Majestic Labs, founded by former Google and Meta chip engineers, announced Prometheus on April 28: a single rack-scale AI server offering up to 128 TB of unified high-bandwidth memory—roughly 1,000× more than current leading GPUs—using proprietary Ignite AIUs (hybrid ARM + RISC-V vector/tensor processors) with a uniform shared memory architecture visible at full bandwidth to all compute elements. The design targets the "memory wall" that forces trillion-parameter models and long-context inference onto fragmented multi-rack GPU clusters; Prometheus is designed to consolidate this into one system with no inter-rack communication overhead. Wide availability is expected in 2027; early-customer development is underway now.