Nvidia has acquired a potential competitor, announcing last night a $20 billion cash deal with AI chip startup Groq. However, the official statement clarified that it's not an acquisition of Groq as a company, but rather a technology license to integrate Groq's products into future offerings. Founded in 2016 by a group of former Google engineers, Groq is the core team behind TPUs and focuses on designing high-performance AI accelerator chips. Nvidia's move can be seen as following in Google's footsteps, accelerating its efforts in inference. Technologically, 1) this represents Nvidia shifting from a single GPU (Graphics Processing Unit) approach to a dual-track model of "GPU + LPU," moving from "general-purpose parallel computing" to "dedicated deterministic inference." 2) Particularly in the storage path, from HBM (High Bandwidth Memory) to SRAM, HBM, while offering large capacity, suffers from bandwidth limitations. Groq's LPU relies almost entirely on on-chip SRAM (Static Random Access Memory), with bandwidth reaching 80 TB/s (approximately 10 times that of HBM). There's a view that the stronger the Groq narrative becomes, the weaker the storage narrative may become. This is something to watch for regarding the future trend of storage copyright.

This seems like a positive development. This merger appears to be more than just a resource consolidation; it also seems to involve technological innovation. NVDA should be in a period of rest. After that rest, it's time to get back on track.