Nvidia strategic acquisition of Groq, $20 billion acquisition of 'competitor'

On December 24th local time, it was rumored that NVIDIA would acquire the core assets of artificial intelligence chip startup Groq for approximately $20 billion in cash. This not only sets a record for Nvidia's largest acquisition in history (far surpassing the $7 billion acquisition of Mellanox in 2019), but also means that this GPU giant has completed a profound strategic layout in the AI inference arena.

This news was revealed by Alex Davis, CEO of Disruptive, a US investment firm that led Groq's $750 million financing. It is reported that Groq has not been completely absorbed, and its cloud business GroqCloud still operates independently; Founder Jonathan Ross and other core executives will join Nvidia, but the company itself will continue to be led by new CEO Simon Edwards; Nvidia has acquired all hardware assets and intellectual property rights.

Nvidia CEO Huang Renxun also explicitly stated in an internal email: 'We did not acquire Groq, but instead obtained authorization for its low latency processor technology and recruited its top talent.'. ”But there is no doubt that Nvidia's $20 billion purchase is not only a set of "chip designs", but also a key puzzle leading to the next generation of AI infrastructure.

Groq: From TPU Father to LPU Revolutionary

The rise of Groq stems from the foresight of a 'genius engineer'. The founder of the company, Jonathan Ross, was once the core creator of Google's TPU (Tensor Processing Unit) project. In the early days of working on Google's advertising system, he keenly sensed the huge demand for computing power in AI tasks such as speech recognition and search ranking.

With the help of Google's famous "20% free time" policy, Ross secretly launched an experimental project - which eventually evolved into a TPU that changed the industry landscape and once supported 50% of Google's global computing power demand.

In 2016, Ross left Alphabet's X Lab (formerly known as Google X) and founded Groq, with the goal of directly targeting Nvidia's dominant position in the field of AI inference. He brought a disruptive architectural concept: Language Processing Unit (LPU). Unlike traditional GPUs that rely on external high bandwidth memory (HBM), Groq's LPU integrates up to 230MB of on-chip SRAM on a single chip, providing an astonishing 80TB/s memory bandwidth - far exceeding GPU solutions.

Image source: Groq

More importantly, Groq abandons complex mechanisms such as out of order execution and branch prediction commonly found in CPU/GPU, and adopts a "deterministic execution" microarchitecture. Its core is the Tensor Streaming Processor (TSP) working in conjunction with a powerful compiler: the compiler precisely schedules each computation and data stream, achieving cycle level precision control. This "software defined hardware" model enables the system to maintain extremely low latency and high throughput in distributed deployment, making it particularly suitable for large-scale model inference scenarios.

Actual test data shows that when running open-source models such as Llama 2 and Mixtral, Groq can achieve an inference speed of up to 500 tokens per second, a response speed several times faster than mainstream GPUs, and a unit token energy consumption of only one-third of GPUs. Groq's goal is not to replace training chips, but to capture a key trend: AI inference demand will soon exceed training demand. In this new era of 'reasoning is king', Groq attempts to become an efficient engine at the infrastructure level.

NVIDIA's Strategic Anxiety: Defensive Expansion and Ecological Closed Loop

Despite Nvidia's absolute dominance in the AI training market (with a market share of over 80%), the inference market is showing a diverse trend. Startup companies such as Cerebras, SambaNova, and Mythic have launched dedicated inference chips one after another; Google, Amazon, and Microsoft are also vigorously developing alternative solutions such as TPU, Trainium, and Maia.

What is even more alarming for Nvidia is that Groq's LPU has demonstrated performance advantages over A100/H100 in specific scenarios, especially in low latency, high certainty tasks.

However, in addition to Groq, Cerebras is intensifying its competition with Nvidia, striving to build processors for running generative AI models. The company submitted its IPO application at the end of 2024, originally planned to go public this year, but withdrew its IPO application after announcing the completion of a round of financing and raising over $1 billion in October. A spokesperson for the company stated that Cerebras still hopes to go public as soon as possible.

For Nvidia, instead of waiting for Groq and others to grow and become true competitors, it is better to incorporate its core technology into its portfolio in advance.

It is worth mentioning that this transaction is not a traditional merger and acquisition, mainly acquiring all IP and hardware assets, including LPU architecture, TSP microarchitecture, compiler technology, and absorbing core teams, especially architecture masters like Ross. However, this transaction excludes GroqCloud business to avoid direct competition with its own DGX Cloud and partners such as CoreWeave.

NVIDIA's strategy not only strengthens its own technology stack, but also avoids integration risks. Huang Renxun emphasized in the email that Groq's low latency processors will be integrated into the "NVIDIA AI Factory" architecture to serve a wider range of real-time AI workloads - meaning that LPU technology may be integrated into the next generation Blackwell or Rubin platforms in the form of IP modules and become part of the CUDA ecosystem.

It is reported that Groq had no intention of selling before being approached, indicating that this transaction was more of a defensive acquisition initiated by Nvidia.

Building deeper ecological moats from Enfabrica to Groq

In fact, apart from Nvidia, in the past few years, tech giants such as Meta, Google, and Microsoft have invested heavily in hiring top artificial intelligence talent through various types of licensing agreements.

And this "talent+IP" model is becoming Nvidia's standard practice for addressing innovation challenges. In September of this year, Nvidia conducted a similar but smaller transaction, in which the company spent over $900 million to hire Rochan Sankar, CEO of Enfabrica, and other employees of the AI hardware startup, and obtained technology licenses from the company.

Currently, Nvidia has increased its investment in chip startups and a broader ecosystem. The company has invested in artificial intelligence and energy infrastructure company Crusoe, artificial intelligence model developer Cohere, and increased its investment in CoreWeave.

In addition, Nvidia previously announced plans to invest up to $100 billion in OpenAI, which has committed to deploying at least 10 gigawatts of Nvidia products. But the two companies have not officially announced an agreement yet. Meanwhile, Nvidia's plan to invest $5 billion in Intel has been approved by the Federal Trade Commission (FTC) in the United States.

The reason why Nvidia is "throwing a lot of money" is due to its strong cash reserves. As of October 2025, its cash and short-term investments amounted to $60.6 billion, significantly higher than the $13.3 billion at the beginning of 2023. Huang Renxun, who holds a large sum of money, obviously does not want to sit idly by and watch all potential threats grow, and instead continues to build a stronger ecological moat.