Majestic Labs is tackling the significant memory limitations of large language models (LLMs) with its upcoming Prometheus server, which boasts a remarkable 128 terabytes of memory—more than 60 times that of Nvidia's DGX B300 server. Co-founder Sha Rabii emphasizes that traditional AI servers often over-provision compute resources while underutilizing memory. To address this, Prometheus employs a DRAM-centric architecture, utilizing LPDDR6 and a unique memory interface with miniature copper cables to achieve bandwidth of up to 25.6 terabytes per second. The server features 12 Ignite chips, which combine ARM and RISC-V cores for efficient LLM processing. Importantly, Prometheus will support existing AI frameworks like PyTorch and Triton without code modifications, easing customer adoption. Majestic Labs claims that this innovative design will significantly lower capital and power costs, with expectations of a 10 to 50 times reduction. Although pricing details remain undisclosed, Prometheus is anticipated to ship in 2027.