Home / Technology / AI Server Dethrones Nvidia with Arm Power
AI Server Dethrones Nvidia with Arm Power
2 Aug
Summary
- New server uses Arm cores and RISC-V engines instead of GPUs.
- Prometheus offers up to 128TB of unified LPDDR6 memory.
- Startup claims 50x more fast memory than Nvidia's DGX system.
Majestic Labs, a startup established in 2023 by former engineers from Google and Meta, has introduced its Prometheus server, designed to compete with Nvidia's GPU and high-bandwidth memory offerings. The Tel Aviv-based company believes the conventional approach of pairing GPUs with HBM is a bottleneck for AI inference.
Instead, Prometheus utilizes Ignite AI Processing Units, which integrate Arm cores with RISC-V vector and tensor engines. Each server supports up to 12 AIUs and a unified memory pool ranging from 8 TB to 128 TB of LPDDR6. This architecture accesses memory through custom chiplets, aiming for greater efficiency.
Majestic Labs claims their server provides over 50 times the fast memory capacity of an Nvidia DGX B300 system. They project that a single Majestic rack could match the memory capacity of 25 Nvidia NVL72 Vera Rubin racks, consuming less power. The company asserts this could enable organizations to run any workload previously reserved for hyperscalers.
The Prometheus server is OCP-compliant and will support major AI frameworks like PyTorch and vLLM, allowing existing models to run without changes. Founded by Ofer Shacham, Sha Rabii, and Masumi Reynders, Majestic Labs has around 40 employees and secured $100 million in Series A funding in late 2025.
While Majestic Labs projects significant cost reductions and lower power consumption per rack compared to GPU systems, these performance claims are based on their own projections. Independent benchmarking and broader testing will be necessary to validate these assertions. The company has reportedly received substantial orders from major enterprises and hyperscalers.
