Groq tech folded into Nvidia's inference push
Nvidia's new Vera Rubin / inference platform incorporates Groq's LPX/LPU technology and is being positioned as a high-efficiency inference solution, with coverage noting Groq integration into Nvidia's stack reported reported. Social posts also name Samsung as a manufacturing partner for Groq LPUs destined for H2 2026 shipments reported.
NVIDIA said) Vera Rubin ships with seven new chips now in production, including an NVL72 GPU rack and a dedicated Groq 3 LPX inference-accelerator rack listed in the company’s March 16, 2026 press materials. The integration follows a late‑2025 strategic deal in which NVIDIA paid) roughly $20 billion for a licensing-and-acqui‑hire package around Groq’s assets, a transaction that included hiring Groq founder Jonathan Ross and other senior engineers into NVIDIA roles. NVIDIA CEO Jensen Huang confirmed) onstage that Samsung Electronics is manufacturing the Groq 3 LPU and said first shipments are expected in the third quarter of 2026 (Q3), aligning the LPU ramp with Vera Rubin deployments. Samsung’s foundry relationship with Groq dates to a 2023 contract to use Samsung’s SF4X 4nm process, and industry reporting now forecasts Samsung wafer volumes for Groq to rise from roughly 9,000 wafers in the prior cycle to about 15,000 wafers as mass production begins. Groq’s own recent financial moves remain material: the company cut) its 2025 revenue projection by about $1.5 billion and published outlooks in late 2025 targeting roughly $1.2 billion in 2026 and $1.9 billion in 2027, figures market coverage tied to the Samsung partnership. Technical differentiation behind the tie‑up centers on the LPU’s memory approach: Groq’s LPU designs emphasize on‑die SRAM and extreme internal bandwidth (tens of terabytes/sec), a tradeoff noted in Groq’s engineering posts and independent coverage that positions LPUs as latency‑focused complements to GPU/HBM architectures.
