en.Wedoany.com Reported - According to an announcement issued by Majestic Labs via Business Wire on April 28, 2026, the company officially launched the Prometheus AI Server—the first server system purpose-built to overcome the AI "Memory Wall." Prometheus integrates up to 128 TB of shared, contiguous memory within a single standard-sized server chassis, connecting the entire memory space to all processing units at full bandwidth. The announcement noted that the performance a single Prometheus system can deliver previously required multiple racks of traditional servers. The product is positioned as an alternative to traditional GPU-centric designs, addressing the problem of processors being underutilized while waiting for data movement across fragmented memory hierarchies.
Prometheus is powered internally by the custom-designed AI processing unit, Ignite. According to the Majestic Labs announcement, Ignite is the first multi-processor chip built around a memory-first architecture, integrating data center-class ARM application cores with RISC-V vector and tensor engines on the same silicon and within a unified memory space. This architecture supports trillion-parameter models, ultra-large context windows of hundreds of millions of tokens, and emerging workloads such as Mixture-of-Experts, agentic AI systems, and graph neural networks within a single node. Previously, such workloads typically required distributed infrastructure to run. 
"Prometheus represents the first time AI infrastructure has been completely reimagined from scratch with memory as a first-class citizen," said Ofer Shacham, Co-founder and CEO of Majestic Labs, in the announcement, adding that as AI model sizes and context windows grow larger and multimodal coordination deepens, the system's advantages become increasingly significant. He stated that Prometheus is built to strip away capacity and bandwidth limitations, enabling organizations to deploy complex AI systems at scale that were previously impossible to run. Sha Rabii, Co-founder and President, stated in the announcement that the industry can no longer afford the efficiency penalties caused by mismatches between machines and workloads, and that AI systems must simultaneously possess massive memory, efficient vector/tensor processing capabilities, and tightly coupled high-performance CPUs.
Regarding software compatibility, Prometheus supports standard frameworks such as PyTorch, vLLM, and OpenAI Triton. According to the Majestic Labs announcement, developers can migrate existing workloads without any code modifications. Masumi Reynders, Co-founder and COO, noted in the announcement that the core insight gained from over a decade of working with internal developers at Google and Meta is foundational: "When customers are forced to choose between performance and productivity, they always choose productivity." Therefore, "Day 1 Productivity" became the design's North Star, meaning the system must work out of the box, with zero switching costs.
According to the Business Wire announcement, Majestic Labs was founded in 2023 by Ofer Shacham, Sha Rabii, and Masumi Reynders, the trio who previously co-led the custom silicon team at Meta Reality Labs and Google's GChips division, cumulatively shipping hundreds of millions of custom chips. Headquartered in San Francisco and Tel Aviv, Israel, the company completed a $100 million Series A funding round in 2025. The announcement noted that the company's team is currently focused on scaling its engineering force in chip design, systems engineering, and AI data science, as well as refining its accompanying software stack.
Jason Bennett, Head of AWS Global Startups and Venture Capital, commented in the announcement that as the scale and complexity of AI models continue to grow, memory bandwidth has become a critical bottleneck. He noted that Majestic Labs is choosing to rebuild AI infrastructure from scratch on AWS, representing a memory-first mindset that is exactly what the industry needs. According to the announcement, Prometheus is currently in an early-stage customer development phase, with plans to open to the broader market within 2026.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
