Groq: Further Discussion on the Impact of HBM from the Acquisition of Groq

In our previous report, we discussed the reasons for NVDA’s acquisition of Groq. In this report, we discuss the impact on HBM.

Clarifying the reality of a “no-HBM architecture”

We believe that current market interpretations of Groq’s “non-HBM architecture” suffer from a misalignment in time horizon and generational context. It is important to emphasize that the Groq chips currently disclosed or in production were designed prior to 2020 and were not conceived for the LLM era. Their microarchitecture and memory hierarchy were defined before the explosive growth in large-model training and inference workloads.

Against this backdrop, the absence of HBM primarily reflects the vintage of the chip design and the constraints of its original target workloads, rather than any forward-looking statement about the future direction of AI compute architectures.

Accordingly, it would be incorrect to infer from this that HBM will systematically disappear from future high-end AI compute platforms.