TL;DR
Nvidia paid $20 billion for Groq’s intellectual property and hired its senior leadership, while explicitly not buying the Groq operating company or its cloud business. The structure — a non-exclusive IP license plus talent hire — sidesteps traditional merger reviews and leaves GroqCloud independent amid ongoing Saudi contracts.
What happened
In late 2025 Nvidia agreed to a $20 billion transaction that purchased Groq’s patents, IP and non-exclusive rights to its inference technology, and brought Groq’s CEO Jonathan Ross and the entire senior leadership team into Nvidia. The purchase did not include Groq as a corporate entity and explicitly excluded GroqCloud, which remains an independent business led by CFO Simon Edwards. Groq had recently raised $750 million at a $6.9 billion post‑money valuation; Nvidia’s payment implies roughly a $13.1 billion premium. The deal was structured as IP licensing and talent acquisition rather than a conventional acquisition, a choice that the reporting connects to avoiding multi‑year regulatory processes and to leaving GroqCloud’s Saudi‑backed infrastructure and contracts outside Nvidia’s control. The source also details Groq’s LPU architecture — SRAM‑centric chips optimized for inference with high on‑chip bandwidth — and notes practical limits such as reduced model size support and no training capability.
Why it matters
- Structuring the transaction as IP licensing plus hiring avoids the formal merger processes that would normally trigger CFIUS and antitrust reviews, shortening regulatory delay.
- Owning Groq’s IP and its engineering leadership gives Nvidia immediate access to technology and expertise while leaving a cloud competitor operationally separate.
- The deal blunts potential distribution channels for competing open‑source stacks and reduces the chance of partners scaling Groq’s stack as an alternative inference path.
- Groq’s SRAM‑based inference approach highlights an architectural tradeoff that could matter if inference demand and DRAM/HBM price pressures continue to grow.
- Separating GroqCloud preserves Saudi‑funded infrastructure and contracts with a standalone entity, limiting Nvidia’s geopolitical and regulatory exposure.
Key facts
- Nvidia paid $20 billion to acquire Groq’s intellectual property, patents, and a non‑exclusive license to its inference technology.
- Jonathan Ross (Groq CEO), Sunny Madra (President) and the full senior leadership team joined Nvidia as part of the deal.
- Nvidia’s statement emphasized it did not purchase Groq as a company; GroqCloud was excluded and remains independent under CFO Simon Edwards.
- Groq raised $750 million in September 2025 at a $6.9 billion post‑money valuation; Nvidia’s $20B payment represents a roughly $13.1 billion premium.
- Groq’s LPU architecture uses large on‑chip SRAM (reported as 230 MB per chip with ~80 TB/s bandwidth) to keep models on‑chip for deterministic, low‑latency inference.
- Reported inference throughput examples for LPUs: Llama 2 7B at 750 tokens/sec (2,048 context), Llama 2 70B at 300 tokens/sec (4,096 context), Mixtral 8x7B at 480 tokens/sec (4,096 context).
- LPUs are described as inference‑only: they cannot train models and have limits on maximum model size (example: 14 GB SRAM per rack limitation cited).
- GroqCloud had a $1.5 billion commitment from Saudi Arabia to expand a Dammam data center and was operating inference clusters across multiple regions.
- Groq operated 13 facilities across the US, Canada, Europe and the Middle East, and reportedly expanded capacity by more than 10% in the month before a funding announcement.
What to watch next
- Whether other companies will take up Groq’s non‑exclusive license terms or integrate the IP — not confirmed in the source.
- Any follow‑on regulatory scrutiny or formal reviews prompted by the deal structure — not confirmed in the source.
- The future viability and commercial trajectory of GroqCloud as an independent business after the leadership departure — not confirmed in the source.
Quick glossary
- LPU (Language Processing Unit): A chip architecture focused on inference that emphasizes large on‑chip SRAM and deterministic execution to reduce latency and energy for language models.
- SRAM: Static RAM, an on‑chip memory type that provides low latency and high bandwidth compared with off‑chip DRAM/HBM but is more area‑ and cost‑intensive per byte.
- DRAM / HBM: Off‑chip memory technologies (dynamic RAM and High Bandwidth Memory) commonly used with CPUs, GPUs, and accelerators to store model weights and activations.
- CFIUS: The U.S. Committee on Foreign Investment in the United States, which reviews foreign investments for national security implications.
- Inference: The run‑time execution of a trained machine learning model to generate outputs (such as text), distinct from training, which builds model parameters.
Reader FAQ
Did Nvidia acquire Groq as a company?
No. Nvidia bought Groq’s IP and licensed its inference technology and hired the leadership team, but explicitly did not purchase Groq as a corporate entity.
Is GroqCloud part of the Nvidia deal?
No. GroqCloud was excluded from the transaction and remains an independent business under CFO Simon Edwards.
Was the license exclusive to Nvidia?
Nvidia described the license as non‑exclusive; whether that will have the practical effect of exclusivity due to IP ownership and talent hire is not confirmed in the source.
Can LPUs train large models?
No. The source describes LPUs as inference‑only architectures and states they cannot perform model training.
"You're taking on a giant. What gives you the audacity?" On November 5th, 2025, Groq CEO Jonathan Ross was asked why he was even bothering to challenge Nvidia. He didn't…
Sources
- Nvidia's $20B Antitrust Loophole (Not an Acquisition)
- Nvidia: What Should Investors Make Of The Groq Deal …
- Groq Signs Non-Exclusive Licensing Deal with Nvidia
- Nvidia Stock: Has A Cash Problem (NASDAQ:NVDA)
Related posts
- OrangePi 6 Plus review — 12-core ARM64 SBC with high-end IO and NPU
- Nvidia Pays $20B for Groq After Firm Cut Revenue Forecast by 75%
- Nvidia Pays $20B to Acquire Groq After Firm Cuts Revenue Forecast 75%