Every architecture generation, the question is the same: what's next? For NVIDIA, the answer after Blackwell is Vera Rubin, named for Vera Rubin, the astronomer who confirmed dark matter. It's not just a GPU, it's a platform: a custom CPU (Vera) connected to a next-gen GPU (Rubin) over NVLink-C2C, with six other chips codesigned as one system. It's in production ramp as of mid-2026.
The Vera CPU
Vera is NVIDIA's own CPU, built on 88 custom "Olympus" cores (Armv9.2). It's the host processor, designed to feed the GPU. The key innovation is NVLink-C2C (chip-to-chip), which connects Vera to Rubin at 1.8TB/s, far faster than PCIe. That means the CPU can stream data to the GPU at memory-fabric speeds, and CPU-side pre/post-processing has low-latency access to GPU memory without PCIe bottlenecks.
Vera's Olympus cores deliver 2x single-threaded performance and 40% lower memory latency vs competing designs. For inference, that matters: the CPU handles scheduling, KV-cache management, tokenization, and orchestration, and CPU bottlenecks are real in production inference.
The Rubin GPU
Rubin is Blackwell's successor, with 5th-gen tensor cores, 50 PFLOPS of NVFP4 (vs B200's ~9 PFLOPS dense FP4), a 3rd-gen Transformer Engine with hardware-accelerated compression, and NVLink 6 at 3.6TB/s per GPU, double NVLink 5. The Rubin Ultra variant is a multi-die design for the Kyber rack with 576 GPUs.
The NVL72 rack
The Vera Rubin NVL72 rack is the headline form factor: 36 Vera CPUs + 72 Rubin GPUs + 18 BlueField-4 DPUs per rack, with 260TB/s aggregate NVLink and 1,046TB of fast memory. A SuperPOD is 14 racks = 1,008 Rubin GPUs = 50.4 exaflops FP4. The entire rack operates as a single coherent AI engine with unified memory, eliminating model partitioning.
Why it matters for inference
Vera Rubin is the disaggregation play taken to its logical end. The CPU is fast enough to be a real prefill engine, and the GPU is fast enough to be a real decode engine. The NVLink-C2C between them means the KV cache transfer between CPU and GPU is no longer a bottleneck.
The numbers that matter: 10x reduction in inference token cost vs Blackwell, and 10x throughput per megawatt on DeepSeek-R1 (CoreWeave benchmark). For anyone running inference at scale, the roadmap matters: Blackwell is the present, Rubin is the near future, and the architectures you learn now transfer directly. The same roofline, the same KV cache math, the same disaggregation, just faster.
Sources
- NVIDIA: DGX SuperPOD sets the stage for Rubin: the CES 2026 announcement.
- NVIDIA: Vera Rubin driving performance per watt: production ramp, 10x per-megawatt.
- NVIDIA: gigawatt AI factories: Kyber, Rubin Ultra, 800V DC.
- NVIDIA: Blue Lion on Vera Rubin: the superchip concept.