Majestic Labs' Prometheus Aims to Shatter AI's Memory Wall

SAN FRANCISCO & TEL AVIV, Israel – April 28, 2026 – Startup Majestic Labs today emerged from stealth to announce Prometheus, a new class of AI server designed to confront what it calls the single greatest obstacle to AI's progress: the 'memory wall.' The company claims its memory-first architecture provides an unprecedented 128 terabytes (TB) of high-speed memory in a single server, a figure that dwarfs current industry offerings and promises to radically alter the economics and capabilities of artificial intelligence.

Prometheus is a direct challenge to the GPU-dominated landscape, which has focused on an arms race for computational power. Majestic Labs, founded by a team of veteran chip designers from Google and Meta, argues this focus has created a massive inefficiency, with powerful processors often sitting idle while waiting for data. By rebalancing the system around memory, the company aims to deliver the performance of multiple data center racks in one unit, drastically cutting costs and power consumption.

“In the early days of AI, the industry ran workloads on machines that were never actually built for AI,” said Sha Rabii, Co-Founder and President of Majestic Labs, in a statement. “The industry can no longer afford the compromise in efficiency that results from this ill-fitting pairing considering the scale that AI is reaching today.”

The Billion-Dollar Bottleneck

The AI industry is grappling with an escalating crisis of cost and energy. The relentless growth of model size and complexity has driven a data center construction boom of unsustainable proportions. Market analyses project that capital expenditures from just the top five hyperscalers will surge past $600 billion in 2026, with over 75% of that investment earmarked for AI infrastructure. Simultaneously, global data center electricity consumption is on track to double by 2030, with AI as the primary driver.

This explosive growth is constrained by the 'memory wall'—the widening gap between a processor's ability to compute and the system's ability to feed it data. While GPUs have become exponentially more powerful, their on-chip memory capacity has grown at a much slower pace. This forces developers to use complex and inefficient techniques to spread large AI models across dozens or even hundreds of GPUs, creating a significant data-shuffling bottleneck that consumes power and limits performance.

Majestic Labs' approach tackles this problem head-on. The Prometheus server is designed to create a vast, unified pool of memory that is equally and instantly accessible to all processing elements, eliminating the data traffic jams that plague conventional systems.

A Memory-First Revolution

The specifications claimed for Prometheus represent a paradigm shift. The headline figure of 128 TB of memory capacity in a standard server rack is staggering when compared to the current market leaders. For instance, NVIDIA's state-of-the-art DGX B200 system, considered a powerhouse for AI, offers a total of 1.4 TB of specialized GPU memory across eight of its most advanced B200 accelerators. Prometheus claims to offer nearly 100 times that capacity in a single, unified space.

“Prometheus represents the first ground-up reimagination of AI infrastructure with memory as a first-class citizen,” added Ofer Shacham, Co-Founder and CEO. “AI brains become better with larger model and context sizes and when multi-modalities work in tandem. Bigger truly is better in this context. We built Prometheus to remove capacity and bandwidth limits so organizations can deploy sophisticated AI systems that were previously impractical, if not impossible, to run at scale.”

At the heart of the server are Majestic's proprietary AI Processing Units (AIUs), named Ignite. These custom chips feature a unique multiprocessor design that combines datacenter-class ARM cores for general-purpose tasks with highly efficient RISC-V vector and tensor cores optimized for AI workloads. According to the company, this holistic design allows all processing to occur on the same silicon and within the same memory space, further reducing data movement and improving efficiency.

Unlocking the Next Frontier of AI

The practical implications of such a massive memory pool are profound. It could enable organizations to run cutting-edge AI models that are currently confined to research papers or the world's largest tech companies. This includes multi-trillion-parameter models, which today require vast and costly clusters of GPUs.

With Prometheus, a model of this scale could potentially reside within a single server, dramatically simplifying deployment and accelerating inference speed. This architecture is also ideal for Large Language Models (LLMs) with massive context windows—the ability to remember and process hundreds of millions of tokens of information, far beyond the few hundred thousand typical of today's models. This could lead to AI assistants with near-perfect long-term memory and sophisticated reasoning capabilities.

Furthermore, the design is well-suited for complex Mixture-of-Experts (MoE) models and large-scale graph neural networks, which are notoriously difficult to run efficiently on conventional hardware due to their sparse data access patterns.

The Critical Path to Adoption

For any new hardware to succeed, however, it must be easy for developers to use. NVIDIA's dominance is built not just on its chips, but on its mature CUDA software ecosystem. Majestic Labs appears acutely aware of this challenge, emphasizing a commitment to seamless integration.

“After over a decade of collaboration with internal developers and researchers working at the intersection of AI algorithms and custom silicon at Google and Meta, we learned that when customers have to choose between performance and productivity, they choose productivity,” said Masumi Reynders, Co-Founder and Chief Operating Officer. “The system just has to work, with no switching cost.”

To that end, the company states its fully programmable system supports industry-standard frameworks like PyTorch, vLLM, and OpenAI’s Triton, allowing developers to deploy existing code without modification. This “Day 1 productivity” claim, if validated, would significantly lower the barrier to adoption. Bolstering its credibility, the startup also announced a collaboration with Amazon Web Services (AWS).

“As AI models continue to grow in scale and complexity, memory bandwidth has emerged as a critical bottleneck,” said Jason Bennett, VP and Global Head of Startups and Venture Capital at AWS. “We are excited that Majestic Labs has chosen AWS to help them reimagine AI infrastructure from the ground up.”

Explaining the product's name, Reynders noted the parallel between the Greek Titan who gave fire to humanity and Majestic's mission to bring efficient, advanced AI infrastructure to any organization. This philosophy is reflected in the server's design. “We had to reimagine the entire architecture from first principles, putting memory first and building everything else around it,” Rabii stated. “The name reflects that ground-up and almost defiant design philosophy.”

Prometheus is currently in development with early access customers, and the company expects wide availability to begin next year.