Examining Nvidia's 60 exaflop Vera Rubin POD — how seven chips underpin company's 40 rack AI factory supercomputer
Every chip in the stack has a defined job, from training trillion-parameter models to caching inference tokens on NVMe storage.
Nvidia announced seven chips in full production at GTC 2026 on Monday, composing the Vera Rubin platform that the company intends to ship in the second half of this year.
Rather than a single product launch, several announcements covered the full silicon stack required to build what Nvidia now calls an AI factory: GPUs, CPUs, a dedicated inference accelerator, networking ASICs, a data processing unit, and an Ethernet switch. All seven are designed to operate as a single co-designed system across five rack types, scaling from individual racks to 40-rack PODs delivering 60 exaflops of compute.
The so-called AI factory is a massive shift in how Nvidia packages and sells its hardware, where the unit of compute is no longer a GPU or even a server; it’s the rack, and increasingly, the POD. Each of the seven chips fills a specific architectural role — and understanding what each does is the fastest route to understanding what Vera Rubin fundamentally is.
The compute layer: Rubin GPU, Vera CPU, and Groq 3
Three chips handle the core compute workload, each optimized for a different phase of the AI pipeline.
The Rubin GPU is the training and inference workhorse built on TSMC's 3nm process. Each GPU uses a dual-die design packing 336 billion transistors, carries 288 GB of HBM4 memory with 22 TB/s of bandwidth, and delivers 50 PFLOPS of inference compute and 35 PFLOPS of training compute in the NVFP4 format. Those figures represent 5 and 3.5 times improvements over Blackwell, respectively.
In the flagship Vera Rubin NVL72 rack, 72 Rubin GPUs connect via NVLink 6 to behave as a single accelerator. Nvidia claims that the NVL72 can train Mixture-of-Experts models with one quarter the GPU count required by Blackwell, and cut inference token costs by 10 times.
The Vera CPU, meanwhile, is Nvidia's first data center CPU built from the ground up. It uses 88 custom Arm-based Olympus cores with Spatial Multithreading for 176 threads, up to 1.5TB of SOCAMM LPDDR5X memory, and 1.2 TB/s of memory bandwidth. Vera connects to Rubin GPUs via NVLink-C2C at 1.8 TB/s of coherent bandwidth, which is seven times faster than PCIe Gen 6. Its role in the rack is orchestration: scheduling workloads, routing KV cache data, managing context, and running the control plane for agentic AI workflows. It also handles reinforcement learning environments and CPU-native workloads.
The Groq 3 LPU — purpose-built for low-latency decode-phase inference — is the most unexpected addition to the platform, and a direct product of Nvidia's $20 billion acquisition of Groq in December. Where Rubin GPUs offer massive memory capacity through HBM4, the Groq 3 trades capacity for bandwidth: Each LPU can carry roughly 500MB of stacked SRAM and delivers approximately 80 TB/s of bandwidth per chip. The Groq 3 LPX rack houses 256 LPUs with about 128GB of aggregate on-chip SRAM and 640 TB/s of scale-up bandwidth.
Rubin GPUs handle the compute-heavy prefill phase of inference, processing long input contexts, with the Groq 3 LPUs stepping in to handle the decode phase, generating output tokens at low latency. Nvidia claims the combination delivers 35 times higher inference throughput per megawatt and 10 times more revenue opportunity for trillion-parameter models, compared with running both phases on GPUs alone.
The fabric: NVLink 6, ConnextX-9, and Spectrum-6
As for moving data between chips at rack scale and between racks at cluster scale, Nvidia has architected this via three dedicated networking ASICs.
The NVLink 6 Switch, due to be upgraded to a 7th generation, handles scale-up connectivity within the rack. Each switch delivers 3.6 TB/s of bidirectional bandwidth per GPU, doubling Blackwell's NVLink performance, while a single switch tray provides 28.8 TB/s of total switching bandwidth and 14.4 TFLOPS of FP8 in-network compute, which accelerates collective operations like the all-to-all communication patterns used in MoE routing. Nine switch trays per NVL72 rack deliver 260 TB/s of aggregate scale-up bandwidth.
The ConnectX-9 SuperNIC provides a scale-out networking endpoint at 1.6 Tb/s throughput per GPU. Where NVLink 6 connects GPUs within a rack, ConnectX-9 connects racks, linking NVL72 systems into multi-rack clusters via either Nvidia's Spectrum-X Ethernet or Quantum-X800 InfiniBand fabrics. Each compute tray uses eight ConnectX-9 NICs to deliver the aggregate bandwidth quoted for the tray's four GPUs.
The Spectrum-6 Ethernet Switch is the switching silicon and the backbone of the Spectrum-6 SPX networking rack, delivering 102.4 Tb/s of aggregate bandwidth, and is Nvidia's first switch to use co-packaged optics, employing silicon photonics to reduce optical power consumption. Nvidia claims five times improved power efficiency and 10 times improved resiliency over prior Spectrum-X generations. Available in two configurations, the SN6800 offers 512 ports of 800G Ethernet or 2,048 ports at 200G.
The infrastructure layer: BlueField-4
The final chip, BlueField-4 DPU, functions as a specialized processor that handles networking and storage tasks that would otherwise consume CPU and GPU cycles.
A dual-die package that combines a 64-core Grace CPU with an integrated ConnectX-9 NIC, the BlueField-4 DPU offloads networking, storage, encryption, virtual switching, telemetry, and security enforcement from the main compute path, running Nvidia's DOCA software framework for infrastructure services. Nvidia says it features twice the bandwidth, three times the memory bandwidth, and six times the compute of BlueField-3.
BlueField-4 underpins the new BlueField-4 STX storage rack, which implements Nvidia's CMX context memory storage platform, essentially extending GPU memory into NVMe storage to cache the key-value data generated by agentic AI workflows. As context windows grow to hundreds of thousands, and in some cases millions, of tokens, operations on the KV cache are becoming a bottleneck, and the STX rack is designed to break through it. Nvidia claims up to five times higher inference throughput when the STX storage tier is deployed alongside compute racks, via the new DOCA Memos software framework.
Five racks, one supercomputer
All seven chips map into five rack types that compose the Vera Rubin POD: the NVL72 for core training and inference (72 Rubin GPUs, 36 Vera CPUs); the Groq 3 LPX rack for decode acceleration (256 LPUs); the Vera CPU rack for RL and orchestration (256 Vera CPUs); the BlueField-4 STX rack for KV cache storage; and the Spectrum-6 SPX rack for Ethernet networking. A full POD spans 40 racks, 1,152 Rubin GPUs, nearly 20,000 Nvidia dies, 1.2 quadrillion transistors, and 60 exaflops.
Vera Rubin-based products are scheduled to ship in the second half of 2026.
Luke James is a freelance writer and journalist. Although his background is in legal, he has a personal interest in all things tech, especially hardware and microelectronics, and anything regulatory.