Elon Musk showed the Cortex AI supercomputer – 50 thousand Nvidia H100 will train AI for Tesla autopilot

Elon Musk posted a video on social network X, shot at his new facility – the Cortex artificial intelligence training supercluster, which is located near Tesla’s Giga Texas plant. The facility will operate 70,000 AI servers, which will consume a total of 130 MW. And by 2026, the supercomputer will be expanded to 500 MW.

Image source: x.com/elonmusk

The video demonstrates the process of assembling server racks – rows of 16 units are interspersed with approximately four racks without AI accelerators. Each rack includes eight servers. The 20-second video included about 16–20 rows of server racks, which, at a rough estimate, gives about 2,000 servers with accelerators or 3% of the total estimated capacity of the facility.

Video of the inside of Cortex today, the giant new AI training supercluster being built at Tesla HQ in Austin to solve real-world AI pic.twitter.com/DwJVUWUrb5

—

Cortex should become Tesla’s largest supercluster for training AI systems – 50 thousand Nvidia H100 accelerators and 20 thousand accelerators of the company’s own design will work here, although it was previously assumed that there would also be 50 thousand of them. Tesla accelerators will be installed a little later, and at launch only Nvidia equipment will work here. The system is being created to “solve AI problems in the real world.” We are talking about training the Tesla Full Self Driving (FSD) autopilot system for consumer cars and Cybertaxi, as well as AI training for the Optimus robot, whose small-scale production is expected to be launched in 2025.

Previously, Musk posted a photo of the Cortex facility’s giant fans connected to a Supermicro liquid cooling system that will handle the entire 500-MW facility. Musk’s first data center to be put into operation will be his startup xAI Memphis Supercluster with 100 thousand Nvidia H100 in a single RDMA structure and with Supermicro cooling – in the future, another 300 thousand B200 will be connected to them, but due to Due to design flaws, their commissioning is delayed for several months. In addition, the Tesla-owned Dojo supercomputer, worth $500 million, is preparing to launch in Buffalo, New York.