Why did Nvidia spend $20bn on Groq without buying the company?

Nvidia has reached a reported $20 billion agreement to license technology and hire key personnel from AI chip specialist Groq, in what is being described as the largest strategic transaction in its history.

The deal was framed as licensing and not a full acquisition. Nvidia will pay to secure a perpetual, non-exclusive license to Groq’s core intellectual property and bring much of Groq’s leadership and engineering team into Nvidia’s organisation.

Reports from multiple outlets indicate that Groq’s founder and CEO Jonathan Ross, who previously worked on Google’s Tensor Processing Unit (TPU), and president Sunny Madra are among those joining Nvidia. Groq will retain its cloud service, GroqCloud, under new leadership.

Groq’s technology focuses on specialised inference processors called Language Processing Units (LPUs), custom chips designed for high-speed, low-latency execution of trained AI models.

Unlike traditional GPUs, graphics processors that excel at training and broad-based compute tasks, LPUs use on-chip SRAM memory, which can deliver faster response times and higher energy efficiency for certain real-time AI tasks.

While GPUs remain central to model development, the rapid commercial deployment of AI systems is shifting spending toward low-latency, energy-efficient inference. By licensing Groq’s technology and absorbing key talent rather than pursuing a full acquisition, Nvidia is reinforcing its grip on the fastest-growing segment of the AI compute market while blunting a credible competitive threat.

Analysts note that in benchmarks, Groq’s designs have shown significant gains in “tokens per second” throughput compared with conventional GPU-based solutions, which makes them attractive for real-time applications such as live translation, autonomous systems, or interactive AI services.

AI inference is projected to account for the majority of AI compute spending in the coming years, as models are deployed across commercial and consumer applications. Securing advanced inference capability positions Nvidia to capture a larger share of this expenditure. At the same time, it neutralises a credible competitor: Groq’s architecture was widely seen as one of the few that could challenge Nvidia’s chips in specific use cases.

The transaction comes at a point of intense competition in the AI hardware sector. Nvidia’s GPUs have been widely adopted by cloud providers, enterprises, and AI developers for both training and inference. However, custom silicon from companies such as Google (with its TPUs), AMD, Intel, and specialised startups like Cerebras has steadily gained attention as organisations seek greater performance and energy efficiency.

Groq was co-founded in 2016 by Ross, who previously helped lead Google’s Tensor Processing Unit development. The company provides dedicated chips and accelerators within its own servers and racks designed for artificial intelligence, machine learning, and high-performance computing.

In August 2024, Groq closed a $640m funding round that was led by BlackRock and saw participation from Neuberger Berman, Type One Ventures, Cisco, KDDI, and Samsung Catalyst Fund.

In February 2025, Groq secured $1.5bn from Saudi Arabia to expand AI inference infrastructure in the region.

Groq equity ownership participants include the founders, employees, and several venture capital firms.

The coming year will test whether Nvidia successfully weaves Groq’s technology into its next generation of AI hardware offerings and whether regulators push back against the practice of structuring major strategic bets as licensing arrangements.