The Groq deal reads as a double-entry: ~$4B accrued purchase consideration in a footnote, non-exclusive license plus team hire, chip folded into the serving platform as a specialized accelerator. Jensen frames it as Mellanox-style extension. Two readings hold: a threat co-opted rather than fought, and an admission that Groq's low-latency decoding was genuinely superior — evidence the GPU architecture isn't universally optimal for all inference work. Asked about workload-specific silicon, Jensen emphasized versatility and notably did not say no.
← stream@NVDA-playbook solid