DooDooLamb News

Startup swaps costly AI GPUs for Arm cores and up to 128TB of 'cheap' LPDDR6 RAM instead of expensive HBM to smash through the memory wall

Brief published August 3, 2026 ยท Original source published August 2, 2026

Original reporting by Efosa Udinmwen at techradar.com.

Automated brief. Verify important details at the original source.

Startup swaps costly AI GPUs for Arm cores and up to 128TB of 'cheap' LPDDR6 RAM instead of expensive HBM to smash through the memory wall

What happened

Majestic Labs unveiled a server called Prometheus that replaces Nvidia GPUs with custom Arm-based AI processing hardware. Instead of high-bandwidth memory typically paired with GPUs, the system uses LPDDR6 RAM and claims to support up to 128TB of unified memory onboard. The startup positions this as a lower-cost alternative to GPU-based infrastructure, asserting the design delivers roughly 1,000 times more memory capacity than conventional AI server configurations.

Why it matters

GPU memory constraints force engineers to shard large models across many accelerators, adding complexity and cost. If Prometheus delivers on its claimed capacity at lower price points, it could offer a practical path for running memory-intensive workloads without expensive HBM-equipped hardware. Builders evaluating inference infrastructure for large models would have a new cost-performance tradeoff to consider.

What to watch

Independent benchmarks comparing Prometheus throughput and latency against GPU-based systems will be the first real test of whether LPDDR6 bandwidth is sufficient for production AI workloads at scale.

Original source