Majestic Labs unveils Prometheus server to replace a rack of Nvidia GPUs

Israeli startup Majestic Labs unveiled its Prometheus server in early August 2026, claiming it can do the work of an entire rack of Nvidia GPUs. The company's central thesis: the main bottleneck in AI infrastructure is not a shortage of compute but limits on memory.

Why does memory matter more than GPUs?

Majestic Labs is going after the "memory wall", a phenomenon well known to computer architects: processors sit idle waiting for data to load from memory instead of doing useful work. The company's founders, who came from Google and Meta and started Majestic Labs in 2023, are convinced that this barrier limits the real-world performance of AI systems more than the number of compute cores does.

Key facts about the company and the product:

• Majestic Labs was founded in 2023 and is headquartered in Tel Aviv
• The team consists of former Google and Meta engineers
• The Prometheus server is positioned as a replacement for an entire rack of Nvidia GPUs
• Architectural bet: priority on memory speed rather than compute

What is wrong with Nvidia GPUs?

Nvidia's accelerators, the H100 and B200, dominate AI infrastructure, but they run into a fundamental limit when working with large language models. The high-bandwidth HBM memory on each chip is limited in capacity: when running inference on a model with tens of billions of parameters, the compute cores often sit idle waiting for data. The typical industry answer is to cluster GPUs together and increase interconnect bandwidth. Majestic Labs considers that a half-measure, not a solution to the problem.

Prometheus rethinks the balance of the system: the server's architecture is built around memory, not GPUs. According to the company, this lets a single machine do work that normally requires an entire rack of accelerators.

"The conversation about AI hardware is stuck on one word: compute.

We want to switch it to another word: memory," is how the company put its position; according to The Next Web, this thesis underlies the Prometheus architecture.

How serious is the challenge to Nvidia?

Several companies already offer alternatives to traditional GPUs in the AI accelerator market, including Groq, Cerebras and SambaNova. Majestic Labs occupies a distinct spot in that group: its bet is not on the speed of matrix operations but on eliminating the memory wall as a systemic constraint. This approach is especially relevant for inference workloads, where large models need fast loading of weights, and that is exactly where GPUs most often lose out on performance per cost.

What this means

If Prometheus holds up under independent testing, Majestic Labs could change the logic of AI infrastructure purchasing: engineers from Google and Meta may pose a real challenge to Nvidia's dominance in AI servers. The main question now is not "is an alternative possible?" but "when will it become cheaper and faster than GPU clusters?"

*Meta has been designated an extremist organization and is banned in Russia.