Shay, link here. Once you've read the headline(s), this is the post I found most helpful.
- Joann Stern, WSJ, broke the story on my x thread, but it was Shay who really provided the back story.
Jonathan Ross, linked here.
The wealth wizard, link here. Note Groq.
1. This was about owning inference economics, not fixing a chip gap
Nvidia didn't aqui-hired Groq because it was behind on chips since Nvidia already dominates training and most inference & its roadmap (GB300, Rubin) continues to push cost-per-token down while expanding performance faster than nearly anyone else. Training is a one-time event while inference is where the new AI business model lives so as AI moves into real products the money shifts to whoever controls runtime.
2. The future where inference escapes Nvidia just got absorbed
Groq was one of the few credible proofs that latency-sensitive inference could eventually move off GPUs and over time that would have chipped away at Nvidia’s “unavoidable” status. The risk was amplified by Groq’s founder Jonathan Ross, who previously built TPUs at $GOOGL and already proved that custom silicon can compete in real workloads. This deal shuts that door before it could scale.
3. Deterministic inference was the missing layer
GPUs excel at flexibility and scale but they were never designed to guarantee perfectly consistent response times. That matters because real-world AI breaks when latency jitters: voice assistants pause, live translation lags, agentic workflows compound delays. Groq solved this by designing around large amounts of SRAM by keeping data close to the processor and delivering quick responses every time. That made Groq uniquely suited for real-time AI where latency matters more than peak throughput.
At this point, it’s hard to argue Nvidia just sells chips when it’s clearly building the platform that owns training, networking, and now real-time inference.
$20B today to avoid a $200B problem later.
No comments:
Post a Comment