When Ofer Shaham left his role leading chip development for Meta’s metaverse glasses, he co-founded Majestic Labs with two former executives, Shah Rabii and Massumi Reinders. The startup, founded in 2023 and now based in a new office tower in Ra’anana, is trying to challenge Nvidia from an unexpected angle, with a graphics processor and server architecture designed to deliver far more AI work per unit of electricity.
Majestic says it has raised more than $100 million, employs 50 people, including 25 in Israel, and counts Lux Capital, Bow Wave, Hetz, Grove, Tal Ventures, Upfront, Idanlayr Capital and SBI among its investors. The company recently unveiled its products: the “Ignite” GPU, built on Arm, open-source RISC-V and proprietary technology, and the “Prometheus” server, which packs in large amounts of memory to avoid the industry’s “memory wall.” Shaham says the goal is to “bring AI to everyone, not just the cloud giants.”
The company plans to begin manufacturing in the coming months and launch its servers and chips in mid-2027. Industry estimates suggest it has already received paper orders worth several hundred million dollars, with customers expected to get up to 50 times more AI tokens per megawatt than current systems. Shaham says one customer is building a 500-megawatt server farm and cares more about tokens per megawatt than hardware price.
Shaham argues that large language models from Gemini, Claude and GPT are constrained by limited GPU memory, forcing cloud companies to buy huge numbers of underused servers and driving an energy crisis. He said hyperscalers Amazon, Google, Microsoft, Meta and Tesla spent $443 billion on capex last year and are expected to reach $700 billion this year. He does not see competition from Nvidia, Cerebras, Intel or AMD as a problem, saying the market is moving toward heterogeneous computing, where different chips handle different workloads.
Shaham, Rabii and Reinders bring more than 20 years of hardware experience from IBM, Stanford, Google and Meta. Shaham said the Israeli and U.S. teams work in shifts across Ra’anana and Silicon Valley, and that the company is focused on solving the memory bottleneck rather than chasing raw speed.
