• Announced at CES 2026, full production as of Q1 2026, availability 2H 2026.
  • A suite of 6 platform products: Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, and Spectrum-6.
  • NVIDIA have been focusing more on system & rack level design. They hope to make racks a unit of compute, with the hardware and software in extreme co-design.
ChipWhat it isWhy it exists
Rubin GPUFlagship datacenter GPU (dual-die, HBM4)The core compute engine; ~5x Blackwell inference
Groq 3 LPULanguage Processing Unit. SRAM-based inference chip (from NVIDIA’s Dec 2025 Groq acquisition)500MB on-chip SRAM @ 150 TB/s per chip. Fights the HBM memory wall for decode. The 7th chip of the platform
Vera CPU88-core custom Arm CPUPower-efficient host CPU, tightly coupled to GPUs via NVLink-C2C (chip-to-chip interconnect)
NVLink 6 SwitchScale-up interconnectMakes 72 GPUs in a rack act as one memory/compute domain
ConnectX-9 SuperNICScale-out Network Interface CardNode-to-node networking across racks
BlueField-4 DPUData processing unitOffloads networking/storage/security from CPUs; software-defined infra
Spectrum-6 Ethernet SwitchScale-out Ethernet (co-packaged optics)Connects racks into pods/clusters efficiently
BoardWhat it isWhy it exists
HGX Rubin NVL88-GPU baseboard with NVLinkFor orgs that want Rubin GPUs but their own servers/CPUs — pairs with x86 (Intel/AMD) or Vera. The OEM building block
Vera Rubin NVL44 GPUs + 2 Vera CPUs moduleSmaller HPC/scientific-computing form factor for MGX servers
RackWhat it isWhy it exists
Vera Rubin NVL72Flagship rack: 72 GPUs + 36 Vera CPUs, one NVLink domainMax scale-up — whole rack runs as a single coherent AI engine, no model partitioning. Fully liquid-cooled, cable-free modular trays
Groq 3 LPX256 Groq 3 LPUs in a rack; deployed alongside NVL72Decode is bandwidth-bound — LPX offloads token generation (FFN/MoE) to SRAM while Rubin GPUs handle prefill + attention. ~35x throughput/MW claim. Supersedes the earlier Rubin CPX concept
Vera CPU rack256 Vera CPUs, no GPUsCPU capacity for RL and agentic AI — sandboxes, tool calls, evals, orchestration (~22,500 concurrent sandboxes)
TurnKeyWhat it isWhy it exists
DGX Vera Rubin NVL72The NVL72 rack as a full NVIDIA applianceHardware + Mission Control software + support, zero integration work
DGX Rubin NVL8NVL8 in a liquid-cooled x86 systemOn-ramp to Rubin for orgs staying on x86
DGX SuperPODMulti-rack pod: 14x NVL72 (1,008 GPUs) or 64x NVL8 (512 GPUs)The “AI factory in a box” — networking, storage, software included. Also the reference blueprint for hyperscale deployments

When will the first models trained on either be released?

Rubin GPU