Majestic Labs unveils 128TB memory AI server Prometheus alongside new autonomous ASI-EVOLVE framework.

What Happened

Three distinct AI advancements surfaced today across hardware, open-source frameworks, and applied medical AI. Majestic Labs introduced Prometheus, an AI server boasting up to 128TB of memory capacity. Concurrently, a new open-source framework called ASI-EVOLVE was released for autonomous AI optimization. Finally, Beijing Tiantan Hospital launched XiaoJun Doctor 2.0, an applied AI system for rapid brain disease diagnosis via CT scans.

Technical Details

The standout hardware announcement is Majestic Labs' Prometheus, which claims a 1,000x memory capacity advantage over standard Nvidia GPUs, allowing scaling up to 128TB. This directly addresses the von Neumann bottleneck and the "memory wall" that currently forces engineers to rely on complex tensor parallelism and pipeline sharding for large parameter models.

On the software side, ASI-EVOLVE introduces autonomous optimization of training datasets, neural architectures, and algorithms. It effectively automates hyperparameter tuning and neural architecture search (NAS) processes to outperform human baselines. In the applied sector, XiaoJun Doctor 2.0 processes CT imaging to diagnose 94 distinct brain diseases across 11 regions in just 60 seconds.

Why It Matters

From an engineering perspective, the memory wall is the primary constraint in scaling LLMs. If Prometheus can deliver 128TB of unified memory with sufficient bandwidth, it fundamentally changes how we deploy trillion-parameter models, shifting the bottleneck from memory capacity back to raw compute. Coupled with ASI-EVOLVE's ability to self-optimize architectures, the barrier to training state-of-the-art models is lowering drastically. The XiaoJun Doctor 2.0 release validates that these foundational improvements are rapidly translating into highly performant, domain-specific applications.

What to Watch Next

Monitor independent benchmarks for Majestic Labs' Prometheus, specifically focusing on memory bandwidth and interconnect latency—capacity alone doesn't solve the memory wall if data transfer is too slow. Additionally, watch the open-source community's adoption of ASI-EVOLVE to see if its autonomous architecture search scales efficiently without prohibitive compute costs.