--- title: "Report: Groq's next LPU stacks DRAM, dropping its SRAM-only design" description: "ZDNet Korea, citing industry insiders, says Groq's next inference chip will stack DRAM in 3D to chase memory capacity. That reframes what Nvidia licensed for $20 billion, while a Delaware suit fights over that price." publisher: "The Inference" section: "Silicon" published: 2026-10-21T13:14:49.201Z modified: 2026-10-21T13:14:49.201Z canonical: https://theinference.org/article/report-groq-s-next-lpu-stacks-dram-dropping-its-sram-only-design language: en keywords: "Groq, Nvidia, Chips, Inference, Memory, Litigation" --- # Report: Groq's next LPU stacks DRAM, dropping its SRAM-only design > ZDNet Korea, citing industry insiders, says Groq's next inference chip will stack DRAM in 3D to chase memory capacity. That reframes what Nvidia licensed for $20 billion, while a Delaware suit fights over that price. ## The SRAM-only bet hits a capacity wall [Groq](https://theinference.org/markets/companies/groq) built the Language Processing Unit, its AI inference chip, on one rule: SRAM only, no DRAM, no HBM. [On October 20, ZDNet Korea quoted a semiconductor insider on the next generation](https://x.com/jukan05/status/2112665161194672181): "Groq plans to introduce 3D packaging to its next-generation chips [1]. It involves stacking DRAM rather than existing SRAM." [1] The reported package, called 3DS, stacks DRAM layers on through-silicon vias and keeps standard DRAM interfaces, chasing capacity where HBM chases speed. [2] Bandwidth was never the problem. Each Groq 3 LPU holds 500 MB of on-chip SRAM moving 150 TB/s, about seven times [the 22 TB/s a single Rubin GPU draws from its HBM4](https://www.nvidia.com/en-us/data-center/hgx/), plus 1.20 PFLOPs of FP8 compute. [1], [3], [4] Capacity is: a full 256-chip LPX rack totals 128 GB of SRAM, less than one Rubin GPU's 288 GB, so a trillion-parameter model's weights must be spread across racks just to be held. [1], [3], [4] *Chart: **A 256-chip Groq rack holds less memory than one Rubin GPU** Bar chart comparing memory capacity: one Groq 3 LPU chip holds 0.5 GB of SRAM, a full 256-chip Groq LPX rack holds 128 GB, and a single Nvidia Rubin GPU holds 288 GB of HBM4.* *Memory carried by one unit of each: on-chip SRAM for the Groq parts, on-package HBM4 for the Rubin GPU. Sources: Wccftech, Crypto Briefing, Nvidia HGX specifications. [1][3][4]* ## A $20 billion license, not an acquisition Those racks entered full production in August inside [Nvidia](https://theinference.org/markets/companies/nvidia)'s Vera Rubin platform and deliver roughly 3,400 tokens per second beside NVL72 servers. [1], [6] Several writeups of the DRAM report call Groq "Nvidia-acquired." [1], [2], [8] The record says otherwise: [December 2025 brought a non-exclusive IP license plus hires worth about $20 billion](https://finance.yahoo.com/technology/ai/articles/ex-groq-engineers-sue-board-171526629.html), and Groq kept operating independently, GroqCloud included. [5], [6] ## A new exhibit for the Delaware fight Two former engineers, Joshua Rubin and Benjamin Serebrin, sued Groq's board on October 2 in Delaware's Court of Chancery, alleging it skipped a stockholder vote Delaware law requires and ran no process to maximize value [6]. Their complaint counts $17 billion for the license and $3 billion in restricted stock for the 150 to 200 engineers who moved [6]. Groq calls the suit meritless and says it remains standalone. [6] Neither Groq nor Nvidia has confirmed the DRAM plan, and [one report hints the SRAM and DRAM lines may survive as separate products](https://cryptobriefing.com/groq-stacked-dram-next-inference-chip/). [1], [3] If the next chip replaces the architecture that license covers, both sides in Delaware get a fresh exhibit on what was actually sold and what it was worth. [Cerebras, the other SRAM maximalist](https://theinference.org/article/cerebras-cuts-its-own-moat-in-half-and-calls-it-a-5x-upgrade), is [reported to be making the same turn to vertical DRAM for its CS-6](https://www.kucoin.com/news/flash/nvidia-s-next-gen-groq-lpu-chip-to-shift-to-3d-stacked-dram). [7] Watch three things: confirmation, the next LPU's per-chip capacity, and the rack configuration. [3] The tell is gigabytes per chip: clear a trillion-parameter model on a single rack and the DRAM detour worked. ## Takeaway Groq is reportedly abandoning the SRAM-only design behind its LPU and stacking DRAM in 3D to chase capacity, a pivot that reframes the $20 billion Nvidia license and hands both sides in the Delaware lawsuit a fresh exhibit. ## Sources 1. [Wccftech, "Groq Aims To Roll Out 3D Stacked DRAM Memory Technology In Next-Gen Chips, Pushing Capacity Over Speed", 21 October 2026](https://wccftech.com/nvidia-groq-moving-next-gen-lpus-from-sram-to-3d-stacked-dram/) 2. [ChainCatcher, "AI chip company Groq plans to introduce 3D stacked DRAM in the next generation of chips", 21 October 2026](https://www.chaincatcher.com/en/article/2297745) 3. [Crypto Briefing, "Groq plans stacked DRAM for its next inference chip to fix a memory problem", 21 October 2026](https://cryptobriefing.com/groq-stacked-dram-next-inference-chip/) 4. [NVIDIA, HGX platform specifications (Rubin GPU: 288 GB HBM4, 22 TB/s), accessed 21 October 2026](https://www.nvidia.com/en-us/data-center/hgx/) 5. [Softonic, "Groq's next LPU will use 3D stacked DRAM: bigger AI models ahead", 21 October 2026](https://en.softonic.com/articles/groqs-next-lpu-will-use-3d-stacked-dram-bigger-ai-models-ahead) 6. [Quartz via Yahoo Finance, "Ex-Groq engineers sue board over $20B Nvidia asset deal", 5 October 2026](https://finance.yahoo.com/technology/ai/articles/ex-groq-engineers-sue-board-171526629.html) 7. [KuCoin News, "NVIDIA's Next-Generation Groq LPU Chip to Transition to 3D Stacked DRAM", 21 October 2026](https://www.kucoin.com/news/flash/nvidia-s-next-gen-groq-lpu-chip-to-shift-to-3d-stacked-dram) 8. [@jukan05 on X, "Groq, Acquired by Nvidia, Seeks to Introduce 3D-Stacked DRAM in Next-Generation Chips", 20 October 2026](https://x.com/jukan05/status/2112665161194672181) --- Report: Groq's next LPU stacks DRAM, dropping its SRAM-only design — The Inference. Canonical: https://theinference.org/article/report-groq-s-next-lpu-stacks-dram-dropping-its-sram-only-design