A new era for artificial intelligence: Nvidia launches Groq 3 LPX servers

Nvidia has officially confirmed that its dedicated AI inference accelerator, called Groq 3 LPX, has reached full mass production. The technology was created through the acquisition of Groq's staff and licenses worth almost €18.5 billion. The new Nvidia Groq 3 LPX accelerators are designed to plug directly into the existing Nvidia Vera Rubin platform. A single server rack can accommodate up to 256 LPX accelerators, enabling extreme computing power density.

The first cloud customer to use these new server systems in its infrastructure is Nebius. As part of its so-called “chip factory,” the LPX cabinets will be installed in parallel with Nvidia Vera processors and Nvidia Rubin graphics accelerators. This combination of equipment is expected to provide exceptional efficiency when processing the most demanding artificial intelligence models. Despite the high performance promises, it remains to be seen in practice how the entire system will perform under long-term loads and whether the integration between different chip types will be carried out without bottlenecks.

Measurements by the independent Artificial Analysis platform have already shown impressive results. When testing the Google Gemma 4 31B model with a context volume of 100K, the equipment achieved a speed of 3400 output tokens per second. Such figures represent a significant leap in the response speed of AI systems.

The space industry is also showing interest in Nvidia's new hardware. SpaceX has revealed that its next advanced artificial intelligence system will run on Nvidia Vera processors. This plan includes operating within future orbital data centers, which represents a completely new and rather risky technological challenge.