Groq, an AI startup known for its ultra-fast reasoning chip LPU, earlierAnnounceThe company has signed a non-exclusive licensing agreement with NVIDIA regarding inference technology.
Even more surprisingly, Jonathan Ross, the key figure behind Groq and a former core developer of Google TPU, will lead a portion of the core team directly to NVIDIA. While this isn't a traditional corporate acquisition, the substantial transfer of "talent and technology" will undoubtedly have a significant impact on the AI inference market.
Founder "jumps ship" to partner company? Groq emphasizes company independence
According to the announcement, the agreement is non-exclusive, meaning Groq will continue to operate independently as a company. However, the personnel changes are quite dramatic: key members, including CEO Jonathan Ross and COO and President Sunny Madra, will move to NVIDIA to lead the development and adoption of licensed technologies. Former CFO Simon Edwards will become the new CEO of Groq, continuing to lead the company's operations, and its GroqCloud cloud service will remain unaffected.
About Groq: An LPU designed specifically for LLM inference.
Groq was founded in 2016 and is headquartered in Mountain View, California. Founder Jonathan Ross was one of the founding members of Google's TPU (Tensor Processing Unit) team.
Unlike NVIDIA GPUs, which use parallel computing to process graphics and AI, Groq developed a dedicated processor called the LPU (Language Processing Unit), which employs a deterministic architecture and is specifically optimized for the "inference" stage of large language models (LLMs). Its token generation speed often amazes the market. Jonathan Ross once proudly stated, "Others rely solely on GPUs, while our advantage lies in customized chips."
Analysis: NVIDIA's "Counter-Defense" and "Technology Acquisition"
In my opinion, although this deal is nominally a "license", it is essentially more like a precise "acqui-hire".
While NVIDIA currently holds absolute dominance in AI training, Groq's LPU architecture does demonstrate higher efficiency and lower latency than GPUs in inference. Through this collaboration, NVIDIA not only acquired a license for Groq's key inference technology but also brought Jonathan Ross, a leading expert in this field, into its fold.
There are two advantages to doing this:
• Avoiding antitrust scrutiny:If NVIDIA were to acquire Groq outright, it would inevitably face strong opposition from regulatory agencies in various countries. By using a combination of licensing and poaching, NVIDIA has secured both the technology and personnel while preserving Groq as an independent shell company, which is a very clever business strategy.
• Reinforce the inference map:As AI applications become more widespread, the demand for inference computing power will exceed that for training. NVIDIA's move will allow it to quickly strengthen its technological moat in inference, preventing Groq or other ASIC vendors from gaining a dominant position.
As for what will happen to Groq after losing its founder, although the company emphasizes that operations will continue as usual, it remains a big question mark whether this once-star unicorn can maintain its innovative momentum after the departure of its key figure.
