Part 2/8:
The growth rate of GPU clusters in data centers has been staggering. Major players have employed standard data center designs and connected them through high-capacity fiber networks. For instance, OpenAI and Microsoft have set up operations in Arizona featuring approximately 100,000 GPUs. Similarly, Meta converted existing designs into a more efficient layout to manage an impressive 128,000 GPUs for training models like Llama 4.
With each GPU consuming upwards of 1,400 watts, the power demands have swelled from just 15-20 megawatts a couple of years ago to an astonishing 150 megawatts today. This exponential growth is accompanied by an urgency to build data centers that not only meet current needs but anticipate future demands.