NVIDIA buys Groq
NVIDIA agreed to acquire Groq for $20 billion and said Groq 3 LPU will ship in Q3 — a major M&A move announced around GTC. (x.com)
NVIDIA structured the arrangement as a non‑exclusive licensing and asset-and‑talent transfer rather than a straight equity buyout, according to the companies’ statements. (groq.com) Groq founder Jonathan Ross and president Sunny Madra — along with the bulk of Groq’s engineering team — have joined NVIDIA to help scale the licensed inference technology. (groq.com) Groq will continue to exist as an independent company under new leadership, with Simon Edwards named CEO and GroqCloud expected to keep operating. (groq.com) The first chip from the arrangement, Groq 3, departs from HBM‑centric designs by packing roughly 500 MB of on‑die SRAM per LPU and delivering about 150 TB/s of local bandwidth, enabling high‑bandwidth decode operations. (tomshardware.com) NVIDIA plans to scale Groq 3 into LPX racks that pair hundreds of LPUs — Tom’s Hardware describes a 256‑LPU rack delivering about 128 GB of SRAM and roughly 40 PB/s of aggregate bandwidth for inference acceleration. (tomshardware.com) Samsung Foundry was announced as the production partner for the Groq 3 design and confirmed mass production on a 4‑nanometer process during NVIDIA’s GTC presentations. (dataconomy.com) U.S. Senators Elizabeth Warren and Richard Blumenthal sent NVIDIA a letter on March 23, 2026 asking whether the deal was structured to evade antitrust review, and NVIDIA has publicly described the transaction as a licensing arrangement rather than an acquisition. (warren.senate.gov)
