According to CoinMeta, Musk revealed that the Memphis supercomputing cluster Colossus plans to deploy approximately 1.2 million NVIDIA Blackwell GPU units by the end of the year. Additionally, two more facilities will be put into operation: gun and range are expected to launch with 220,000 GB300 units by the end of November for model training. spam was converted from a warehouse that originally stored luncheon meat and other food products and is planned to deploy about 200,000 GB300 units, specifically for AI inference tasks. Currently, macrohard is equipped with 110,000 GB200 and 220,000 GB300 units; macroharder and minihard each have 220,000 GB300 units. Together with the additional 420,000 units planned to be added in the future, the five clusters will total about 1.19 million units. Training and inference tasks have different network requirements; macroharder, minihard, and gun all use 800Gbps network cards, whereas the inference tasks of spam do not require such high-speed interconnection. Musk stated that its network configuration is simpler, requiring less than half of the fiber optic capacity needed by the training facilities.