Elon Musk's xAI Colossus Supercomputer Is on Its Way to 1 Million GPUs

Phase One of xAI’s supercomputer is operating arsenic planned. Elon Musk’s monolithic AI grooming strategy successful Memphis, Tenn. has transitioned to the city’s main powerfulness grid, according to Tom’s Hardware. The grid gives xAI entree to the 150MW it needs to powerfulness a halfway that operates 200,000 Nvidia GPUs. The institution plans to scope 1 cardinal GPUs arsenic it scales up to vie with rivals similar the Oracle Cloud Infrastructure (OCI), which has astatine slightest 131,072 GPUs.

One of the astir intriguing aspects of xAI’s ambitious task is the velocity with which it’s moving. The institution came onto the AI country successful the summertime of 2023, erstwhile Musk announced xAI connected X, penning that its intent was “to recognize reality.” Fast guardant a year, and operation of Colossus, the company’s AI supercluster, was good underway. Colossus started with 100,000 Nvidia H100 Hopper AI accelerators, which is simply a stunning fig of devices successful its ain right, but that fig doubled successful February 2025 to 200,000.

According to xAI, Colossus was built successful 122 days, which is simply a shockingly accelerated operation period. The institution past bumped to 200,000 GPUs successful 92 days. During this time, the institution faced a monumental challenge: providing powerfulness to the monolithic fig of GPUs successful its cluster. Unfortunately, the powerfulness grid provided conscionable 7MW to Colossus astatine launch. To supply powerfulness portion awaiting a amended transportation to the Memphis powerfulness grid, the institution made usage of earthy state generators, which, arsenic Tom’s Hardware notes, led to complaints from immoderate residents.

Now, the concern is different. Memphis provides 150MW of power, which resulted successful the Colossus reducing the fig of generators it was using. On apical of that, the Colossus seemingly has 150MW successful Tesla batteries to supply powerfulness erstwhile needed. That’s bully quality for Phase One, but Phase Two volition necessitate much power. Another substation volition assistance the Colossus get to 300MW of power, perchance arsenic aboriginal arsenic the autumn of this year. The supercomputer’s powerfulness needs volition proceed to emergence arsenic much GPUs are added to the existing setup.

XAI offers Colossus for grooming ample connection models (LLMs) and uses it to bid its ain Grok, an AI instrumentality disposable connected Musk’s X societal network. The institution precocious released Grok 3 beta, which lets users spot the reply it provides and the reasoning it utilized to get astatine that answer.