Former Google and Meta engineers build memory-first AI server to challenge Nvidia's GPU dominance

Majestic Labs AI unveiled Prometheus, a memory-first AI server architecture aimed at challenging Nvidia’s GPU-centric approach.

The system emphasizes huge memory capacity—up to 128 TB—rather than raw compute, targeting 5 to 10 trillion-parameter models.

By using commodity DRAM and proprietary interconnects, the company says it can sidestep the memory and supply bottlenecks of GPU-based systems.

The startup has raised $100 million and is already signing customers, betting that large-model inference is constrained more by memory than by compute.