What does NVIDIA's Vera Rubin with Groq 3 LPX change for AI factories?

The rack that prints tokens

A server rack usually sits in a row, humming, moving data between GPUs. The Vera Rubin NVL72 rack does that, but it also adds a second kind of silicon dedicated to the decode step of large language models. NVIDIA calls this combination a token factory.

The idea is simple: reasoning models produce one token at a time, and every microsecond of latency multiplies across the chain. Groq 3 LPX handles that decode step while Rubin GPUs chew through context. They are codesigned to share every layer of the model. In an Artificial Analysis benchmark running Gemma 4 31B, the combined system delivered 3,400 output tokens per second for 100,000-token long-context use cases. That is four times faster than the nearest alternative platform. The benchmark uses an open source agentic model. The number comes from NVIDIA's own announcement at Hot Chips in Palo Alto.

LPX sits beside the GPU

Groq 3 LPX is not a replacement for the GPU. It is an extension. A rack-scale deployment can include 256 LP30 accelerators connected through direct chip-to-chip links. Rubin GPUs and LPUs jointly compute every layer. At scale, fleets of LPUs operate as a giant processor optimized for deterministic inference. The architecture eliminates the traditional tradeoff between speed and throughput. NVIDIA says more optimizations, more models, and more performance are coming when paired with Vera Rubin NVL72.

Nebius is the first AI cloud to adopt Groq 3 LPX. Adding it to Vera Rubin NVL72 in Nebius Token Factory will boost inference performance so developers can build highly interactive agents, coding systems, and other real-time AI experiences at scale.

SpaceXAI plans to build its future AI architecture around NVIDIA Vera Rubin, from data centers on Earth to orbital satellites. The company will deploy Vera CPUs to accelerate CPU-intensive work behind agentic AI, including orchestration, tool use, code execution, data processing, and simulation.

Networking without a third tier

As AI factories grow, the network becomes a bottleneck. The traditional fix adds a third network tier. That adds latency, unpredictable slowdowns, and cost for cabling, optics, and power.

Spectrum-X Multiplane takes a different approach. It splits each server's network connection into several independent paths, or planes, each running its own lightweight two-tier network. The result is a flat network that scales to 512,000 GPUs without the added tier. A dedicated hardware engine inside the ConnectX SuperNIC manages traffic across planes and instantly reroutes around any failure. Applications see one fast, reliable connection. In an eight-plane topology, if one plane fails, the network still maintains about 90 percent of its total bandwidth. Hardware recovery is 11 times faster than software-based multiplane load balancing. This translates to 1.6 times higher AI factory output.

The SN6000 series switches use the 102.4 terabit per second Spe

Spectrum-XGS Ethernet extends the same codesign across data centers, letting multiple facilities function as a single AI super-factory and accelerating multi-site NCCL collectives by 1.9 times. CoreWeave is deploying Spectrum-X Multiplane in production, unlocking advances for its AI cloud infrastructure.

Scale-In joins the stack

NVIDIA Scale-In is introduced as the fifth pillar of NVIDIA AI networking. It extends purpose-built acceleration to infrastructure services that secure, manage, and operate the AI factory. Powered by the BlueField-4 processor and DOCA software platform, connected over Spectrum-X Ethernet, Scale-In transforms the traditional north-south access network into a unified, accelerated infrastructure domain. Cloud computing brought software-defined networking, composability, and elasticity. Agentic AI represents the next platform shift. AI factories bring massive accelerated compute together with growing numbers of users, applications, and autonomous agents continuously interacting with data, storage, and services. Networking, storage, cybersecurity, and operations must now be accelerated alongside AI compute. Scale-In delivers multi-tenant networking, high-performance storage access, in-silicon security, elastic provisioning, and real-time observability while keeping infrastructure processing independent of host compute resources.

Custom silicon gets a seat at the table

NVLink Fusion connects custom XPUs and CPUs to NVIDIA's scale-up and scale-out technology stack. The platform includes sixth-generation NVLink and NVLink Switch for scale-up networking, plus NVLink-C2C for energy-efficient connectivity between XPUs and CPUs. Through the MGX ecosystem, adopters can use production-proven rack designs, components, manufacturing partner solutions, and open software for distributed computing, disaggregated workloads, and cluster management. By standardizing GPU- and XPU-based systems on a unified architecture, NVLink Fusion helps decouple data center buildout from silicon readiness. Operators can share rack footprints, networking, cooling, power delivery, and management systems, then adjust the mix of GPUs and XPUs as supply and workload requirements evolve.

The announcement does not name specific partners adopting NVLink Fusion today. It also does not disclose pricing, power draw per rack, or availability dates beyond "full production" for Groq 3 LPX.