational cluster in Musk's data centers is the Memphis supercomputer cluster, owned by xAI and powered by 100,000 Nvidia H100 GPUs. All 100,000 servers in Memphis are connected through a single RDMA (Remote Direct Memory Access) network and are similarly cooled with Supermicro's assistance.