A startup says the AI bottleneck isn’t compute. It’s memory, and it ditched the GPU to prove it.

Majestic Labs, a Tel Aviv startup, claims its Prometheus server replaces GPUs with custom AI processing units and large pooled LPDDR6 memory, offering over 50 times more fast memory than Nvidia's DGX B300. The system supports standard AI frameworks and aims to overcome memory bottlenecks in AI inference, though no hardware has shipped yet and claims await independent verification.