Majestic Labs is rethinking server architecture from the ground up to solve one of AI's most expensive bottlenecks: memory. In April 2026, the company emerged from stealth with Prometheus, the first AI server purpose-built to break the memory wall. Configurable with up to 128 TB of high-speed, shared memory in a single standard-size server, Prometheus connects 1,000 times more memory per processor than a conventional GPU system, and is the first commercial system capable of running multi-trillion-parameter models, mixture-of-experts architectures, agentic systems, and LLMs with context windows of hundreds of millions of tokens inside a single node. At its core is Ignite, Majestic's proprietary AI Processing Unit, which combines data-center-class ARM cores with RISC-V vector and tensor engines in a unified memory space and runs PyTorch, vLLM, and OpenAI Triton workloads without code changes.
For hyperscalers and enterprises running memory-intensive AI, the result is a dramatically smaller footprint, lower power and cooling requirements, and meaningful reductions in data center spend at a moment when capex from the top five hyperscalers alone is expected to climb from $443 billion in 2025 to more than $600 billion in 2026. Majestic has announced a collaboration with AWS, with early-access customers already in development and wide availability expected in 2027.
The company was founded in late 2023 by Ofer Shacham, Sha Rabii, and Masumi Reynders, the team that previously co-led Google's GChips division and built Meta's Facebook Agile Silicon Team (FAST) inside Reality Labs, and that collectively holds more than 120 patents and has shipped hundreds of millions of custom chips. Like NVIDIA, Majestic develops its entire chip system in-house, but the company is focused specifically on the workloads where the fixed compute-to-memory ratio of conventional GPUs becomes a constraint, including frontier AI applications in financial services and pharmaceuticals.
