Majestic Labs unveils Prometheus AI server with 128TB memory to challenge Nvidia's dominance

Majestic Labs Takes Aim at Nvidia's Memory Wall

Majestic Labs, a Tel Aviv startup founded in 2023 by former Google and Meta engineers Ofer Shacham, Sha Rabii, and Masumi Reynders, has unveiled Prometheus, an AI server that abandons GPUs entirely1

1

2

2

. The company's bold pitch targets what it calls the AI memory wall—the bottleneck created when running large AI models is limited by available fast memory rather than raw compute power. With $100 million in funding raised late last year, Majestic Labs now has about 40 staff across Tel Aviv and Los Angeles working to challenge Nvidia's dominance from an unconventional angle1

1

2

2

.

GPU-Free Server Architecture Built on Shared Memory

Prometheus represents a fundamental departure from conventional AI inference server platform design. Instead of pairing expensive GPUs with scarce high-bandwidth memory, the AI server uses what Majestic calls Ignite AI Processing Units1

1

. Each Ignite unit combines Arm cores with RISC-V vector and tensor engines. Up to 12 of these units sit in a single server, all connected to one shared pool of 8TB to 128TB of LPDDR6 memory—the same affordable memory found in smartphones, not the costly high-bandwidth memory that GPUs depend on1

1

2

2

.

The shared memory architecture relies on custom aggregation chiplets connected through copper cables up to one metre long, creating what Majestic claims is a single coherent fast-memory pool far larger than any GPU box can address1

1

. Building a 128TB pool from 2GB LPDDR6 dies would require roughly 64,000 of them, implying more than a hundred aggregation chiplets working in concert within a single server1

1

.

Striking Performance Claims Against Nvidia Hardware

Majestic Labs positions Prometheus as a direct challenger to Nvidia's flagship systems. An Nvidia DGX B300 with eight Blackwell GPUs carries 2.3TB of high-bandwidth memory. Prometheus offers more than 50 times as much fast memory capacity, according to the company, along with 1.7 times the interconnect bandwidth1

1

2

2

. The comparison extends to rack-level deployments: Majestic claims one Prometheus rack matches 25 of Nvidia's Vera Rubin racks for fast memory, at a fraction of the power consumption1

1

2

2

.

On cost and power efficiency, the startup says Prometheus could reduce expenses by 10 to 50 times while also cutting power use significantly2

2

. For AI model serving workloads where inference is often memory-bound rather than compute-bound, these figures suggest a meaningful shift in economics.

Software Compatibility and Enterprise Adoption

Despite the radical hardware departure, Majestic Labs has built Prometheus to open standards. The platform supports PyTorch, vLLM, and OpenAI's Triton, and models developed for GPUs are designed to run on Prometheus without modifications1

1

2

2

. This compatibility matters for enterprises already invested in GPU-based workflows. The company reports it has already secured orders from large enterprises, neoclouds, and hyperscalers, though specific customers haven't been named1

1

.

Buyers will need evidence that the software stack is mature and fits seamlessly into existing MLOps workflows before committing at scale2

2

. Select customers are slated to receive shipments in 2027, giving the company roughly two years to prove the platform in production environments2

2

.

The Unproven Promise and What Comes Next

Every performance figure Majestic Labs has shared carries the same caveat: none has been independently verified, and no hardware has shipped1

1

2

2

. Keeping 64,000 memory dies coherent across more than a hundred aggregation chiplets presents significant engineering challenges. Buyers are likely to wait for independent benchmarks before switching infrastructure1

1

.

Majestic Labs joins a growing line of startups attacking Nvidia from unconventional angles—optical chips, edge silicon for inference, open networking gear. Each picks a different weakness in Nvidia's architecture. What makes Prometheus distinct is its focus on the AI memory wall, betting that inference workloads are increasingly starved for accessible memory rather than compute cycles1

1

. Whether this GPU-free server can deliver on its bold claims will become clear when hardware reaches customers and third-party testing begins in earnest over the next two years.