NVIDIA’s Rubin AI Accelerators Set to Enter Market as Early as September, Just Six Months After Blackwell Ultra
In a surprising acceleration of its already aggressive roadmap, NVIDIA appears poised to release its next-generation Rubin AI GPUs and Vera CPUs as early as September 2025, marking a rapid six-month leap following the debut of the Blackwell Ultra series. The report, originally published by Commercial Times (CTEE), indicates that customer sampling for the Rubin R100 AI chips will begin imminently, signaling a major milestone in NVIDIA’s relentless pursuit of AI leadership.
Rubin R100 GPUs: NVIDIA’s First Chiplet-Based AI Architecture
The Rubin architecture represents a fundamental leap for NVIDIA, not just in terms of performance, but also in how the silicon is constructed. The R100 GPUs will:
Leverage HBM4 memory: A significant upgrade from the current HBM3E, promising massive bandwidth increases essential for large-scale AI workloads.
Adopt a chiplet architecture: Marking NVIDIA's first venture into a modular design paradigm, Rubin will utilize a 4x reticle chiplet design, compared to Blackwell’s 3.3x.
Be manufactured on TSMC’s 3nm N3P node: Combined with CoWoS-L packaging, this allows for higher density and improved thermal and power efficiency.
These changes are not just iterative improvements—they represent a ground-up rework of the GPU stack tailored specifically for AI inference and training needs in modern data centers.
Vera CPUs: The Successor to Grace with Advanced ARM Cores
The Rubin GPU lineup will be paired with the Vera CPUs, which are expected to replace NVIDIA’s Grace processors. Vera chips are rumored to be built on next-generation ARM cores and will likely deliver significant improvements in performance-per-watt and multi-core efficiency, continuing NVIDIA’s move toward integrated AI computing platforms.
This Rubin–Vera combination will define the foundation of NVIDIA’s next-gen DGX and HGX platforms, with a strong emphasis on power efficiency, which has become increasingly critical as data centers face mounting energy demands.
A Condensed Roadmap Raises Supply Chain Questions
NVIDIA’s CEO Jensen Huang previously confirmed a “one-year cadence” for AI GPU architecture updates. However, the Rubin timeline appears even more aggressive. With Blackwell Ultra having just launched earlier this year, Rubin’s September sampling implies a six-month release cycle, cutting the expected interval in half.
While this showcases NVIDIA's engineering prowess, it also presents logistical challenges. The company’s partners, OEMs, and hyperscalers may struggle to integrate new architectures at such a rapid pace, particularly when earlier platforms are still in the ramp-up or optimization phase. Previous rollouts—including Hopper and early Blackwell samples—faced teething issues, and the compressed cycle may heighten those risks.
Performance, Power, and AI Leadership
At the core of Rubin’s design philosophy is performance per watt, aligning with global trends toward more sustainable AI infrastructure. NVIDIA aims to dominate not only in sheer compute power but in the ability to do more with less—an imperative as LLMs and other advanced models continue to scale exponentially.
Despite potential adoption hurdles, Rubin's HBM4 integration, advanced packaging, and modular architecture all suggest a new era of scalable, energy-conscious AI supercomputing, placing NVIDIA further ahead of rivals like AMD and Intel in the data center AI arms race.
Will Rubin set a new standard in AI performance? Can the market keep pace with NVIDIA’s breakneck innovation cycle? Let us know what you think about this upcoming generation in the comments.
