Beyond GPUs: How NVIDIA’s Vera Rubin, LPU, and NemoClaw Redefine AI at GTC 2026

During the annual NVIDIA GTC, CEO Jensen Huang delivered a two‑hour keynote that introduced a suite of new AI hardware and software, positioning the announcements as the next wave of AI infrastructure.

Not Just GPUs

The centerpiece of the hardware roadmap is the Vera Rubin platform, which comprises seven breakthrough chips, five rack configurations, and a supercomputer. Key components highlighted are the Rubin GPU, the Groq‑based LPU (LPX), and the Vera CPU.

The Rubin GPU, built for Agentic AI, debuted in January and features a third‑generation Transformer Engine with hardware‑accelerated adaptive compression, delivering 50 petaflops of NVFP4 compute and supporting NVLink‑72 full interconnect.

Following NVIDIA’s 2025 acquisition of Groq technology for $20 billion, the company confirmed that LPU will complement rather than replace GPUs. In large‑scale deployments, an LPU cluster acts as a single massive processor for deterministic inference, and when paired with Rubin NVL72, the GPU and LPU jointly compute each token layer, boosting decoding throughput. The LPX rack, equipped with 256 LPU processors, targets low‑latency, large‑context agent systems and claims up to 35× higher inference throughput per megawatt for trillion‑parameter models.

The Vera CPU, marketed as the first processor designed for Agentic AI and reinforcement‑learning workloads, offers twice the efficiency of traditional rack‑level CPUs and 50 % faster execution, enabling higher AI throughput and response speed for large‑scale services such as coding assistants and consumer‑grade agents. Jensen Huang emphasized that the CPU is now a driver of models, not merely a supporter.

To help developers navigate the increasingly complex GPU and accelerator landscape, HyperAI launched a "GPU Leaderboard" that standardizes cross‑vendor, cross‑architecture comparisons based on AI, large‑model, and HPC workloads.

NemoClaw: One‑Command Optimization for OpenClaw

On the software side, NVIDIA introduced NemoClaw, an extension of the OpenClaw open‑source project. Described by Huang as the "personal AI operating system," NemoClaw uses the NVIDIA Agent Toolkit to optimize OpenClaw with a single command, integrating it into NVIDIA’s ecosystem.

NemoClaw installs OpenShell, provides open‑source models, and creates an isolated sandbox that enhances data privacy and security for autonomous agents. It supports any programmable agent, allowing local models (including NVIDIA Nemotron) to run alongside cloud‑hosted frontier models via a privacy router, thereby offering a hybrid local‑plus‑cloud execution environment under strict privacy constraints.

HyperAI also offers ready‑to‑use online notebook environments for developers to experiment with OpenClaw and NemoClaw without complex configuration, linking to the provided URLs.

Overall, the announcements illustrate NVIDIA’s strategy of tightly coupling cutting‑edge AI hardware with a unified software stack to accelerate the next generation of agentic AI, while also acknowledging the practical challenge for users to objectively evaluate and select from a rapidly expanding set of AI accelerators.