XAI rapidly building AI data centers with 30 times the compute capacity of Grok 3 completed in 7 months.

XAI has completed the installation of 550,000 Nvidia B200 GPUs in a matter of weeks—a deployment speed far faster than previously anticipated. This represents over 7 times the compute used to train Grok 4, which was released just weeks prior. The achievement is particularly striking when compared to earlier iterations: Grok 3, released in February of this year, utilized 100,000 Nvidia H100 chips and represented 12 times the compute of Grok 2, which had been deployed six months earlier.

Grok 4 employed 230,000 Nvidia chips in total—comprising 150,000 H100s, 50,000 H200s, and 30,000 B200s—equivalent to 400,000 H100-class units. This marked a four-fold increase in compute capacity achieved in approximately three months. By comparison, the 550,000 B200 installation represents 30 times the compute of the earlier Grok 3 system.

The speed of GPU installation has become a defining competitive advantage for XAI. When deploying Grok 3, the company installed 100,000 chips in 120 days, then accelerated to deploy the next 100,000 in approximately 90 days. The current installation rate has increased dramatically to between 200,000 and 300,000 chips within 30 days, positioning XAI as approximately 100 times faster than competitors at deploying GPU chips in data centers.

The infrastructure supporting this expansion extends to power delivery. The 550,000 B200 chips require approximately 1 gigawatt of power, a significant increase from the 400 to 500 megawatts required for Grok 4's 230,000 chips. XAI has announced a goal of bringing 50 million units of H100 equivalent-AI compute—with substantially improved power efficiency—online within five years.