Local time on Wednesday (December 24), Groq, regarded as a "challenger," announced on its official website that it had reached a "non-exclusive licensing agreement" with NVIDIA. $NVIDIA (NVDA.US)$ Groq founder and CEO Jonathan Ross, President Sunny Madra, and other core executives and team members will join NVIDIA.

This is not an acquisition of the entire company. NVIDIA paid approximately USD 20 billion in cash to acquire Groq’s core AI inference technology intellectual property and related assets, while Groq's cloud service business (Groq Cloud) will continue to operate independently.

This is considered a typical way for tech giants to compete for top AI talent and technology, enabling them to quickly acquire key innovations while bypassing complex antitrust reviews. For Groq, this may mean the end of its journey as an independent hardware challenger, but its core technology will gain a broader development platform within NVIDIA’s ecosystem.

Groq, a star startup specializing in AI inference chips, was founded in 2016 and is headquartered in California, USA. Founder Jonathan Ross was a core R&D member of Google’s self-developed AI chip TPU (Tensor Processing Unit) project, and some former members of Google’s TPU team also joined Groq following him.

As a core R&D member of Google’s first-generation Tensor Processing Unit (TPU) project, Jonathan Ross was deeply involved in the design of chips specifically optimized for AI. This project was later used in the AlphaGo match that defeated Go champion Lee Sedol and became a key hardware component of Google’s AI services. In 2016, he led seven out of ten core members of Google’s TPU team to resign and founded Groq. At the time, he realized that traditional computing architectures (such as CPUs/GPUs) could not efficiently handle modern AI tasks, prompting him to establish a company that breaks through conventional limitations.

Groq’s core product is the LPU (Language Processing Unit). This type of chip is primarily used to accelerate the speed at which large language models complete inference-related tasks and is considered one of the alternatives to NVIDIA’s GPUs.

In February 2024, Groq launched a new AI chip, claiming to achieve the “strongest inference on Earth”—with inference speeds for large models running on Groq being 10 times or even higher than those of NVIDIA GPUs.

In November 2025, the latest statements from the White House and the U.S. Department of Energy showed that 24 top artificial intelligence companies have signed agreements with the U.S. government to join the “Genesis Program,” with both NVIDIA and Groq among them.

Currently, Groq has partnered with Meta to provide inference acceleration for its Llama API; collaborated with IBM to integrate its AI inference platform; and signed a substantial agreement with Saudi Aramco to plan the construction of a large-scale AI inference data center.

Groq LPU: Remarkable Inference Speed but High Cost

The remarkable inference speed and differentiated technical approach are considered the foundation of Groq LPU’s competitive advantage. Its text generation speed (up to 500 tokens per second) on large models such as Llama and Mixtral has garnered significant attention, being recognized as far surpassing contemporary GPUs.

Moreover, the operating principle of Groq LPU differs from NVIDIA’s GPUs, employing an architecture known as Temporal Instruction Set Computer and utilizing Static Random-Access Memory (SRAM), which is approximately 20 times faster than the High Bandwidth Memory (HBM) used in GPUs.

In terms of chip specifications, the SRAM capacity is 230MB, bandwidth is 80TB/s, and FP16 computational power reaches 188TFLOPs. This difference accounts for the disparity in generation speed between LPUs and GPUs. According to Groq, generating each token requires approximately 10 to 30 joules (J) for NVIDIA GPUs, whereas Groq only needs 1 to 3 joules.

However, Groq LPU is not without flaws, facing challenges in cost and versatility. The enormous clusters required to run large models result in high acquisition and maintenance costs, and the specialized chips struggle to flexibly adapt to the rapidly evolving AI algorithms.

Jia Yangqing, former Vice President of Alibaba Group and Founder & CEO of Lepton AI, noted on social media that since each Groq card has a memory capacity of only 230MB, running the Llama-2 70B model requires between 305 and 572 Groq cards, while only eight H100 cards would suffice.

Jia Yangqing believes that calculating the cost over three years of operation, Groq’s hardware procurement cost amounts to $11.44 million, with operational costs exceeding at least $762,000. Based on current pricing, this implies that, for equivalent throughput, it is nearly 40 times the hardware cost of H100s and 10 times the energy consumption cost.

It is not just the high cost. The SRAM technology occupies a larger area and has relatively higher power consumption. It has long been integrated into SoCs (System-on-Chip) in the form of IP cores rather than being used independently, making it far less promising than HBM for future development. Industry insiders suggest that, overall, whether comparing unit capacity price, performance, or power consumption, the HBM technology used in NVIDIA GPUs outperforms SRAM.

Valuation Surges to $6.9 Billion, Revenue Reaches $90 Million Last Year

Currently, Groq has completed multiple rounds of financing, with its latest valuation reaching approximately USD 6.9 billion.

2017: Seed round raised USD 10.3 million.

2021: Series C round raised USD 300 million, with a valuation exceeding USD 1 billion, officially becoming a unicorn.

August 2024: Completed a USD 640 million Series D round led by BlackRock, with the valuation reaching USD 2.8 billion.

September 2025: Completed a new USD 750 million strategic financing round, with the valuation surging to approximately USD 6.9 billion.

Groq is backed by top-tier multinational financial institutions, leading technology industry giants, and active venture capital funds:

Financial institutions as cornerstones: Global top-tier asset management firms such as BlackRock and Neuberger Berman have participated in large-scale financings multiple times, joined by D1 Capital, Altimeter Capital, and 1789 Capital.

Deep involvement from industrial capital: Investments from industry giants such as Samsung, Cisco, and DTCP (Deutsche Telekom Capital Partners) are not merely financial acts but also strategic partnerships. For instance, these may involve collaborations in chip production, data center deployment, or market channel expansion.

Continuous lead investments from specialized funds: Disruptive (long-term lead investor) and Infinitum. Among them, venture capital funds represented by Disruptive acted as the lead investor in the latest USD 750 million financing round in 2025.

However, compared with revenue of USD 90 million in 2024, a valuation nearing USD 7 billion reflects an extremely high premium.

Its revenue forecast for 2025 has been significantly reduced. In July 2025, Groq cut its 2025 revenue expectation from $2 billion to $500 million. This adjustment may be attributed to delays in the delivery of some large orders, such as agreements with Saudi Arabia, as well as the progress of data center construction.

Groq previously informed investors that its revenue would increase to nearly $1.2 billion (approximately RMB 8.6 billion) in 2026 and exceed $1.9 billion (approximately RMB 13.6 billion) by 2027, primarily driven by direct hardware sales to other companies.

As of mid-2025, Groq holds over $2 billion in cash reserves, ensuring sufficient funding to support its continued expansion.

Looking to pick stocks or analyze them? Want to know the opportunities and risks in your portfolio? For all investment-related questions,just ask Futubull AI!

Editor /rice