NVIDIA's Largest Acquisition Ever: Invests 640 Billion to Acquire Groq Technology and the Father of Google TPU

Nvidia has agreed to acquire the core assets and technology licensing of AI inference chip startup Groq for approximately $20 billion in cash, marking Nvidia’s largest single transaction in history. However, both parties emphasize that this is not an acquisition of all of Groq’s equity, but rather a non-exclusive licensing agreement for technology and the integration of high-level talent as the core framework.

Groq subsequently confirmed on its official blog that it has signed a non-exclusive inference technology licensing agreement with Nvidia. Groq founder and CEO Jonathan Ross, who was responsible for Google TPU chip design, along with President Sunny Madra and several core technical executives, will join Nvidia to assist in promoting and expanding related technologies. Groq will continue to operate as an independent company, with former CFO Simon Edwards appointed as the new CEO. Its cloud service GroqCloud is not included in this transaction and will continue to operate as usual.

Nvidia’s largest-ever transaction, signing a technology licensing agreement with Groq

The deal is valued at about $20 billion, and Groq was valued at $6.9 billion during its funding round completed in September this year. The round was led by Disruptive, with investors including BlackRock, Neuberger Berman, Samsung, Cisco, Altimeter, and 1789 Capital, a firm with Donald Trump Jr. as a partner.

With rapid accumulation of cash reserves, Nvidia has continued to expand its investments in the AI ecosystem in recent years. As of the end of October, Nvidia’s cash and short-term investments reached $60.6 billion, far exceeding the $13.3 billion at the beginning of 2023.

In addition to Groq, Nvidia has also invested in AI cloud provider CoreWeave, model company Cohere, and energy and compute infrastructure provider Crusoe in recent years. In September this year, Nvidia announced plans to invest up to $100 billion in OpenAI and $5 billion in Intel, strengthening its overall AI industry chain layout.

Groq confirmed on its official blog that it has signed a non-exclusive inference technology licensing agreement with Nvidia. Groq founder and CEO Jonathan Ross, President Sunny Madra, and several core technical executives will join Nvidia to help promote and expand related technologies. Groq will continue to operate as an independent company, with former CFO Simon Edwards appointed as the new CEO. Its cloud service GroqCloud is not included in this transaction and will continue to operate as usual.

Jensen Huang on the Groq case: integrating low-latency processors to strengthen real-time inference layout

CNBC quoted an internal email from Nvidia CEO Jensen Huang to employees, stating that Nvidia plans to integrate Groq’s low-latency processors into the “NVIDIA AI Factory” architecture to support a broader range of AI inference and real-time workloads.

Huang emphasized in the letter: “We are licensing Groq’s intellectual property and recruiting talented individuals, but we are not acquiring Groq itself.” This transaction structure echoes Nvidia’s approach in September this year, when it invested over $900 million to acquire the Enfabrica AI hardware startup through technology licensing and talent integration.

Analysts point out that using technology licensing and talent integration instead of full acquisition has become a common strategy for tech giants under increasingly strict antitrust regulations. For Nvidia, this move not only quickly fills the inference technology gap but also allows it to preemptively secure a key position in the AI second half without triggering regulatory scrutiny.

The AI battlefield shifts to inference, with ASIC architecture becoming critical

Founded in 2016 by former Google engineers, Groq was established by several ex-Google engineers, including Jonathan Ross, who participated in the design of Google TPU (Tensor Processing Unit). Groq focuses on an LPU (Language Processing Unit) architecture specifically designed for AI inference, emphasizing ultra-low latency, stable response times, and high energy efficiency, making it particularly suitable for real-time dialogue, voice assistants, finance, and industrial scenarios.

As AI applications gradually move from model training to large-scale deployment, industry consensus suggests that future compute demand growth will focus on inference rather than training. Compared to the dominance of GPUs in the training market, the inference domain is now facing competition from Google TPU, dedicated ASICs, and other startup chips.

This article, Nvidia’s largest acquisition: investing 640 billion to acquire Groq technology and the father of Google TPU, first appeared on Chain News ABMedia.