NVIDIA Achieves 35x AI Inferencing Breakthrough Ahead of Schedule as Foxconn Accelerates Groq 3 LPX Racks for Trillion-Parameter Models

Foxconn has been designated as the primary supplier for NVIDIA’s newly unveiled Groq 3 LPX chip, an advanced processor engineered specifically to enhance AI inferencing capabilities significantly.

During the GTC 2026 event, NVIDIA introduced the Groq 3 LPX chip alongside the Vera Rubin AI platform. This new chip, classified as a Logical Processing Unit (LPU), aims to amplify AI inferencing performance by an impressive 35x. Each Groq 3 LPX rack will comprise 256 chips equipped with 128 GB of SRAM and 12 TB of DDR5 memory, specifically tailored for multi-trillion parameter AI models.

Side-by-side view of the Rubin GPU and Groq 3 LPU providing performance specifications such as 288 GB HBM4 and 500 MB SRAM, with performance metrics detailing 50 PFLOPs (NVFP4) for Rubin GPU and 1.2 PFLOPs (FP8) for Groq 3 LPU. — Uniting Processors of Extreme Performances.

The NVIDIA Groq 3 LPX server imagery showcasing specs such as 'AI Inference Compute' at 315 PFLOPS, 'SRAM Capacity' of 128 GB, and 'Memory Bandwidth' of 40 PB/s. — Specifications for AI Inference Compute.

Recent reports from Taiwanese media indicate that shipments of the Groq 3 LPX chip are ahead of schedule, expected to commence in the third quarter of 2026. This rapid timeline underscores NVIDIA’s commitment to accelerating the development of its Vera Rubin platform to meet the surging demand for sophisticated AI inferencing solutions—a critical component for the burgeoning Agentic AI era.

NVIDIA is making significant strides in the AI inference sector with its Groq 3 LPX cabinet, which is reportedly scheduled for expedited shipping in Q3. Foxconn serves as the exclusive provider for the computing tray and plays a leading role in assembling the LPX cabinets, competing directly in the AI inference market.

UDN

Initial projections for Groq 3 LPX shipments were quite conservative, but updated supply chain analyses suggest that an estimated 1.5 million units of LP30 and LP35 chips—integral to the LPX racks—are anticipated to ship this year, with numbers potentially climbing to 2.5 million in 2027.

According to these analyses, Foxconn stands out as the “exclusive”supplier for the Groq 3 LPX computing tray, along with taking the lead in LPX cabinet assembly. The exceptional demand for NVIDIA’s Vera Rubin platforms and Groq 3 LPX racks is projected to elevate Foxconn’s market share from 55% to 60% in the latter half of this year, resulting in significant revenue growth.

A slide presentation titled 'NVIDIA Extreme Co-Design Delivering X-Factors Every Year' illustrating various hardware architectures including Blackwell, Rubin, and Feynman, with associated NVLink models and CPUs for the years 2024, 2026, and 2028. — NVIDIA’s Roadmap Across Key Hardware Architectures.

Foxconn plans to deliver approximately 6, 000 Groq 3 LPX racks in the current year, with an additional 10, 000 units expected in 2027. Notably, this figure does not account for the anticipated next-generation LPX racks based on LP40 chips, set to begin shipping next year.

The core Vera Rubin “NVL72″racks are forecasted to reach a production milestone of 12, 000 units by 2026, with major clientele including tech giants Google, Amazon AWS, and Microsoft. Mass production of the Vera Rubin VR200 NVL72 servers is expected to kick off by late Q3 2026.

In response to significant market demand, Foxconn is also ramping up its production capabilities for NVIDIA’s latest AI server offerings. CEO Liu Yangwei has stated that the company has the capacity to manufacture over 1, 000 cabinets weekly, with intentions to boost this capacity to 2, 000 cabinets by the end of 2026.

Source & Images