An end-run around the AI memory shortage

PROMETHEUS, the AI server unveiled on April 28th by Majestic Labs, makes a pair of audacious claims. The first is that it carries a thousand times the memory capacity of an Nvidia GPU. The second is that it does so without using any of the high-bandwidth memory chips the rest of the industry is fighting over.

Majestic is the work of Ofer Shacham, the firm's chief executive, Masumi Reynders, its operating chief, and Sha Rabii. The three met at Google, where they ran custom-silicon programs together, and later at Meta's Reality Labs. They raised $100m last November from Lux Capital, Bow Wave Capital and Grove. Their new server, fitted with hundreds of a proprietary chip called Ignite, can scale to 128 terabytes of high-speed memory per box — enough, the company reckons, to run models with 5 trillion to 10 trillion parameters without sharding across racks. Majestic says it has booked hundreds of millions of dollars in customer revenue starting in 2027, though it has declined to name any of the customers. The conventional read is straightforward: another inference upstart taking a swing at Nvidia, like Cerebras, which filed to go public on April 17th alongside a $20bn supply agreement with OpenAI, or Groq, which Nvidia paid roughly $20bn to absorb at the end of 2025.

Bypass surgery

The more telling fact about Prometheus is what is not inside it. Each server's 128 terabytes are built from commodity DDR DRAM — the older, simpler, more abundant cousin of high-bandwidth memory, the stacked variant the rest of the AI industry cannot get enough of. Majestic's proprietary interconnect, its founders say, links its processors to those large pools of cheap memory at bandwidth that exceeds HBM's, and at lower power and cost. That is, on its face, a technical claim about latency. It is also a strategic one about supply.