Nvidia launches Groq 3 LPX at GTC 2026, adding inference hardware to its platform

Nvidia’s Vera Rubin platform, first announced at CES 2026, was dramatically expanded at GTC 2026 with the addition of the Groq 3 LPX rack, marking the company’s inaugural foray into dedicated inference hardware. According to The‑Decoder, the Groq 3 LPX sits alongside the NVL72 “central compute” rack and introduces a low‑latency inference pipeline that complements the platform’s existing training‑focused components. The new rack is part of a broader seven‑chip, five‑rack ecosystem that Nvidia now calls the Vera Rubin POD, which aggregates 40 racks, 1,152 Rubin GPUs, and 60 exaflops of compute power, delivering up to ten‑fold inference performance per watt over the previous Blackwell generation.

The NVL72 rack, described by The‑Decoder as the POD’s “core compute engine,” integrates 72 Rubin GPUs, 36 Vera CPUs, ConnectX‑9 SuperNICs, and BlueField‑4 DPUs across 18 compute trays and nine NVLink switch trays. Nvidia claims the single 19‑inch‑wide rack houses roughly 1.3 million individual components and weighs about 4,000 pounds, yet it is assembled without traditional cabling, hoses, or fans. A PCB midplane replaces conventional wiring, cutting assembly time per tray from nearly two hours to five minutes, a claim also echoed in Nvidia’s own presentation at GTC. The sixth‑generation NVLink delivers 3.6 TB/s per GPU and 260 TB/s across the full rack, while a modular copper cable backbone spans more than two miles, supporting the massive bandwidth requirements of both training and inference workloads.

Beyond the NVL72, Nvidia unveiled two scaling tiers that extend the POD’s capabilities. The Vera Rubin Ultra NVL576 tier links eight NVL72 racks into a 576‑GPU NVLink domain using a new two‑layer all‑to‑all topology, while the Kyber rack doubles the per‑rack NVLink domain to 144 GPUs by stacking compute hardware vertically. These designs build on a prototype called Polyphe, which was based on the older GB200 architecture, according to The‑Decoder. The scaling architecture is intended to accommodate the growing demand for both massive model training and ultra‑low‑latency inference, a need underscored by Nvidia’s simultaneous launch of the Nemotron Coalition—a partnership with Mistral AI, Perplexity, and Cursor aimed at developing open‑frontier models tightly coupled to Nvidia’s infrastructure.

Security and software integration were also highlighted at GTC. Nvidia introduced NemoClaw, a security stack for AI agents that CEO Jensen Huang likened to the importance of Linux and Kubernetes in the broader computing ecosystem. The company also announced an inference operating system to manage the new Groq 3 LPX hardware, though detailed specifications were not disclosed. In parallel, Nvidia’s DLSS 5, which uses AI to generate photorealistic lighting, is slated for the RTX 50 series later this year, a move that has already sparked debate within the gaming community over artistic fidelity versus AI‑driven enhancement.

Analysts have noted that Nvidia’s end‑to‑end strategy—spanning custom CPUs, GPUs, dedicated inference ASICs, high‑speed interconnects, and a unified software stack—mirrors its broader ambition to dominate the AI data‑center market. A ZDNet feature points out that Huang argues “AI data centers will be more efficient, more economical, and generate more revenue if you buy all the parts from his company.” By integrating inference hardware directly into the Vera Rubin POD, Nvidia not only differentiates its offering from competitors that rely on third‑party accelerators but also creates a tighter coupling between model development and deployment, a point emphasized in VentureBeat’s coverage of the Vera Rubin platform’s open‑model alliances.