Musk powers up a 100k H100 monster cluster to build the next Grok

Elon Musk already runs Tesla, SpaceX, Starlink, Neuralink, and X, but he clearly wanted more firepower for his AI ambitions. Musk announced on X that xAI has officially started training on what he calls the most powerful AI training cluster in the world.

Dubbed the Memphis Supercluster and located in southwestern Memphis, Tennessee, the facility is designed to train the company's Grok large language models.

100,000 liquid-cooled chips on a single fabric

The supercluster packs a massive punch with 100,000 liquid-cooled Nvidia H100 GPUs. That is a staggering collection of top-tier silicon running in one place.

Musk noted that the entire setup operates on a single RDMA fabric. This remote direct memory access architecture lets compute nodes swap data directly without leaning on central processors, keeping latency low and throughput high across the whole network.

According to local outlet WREG, this marks the largest capital investment by a new-to-market business in Memphis history.

There is a notable hitch, however. Local utility Tennessee Valley Authority requires formal contracts for power demands exceeding 100 megawatts, and xAI has not secured that agreement yet.

Eyeing the AI crown by December

Musk made no secret of his target. Replying to a post, he stated that xAI aims to deliver the world's most powerful AI by every metric by December this year.

xAI currently serves Grok to paid subscribers on X, but keeping pace with OpenAI, Anthropic, Google, and Meta requires massive model leaps. Musk is counting on the Memphis cluster to give his team a critical edge.

Still, his track record on aggressive timelines gives reason for skepticism. Between Robotaxis, full self-driving promises, and Mars flights, Musk rarely hits his self-imposed deadlines.

Meanwhile, the rest of the industry is not sitting still. Microsoft and OpenAI are already reportedly mapping out a $100 billion supercomputer codenamed Stargate, meaning xAI’s hardware crown could be temporary.

Why it matters

For everyday users and X Premium subscribers, this means Grok could see a dramatic leap in intelligence before the year ends. If xAI hits its mark, users get a far more capable tool for coding, reasoning, and content generation, though it is not worth upgrading a subscription just for future promises.

For the industry, xAI just proved that the compute race is escalating into brute-force territory. Deploying 100,000 liquid-cooled chips on a unified RDMA network resets the baseline for training frontier models.

We think xAI made an aggressive, clever engineering bet to bypass traditional scaling bottlenecks and catch OpenAI quickly. That said, operating without a finalized power contract with the local utility is a massive operational gamble. Piling up GPUs is one thing, but keeping the lights on and delivering a chart-topping model by December is a whole different test.