On October 31 local time, Musk disclosed the latest progress on the Colossus 2 data center. Beyond the deployed Macroharder and Minihard compute clusters, SpaceX is building two new clusters, Gun Range and Spam, planning to add about 420,000 Nvidia GB300 GPUs.
Gun Range will have about 220,000 GB300s, using the same training cluster configuration and 800Gbps NICs as Macroharder and Minihard, expected to enter operation by end-November. Spam plans about 200,000 GB300s mainly for AI inference, expected to go live in December. Together, the two clusters total about 420,000 GB300s.
The key difference is not GPU count but network architecture. Musk said inference clusters have simpler network architectures, so Spam needs less than half the fiber of training clusters. Large training clusters require massive data exchange and synchronization between GPUs; once compute tasks are split across GPUs and servers, gradients, parameters, and intermediate data must be exchanged over high-speed networks, directly affecting training efficiency. Inference clusters use trained models to handle user requests, with communication patterns depending on model deployment, service architecture, and load characteristics—so they need not replicate training cluster network configurations.
In the broader Colossus 2 timeline, Musk said in September the data center already had about 110,000 GB200s and 440,000 GB300s. SpaceX's June 2026 EU prospectus disclosed that the first 110,000 GB200s and about 210MW came online in 91 days, while the second batch of 110,000 GB300s and about 220MW took only 64 days. Based on public configurations, Macroharder, Minihard, and Gun Range each have about 220,000 GB300s for training, and Spam about 200,000 for inference—four clusters totaling about 860,000 GB300s, of which about 660,000 are for training. Musk said Colossus 2 will have about 1.2 million Blackwell GPUs by end-2026, a year-end target, not deployed actuals.
With Gun Range starting by end-November and Spam in December, deploying about 420,000 GB300s in two months is a major test for power, cooling, network cabling, server integration, and on-site engineering management. For Nvidia, it means not only continued GPU demand but also continued benefits for high-speed networking, switch chips, optical modules, and fiber interconnect as AI infrastructure expands.
ICgoodFind Summary: Colossus 2's 420,000 new GB300s matter most for the network architecture split between training and inference clusters. Training needs 800Gbps NICs and high-density fiber; inference uses less than half the fiber. AI data centers are shifting from stacking GPU counts to configuring compute, storage, and network by workload—a trend that may drive optical modules, switch chips, and high-speed interconnect longer than GPU counts alone.
