SpaceXSI’s expanding Colossus 2 build-out in the Memphis area is expected to bring a large collection of Nvidia Blackwell GPUs online, but public details about the storage supporting the cluster remain limited. VAST Data and DDN have both been associated with storage supplied to Elon Musk’s AI operations; the precise deployments and total capacity across Colossus 2 have not been confirmed.
The five-site cluster is described as a group of data centers in the region around Memphis, Southaven and Tulane Road, near the Tennessee-Mississippi border. Its sites have been given names including Macrohard, Macroharder, Minihard, GunRange and Spam. The names and site descriptions reflect differing functions and histories, but do not by themselves establish which storage systems are installed at each location.
GPU deployments and workloads
Macrohard was announced in March. Musk described the project as combining SpaceXSI’s Grok large language model with a Tesla-developed AI agent that interprets real-time computer-screen video and keyboard and mouse actions. He said the system was intended to emulate the functions of entire companies, explaining the name as a joke referencing Microsoft.
Macroharder and Minihard are each reported to be configured for AI training, with approximately 220,000 Nvidia GB200 GPUs and 800-gigabit network interface cards. GunRange is also described as a training site with roughly 220,000 GB200 GPUs and 800-gigabit NICs. Its name comes from the site’s previous ownership by the Southaven police department.
Spam is intended for inference workloads rather than training. The former cold-storage and food-distribution warehouse at 2929 Stateline Road West was owned by Hormel Foods, whose brands include Spam. The site is reported to be planned for about 200,000 Nvidia GB300 GPUs.
According to the reported schedule, GunRange was expected to become operational by the end of November 2026, while Spam was due to come fully online by the end of December. The five sites are projected to house around 1.2 million Blackwell GPUs by the end of 2026. These figures and timelines describe the reported build-out plans; they do not confirm that every system is already installed or operating.
Storage details remain unconfirmed
Compared with the GPU and networking figures, information about Colossus 2’s storage architecture is sparse. VAST Data first announced in November 2024 that it was supplying storage to the Colossus data center. DDN subsequently became associated with the project as a storage supplier. The available details do not identify the hardware configurations, deployment locations, capacity allocations or division of work between the two companies.
A separate report in October 2025 described VAST Data storage being used in an xAI Frontier AI Factory expansion, with online availability planned for 2026. That timing suggests a possible connection to a Colossus 2 site, but the relationship has not been established publicly. It should not be treated as confirmation that the facility is part of the five-site cluster.
The suggestion that VAST Data and DDN systems are both serving Colossus 2, potentially at exabyte scale, remains an estimate rather than a disclosed capacity figure. No specific total storage capacity or site-by-site deployment information is available in the reported details. That leaves open how the cluster’s storage is organized to support its distinct training and inference workloads.
As the expansion proceeds, storage will be a significant part of the infrastructure story alongside the stated GPU counts. But without confirmed details on installed systems, usable capacity and the roles of VAST Data and DDN, the storage footprint of Colossus 2 cannot yet be described with precision.