On December 25, U.S.-based AI chip powerhouse NVIDIA announced it had reached an agreement to acquire core assets from high-performance AI accelerator startup Groq in a cash transaction valued at approximately $20 billion, according to multiple reports from Business Insider and CNBC.

The transaction marks NVIDIA's largest deal to date, surpassing its 2019 acquisition of Israeli chip designer Mellanox for nearly $7 billion. Rather than a full corporate takeover, the deal is structured as a non-exclusive technology licensing agreement combined with a large-scale asset and talent acquisition.

Under the agreement, NVIDIA will obtain Groq's core technology assets and intellectual property related to AI inference, while Groq's cloud business, GroqCloud, is excluded from the transaction and will continue operating independently. Groq will remain an independent company, with former chief financial officer Simon Edwards appointed as its new chief executive officer.

As part of the deal, Groq founder and former Google TPU architect Jonathan Ross, company president Sunny Madra, and other senior technical leaders will join NVIDIA to help integrate and scale the licensed technology. In an internal email to employees, NVIDIA chief executive Jensen Huang said the agreement would significantly expand NVIDIA's capabilities in AI data centers, which he refers to as "AI factories."

"We plan to integrate Groq's low-latency processors into the NVIDIA AI factory architecture, extending the platform to support a broader range of AI inference and real-time workloads," Huang wrote.

NVIDIA chief financial officer Colette Kress declined to comment on the transaction. Alex Davis, chief executive of Disruptive, which led Groq's most recent funding round, said NVIDIA is acquiring essentially all of Groq's core assets, excluding its cloud operations. Groq confirmed that GroqCloud "will continue to operate without interruption."

Founded in 2016 in Mountain View, California, Groq focuses on high-throughput, low-latency AI inference rather than training. The company was established by former Google engineers from the original Tensor Processing Unit team. In February 2024, Groq introduced its first large-scale inference chip based on its proprietary Language Processing Unit architecture, designed to maximize parallelism and energy efficiency.

Groq's LPU chips, manufactured using a 14-nanometer process, rely on a Tensor Streaming Architecture and integrate 230 megabytes of on-chip SRAM, delivering memory bandwidth of up to 80 terabytes per second. According to the company, its inference performance can exceed traditional GPU- and TPU-based systems by ten to one hundred times on certain workloads, while consuming a fraction of the power.

In September, Groq raised $750 million at a valuation of approximately $6.9 billion, with investors including BlackRock, Neuberger Berman, Samsung, Cisco, Altimeter, and 1789 Capital. At the time, the company was not actively pursuing a sale, Davis said.

The deal reflects a broader trend among major technology companies to secure AI talent and specialized hardware capabilities through licensing and asset acquisitions rather than full takeovers, amid heightened antitrust scrutiny. NVIDIA previously used a similar structure in September, spending over $900 million to license technology and recruit leadership from AI hardware startup Enfabrica.

NVIDIA's aggressive dealmaking has been fueled by its rapidly expanding cash reserves. As of the end of October 2025, the company held $60.6 billion in cash and short-term investments, up sharply from $13.3 billion in early 2023. Beyond Groq, NVIDIA has announced plans to invest up to $100 billion in OpenAI, committed $5 billion to a partnership with Intel, and expanded investments across AI infrastructure, cloud computing, and semiconductor ecosystems.