NVIDIA recently spent about $2 billion to obtain non-exclusive licensing rights to technology from the AI chip startup Groq. Although Groq remains independently operated, key team members such as its founder Jonathan Ross have joined NVIDIA. The deal has been viewed by the outside world as a "de facto acquisition" aimed at circumventing regulatory scrutiny, with the transaction amount being three times Groq's previous valuation.

GPU chip (3)

Strategic Defense: Containing the Threat of Google's TPU

The core objective of NVIDIA's move is to counter the growing challenges posed by Google's Tensor Processing Units (TPUs). Tech giants like Apple and Anthropic have begun to shift toward TPU-based model training, and Google's upcoming TPUv7 is already matching the performance of NVIDIA's Blackwell architecture. More importantly, the cost advantages of TPUs have become a bargaining chip for customers to pressure NVIDIA. By acquiring Groq's LPU (Language Processing Unit) technology, which is optimized for large language model inference, NVIDIA aims to address its shortcomings in the low-latency inference market.

Expanding into the Inference Market: From Training to Applications

NVIDIA CEO Jensen Huang stated that the company plans to integrate Groq's low-latency processors into NVIDIA's AI factory architecture to serve a broader range of real-time workloads. While NVIDIA GPUs still dominate in model training, the company faces increasing competition in the inference market from cost-effective solutions offered by Meta, OpenAI, and various startups. Although Groq has faced challenges in production capacity and revenue, its unique technical architecture is seen as a key incremental factor in enhancing NVIDIA's inference performance.

Industry Trends: Giants Adopt a "Consolidation" Approach