The news, 365 days behind — on purpose Delayed live · replaying 2025

One Year Ago.AI

Remember how fast this is.

22JUL2024replayed
one year on
computexAI · Elon Musk · Nvidia · X · Supermicro

Musk says xAI begins training on Memphis 'Gigafactory of Compute' with 100,000 H100 GPUs

Elon Musk claims the Memphis Supercluster is the most powerful AI training cluster in the world and will produce the 'most powerful AI by every metric' by December.

Elon Musk announced today that xAI has begun training on its Memphis Supercluster, a facility he claims is ‘the most powerful AI training cluster in the world.’ The system houses 100,000 liquid-cooled Nvidia H100 GPUs connected via a single RDMA fabric. Musk posted on X that training started at 4:20 a.m. local time and that the cluster will be used to create ‘the most powerful AI by every metric by December of this year.’

The Memphis site, which Musk has previously called a ‘Gigafactory of Compute,’ was reported in May as aiming to open by Fall 2025. Today’s news suggests the project came online more than a year early, or earlier reports misstated the timeline. The cluster uses current-generation H100 chips rather than waiting for Nvidia’s upcoming H200 or Blackwell GPUs, signaling Musk’s urgency.

The scale dwarfs existing supercomputers: Frontier uses 37,888 AMD GPUs, Aurora has 60,000 Intel GPUs, and Microsoft Eagle relies on 14,400 H100s.

C
Charles Liang

Supermicro CEO praised the team's execution in a tweet, highlighting speed, efficiency, and environmental friendliness.

One year later — open only if you can handle spoilers

The Memphis cluster, later named Colossus, became a key asset for xAI, enabling the training of Grok 3. However, reports emerged of xAI renting compute to rival Anthropic, and the facility's energy use drew continued scrutiny from local regulators and environmental groups.

Replay thisPost on XRedditHNLinkedIn