NVIDIA launches Rubin platform: six chips for AI supercomputers

At CES on Jan. 5, 2026, NVIDIA announced Rubin. It is a rack-scale AI system built from six co-designed chips: the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet switch. NVIDIA offers Rubin as NVL72 rack systems and HGX Rubin NVL8 systems for extreme agentic AI and ultra-dense inference.

NVIDIA says Rubin GPUs deliver about 50 petaflops of NVFP4 inference throughput. The Vera CPU includes 88 Olympus cores, supports Armv9.2, and uses NVLink‑C2C links. NVLink 6 provides multi‑terabyte-per-second GPU-to-GPU bandwidth; NVIDIA lists per‑GPU and per‑rack figures. The NVIDIA Inference Context Memory Storage Platform serves as an AI‑native KV‑cache tier and reportedly improves long‑context inference performance by multiple factors.

NVIDIA positions Rubin as cutting token costs up to 10x and enabling MoE training with roughly 4x fewer GPUs versus Blackwell. Partners and cloud operators are listed as early adopters, with partner systems rolling out through 2026. Independent coverage flags cost, complexity, and export or regulatory questions in some regions. Enterprises should plan for denser racks, NVLink topologies, and new storage tiers when evaluating Rubin in 2026–2027.

Why It Matters

• Rework rack topology and cabling for NVLink‑centric networking; SRE and infrastructure teams should map NVLink domains, power, and cooling before deployment.
• Re-evaluate inference cost models and placement: if NVIDIA’s token‑cost claims hold, dense rack deployments may be cheaper than distributed GPU fleets—run TCO and latency simulations.
• Plan storage and fabric capacity and monitoring for AI‑native KV‑cache tiers and in‑network compute, since bottlenecks can shift from GPU memory to storage and the fabric.
• Validate regional availability, pricing, and export controls with procurement and legal before committing to on‑prem or cloud Rubin deployments.