- Startup replaces Nvidia GPUs with customized Arm-powered AI processing {hardware}
- Prometheus packs as much as 128TB of unified LPDDR6 reminiscence onboard
- New server guarantees 1,000× extra reminiscence accessible per processor
Majestic Labs, a startup based in 2023 by former Google and Meta engineers, has unveiled a server constructed to rival Nvidia’s GPU and HBM mixture.
The Tel Aviv-based firm argues that pairing pricey graphics processors with high-bandwidth reminiscence has develop into a essentially memory-bound and dead-end method for AI inference.
Its reply is the Prometheus server, which swaps GPUs for Ignite AI Processing Models combining Arm cores with RISC-V vector and tensor engines.
Newest Movies FromTechRadar
A distinct option to scale reminiscence
Every Prometheus server can home as much as 12 AIUs, sharing between 8 TB and 128 TB of LPDDR6 reminiscence throughout one contiguous, coherent pool.
That reminiscence pool is accessed by customized reminiscence aggregation chiplets linked by copper cables as much as one metre lengthy as a substitute of reminiscence connected on to GPU packages.
A regular 40U rack can maintain 4 such servers, drawing 120 kW whole and cooled by cold-plate liquid methods relatively than air.
By comparability, an Nvidia DGX B300 system with eight Blackwell GPUs affords 2.3 TB of HBM3e plus as much as 4 TB of DDR5 system reminiscence.
Majestic claims its structure due to this fact delivers over 50 occasions extra quick reminiscence than that rival configuration, alongside 1.7 occasions its interconnect bandwidth.
“One Majestic rack holds the quick reminiscence capability of 25 Nvidia NVL72 Vera Rubin racks at a fraction of the facility,” Majestic Lab stated.
“Organizations that might by no means justify hyperscaler infrastructure can now run any workload. In truth, there may be as much as “1000× extra reminiscence per processor.”
Efficiency claims await broader testing
Majestic Labs says the Prometheus server might price between 10 and 50 occasions lower than a GPU system of equal efficiency as soon as it ships subsequent 12 months, whereas consuming much less electrical energy per rack.
The server is designed to be OCP-compliant and can assist PyTorch, vLLM and OpenAI’s Triton frameworks, letting current AI fashions run with out modification.
Based by CEO Ofer Shacham, President Sha Rabii and COO Masumi Reynders, the corporate employs round 40 folks throughout Tel Aviv and Los Angeles and raised $100 million in an A-round late in 2025.
It claims to have already acquired vital orders from giant enterprises, neoclouds and hyperscalers.
But a number of particulars stay unclear, together with what number of reminiscence aggregation chiplets a single server really requires.
If a 128 TB configuration depends on broadly accessible 2 GB LPDDR6 dies, it could want roughly 64,000 of them, implying properly over 100 aggregation chiplets per server.
The numbers Majestic Labs presents are hanging, however they continue to be the startup’s personal projections forward of any impartial benchmarking or shipped {hardware}.
Enterprise patrons contemplating a shift away from established GPU distributors could have to attend for impartial validation to confirm Majestic’s claims.
Through Blocks and Information
Comply with TechRadar on Google Information and add us as a most well-liked supply to get our professional information, opinions, and opinion in your feeds.
