Majestic Labs raises $100M Series A for memory-pooled AI server architecture
What's the deal? Majestic Labs, an AI chip startup based in Los Altos, California, has raised $100M in Series A funding for a memory-pooled server architecture designed for AI inference. The company's system can offer up to 100TB of DRAM per accelerator — far beyond what current HBM technology delivers — by disaggregating memory from compute.
Founded in 2023 by Masumi Reynders, Ofer Shacham, and Sha Rabii — all former colleagues from Google's and Meta's silicon divisions — Majestic Labs claims it can pack the memory capacity and bandwidth of 10 racks of state-of-the-art GPUs into a single server.
Why now? AI compute is scaling faster than memory bandwidth, but inference for most large models is memory bandwidth-limited. The founders bet that models would keep growing, context lengths would get longer, and the economics of today's compute-first architectures would become untenable.
"We knew from the get-go that it was foolish to try and play Nvidia's game and just out-execute them," co-founder Rabii told EE Times.
Current approaches — HBM for high bandwidth with limited capacity, or CXL for large capacity with low bandwidth — don't fully meet AI's requirements. Majestic's memory-first architecture uses pooling with proprietary high-bandwidth, low-latency interconnects to bridge that gap.
The company is building two chips: a memory interface chiplet that sits next to both compute and memory, and a many-core AI accelerator. A single server pairs over 100TB of standard LPDDR with up to 12 accelerator chips, presenting the entire memory space as a single flat address space accessible at uniform bandwidth and latency.
"That really simplifies programming," Rabii said, noting that GPU-based servers have multiple memory tiers that make workload optimisation a complex software task. "There are entire companies whose reason for existence is to help other companies map their workloads more effectively and efficiently to GPU clusters. We think that is an unnecessary task."
What could go wrong? Majestic is attempting to build both a novel interconnect and a new accelerator chip simultaneously — a massive engineering undertaking. It will also need to convince developers and cloud providers to adopt an unfamiliar architecture in a market where Nvidia's CUDA ecosystem dominates.
Delivering on the promise of uniform memory bandwidth across 100TB is an extraordinary technical claim. If latency or reliability falter at scale, the flat-memory advantage could erode quickly.
The signal: Majestic Labs' $100M Series A is notable for a company founded only in 2023, already classified by Dealroom as "late growth" stage — an unusually rapid trajectory that underscores how urgently investors want alternatives to Nvidia's compute-first paradigm. With AI inference costs increasingly dominated by memory bandwidth constraints rather than raw compute, capital is flowing toward startups that rethink server architecture from first principles rather than iterating on existing GPU designs.
