XAI’s Colossus 1 built in 122 days, rivals gigawatt-scale AI clusters

Colossus 1, built by xAI from scratch in 122 days, incorporated roughly 200,000 H100/H200 accelerators and about 30,000 GB200 NVL72 accelerators. The Memphis facility consumes roughly 300 megawatt—a power level comparable to the electricity use of about 300 thousand homes—and remains the single largest coherent cluster worldwide. Nevertheless, gigawatt-scale clusters under construction at OpenAI, Meta and Anthropic render xAI’s achievement modest by comparison. The major cloud partners of those rivals are financing the projects from their expense budgets to win the market by injecting capital.

The construction pace leaves competitors behind. Project Colossus 2 launched in March 2025 when xAI acquired a one-million-square-foot warehouse (about 93 thousand square meters) in Memphis together with two adjacent sites covering a total of 100 acres (about 405 thousand square meters). By August 2025 the site counted 119 air-cooled chillers delivering roughly 200 megawatt of cooling capacity, enough to operate about 110 thousand GB200 NVL72 accelerators. According to an online report by Elon Musk, “some server racks were already installed in July.” xAI’s construction speed sets a new benchmark: the company erected in six months infrastructure that took Oracle, Crusoe and OpenAI 15 months to build.

A gap between statements and on-site reality persists. Site photographs raise a central question about the power source; no substation or local turbines are visible, leaving the facility’s power solution unclear. The source also notes that the total capacity of a single training cluster will surpass that of Meta Superintelligence and Anthropic by the third quarter of 2025, but explicitly states xAI still needs to raise capital to purchase the accelerators themselves. The company received allocations from Nvidia, and the models are expected to be trained at large scale early next year, yet no performance metrics have been published to gauge the gap with competitors.

The source, presented as an industry intelligence report rather than a peer-reviewed paper, highlights a proprietary reinforcement-learning (RL) method that xAI is developing. The report estimates the methodology could enable the company to overtake OpenAI, Anthropic and Google. It adds that the data are based on a proprietary forecasting model that previously predicted Oracle deals months before official announcements, but those remain market estimates rather than verified company data. No benchmark data on Grok model performance have been released, and the claims of RL-architecture superiority rely solely on external analysis.