Majestic Labs raises $100M

AI servers are running into a boring-sounding problem with huge consequences — memory. Chips keep getting faster, but the data they need is still hard to move, hard to share, and expensive to keep close. That gap is now big enough that a startup can build an entire company around it. Majestic Labs did exactly that. The company used its April 28, 2026 Prometheus launch to show what its earlier $100 million raise was really funding: a memory-first AI server meant to break the “memory wall.” (finance.yahoo.com) ### What is the memory wall? The short version is that processors can only work on data they can reach fast enough. Modern AI systems have tons of compute, but memory is scattered across GPUs, CPUs, switches, servers, and racks. So the expensive chips sit around waiting for data instead of doing useful work. That is the wall. Not a lack of math h(finance.yahoo.com)te and underbuilt the memory side of the machine. (finance.yahoo.com) ### Why does that matter more now? Because models got fatter in every direction. Bigger parameter counts. Bigger datasets. Bigger context windows. Inference also keeps more state live, especially when you want long-context or multi-user workloads. That creates a nasty economics problem. Companies buy more accelerators just to get enough memory at(finance.yahoo.com)ploding while utilization stays worse than it should be. (finance.yahoo.com) ### So what did Majestic actually build? Prometheus is not just a chip. It is a full server architecture with custom silicon and a software stack built around disaggregated, shared memory. The company says one standard-size server can be configured with up to 128 TB of memory, and that the whole memory pool is uniform, contiguous, and available at full bandwidth to all processing elements. (finance.yahoo.com) Basically, instead of treating memory as a bunch of small private closets attached to different processors, Majestic wants one giant room everyone can access efficiently. ### Why is 128 TB such a big deal? Because the comparison point is not ordinary DRAM in a storage box. The comparison point is the amount of fast, useful memory AI processors ca(finance.yahoo.com)ic systems and can collapse the memory capacity of roughly 10 racks into one server. (finance.yahoo.com) If that holds up in real deployments, the win is not just “bigger number.” It means fewer machines, fewer network hops, less orchestration pain, and less wasted compute. ### Who is behind it? The founders are Ofer Shacham, Sha Rabii, and Masumi Reynders. They previously led silicon efforts at Google and Meta, including teams tied to custom chip p(finance.yahoo.com)to redesign the box itself. (cnbc.com) ### Where does the $100 million fit in? The funding was announced in November 2025, with Bow Wave Capital leading the Series A and Lux Capital among the backers. The April 2026 news is the product reveal — Prometheus is the thing that money went into building. (cnbc.com) That date split matters because this is not (cnbc.com)ncrete. ### What is the catch? The catch is always the same in infrastructure. Claims are easy. Adoption is hard. Majestic says the system keeps programmability and cuts total cost and power, but the real test is whether hyperscalers and big enterprises will cha(cnbc.com)customers in 2027. (finance.yahoo.com) ### Bottom line? Majestic is making a sharp bet: the next AI bottleneck is not more flops, but more usable memory. If that bet is right, Prometheus is not just another AI server — it is a challenge to the GPU-centered shape of the whole data center. (finance.yahoo.com)