BREAKING NEWS🚨: HBM CHIPS CAN REACH OVER 1000 tok/s/user on giant 700B+ models via agent swarm-written software, according to a new NVIDIA AgentX submission. Furthermore, on Rubin NVL72, it can achieve 20.9x better perf per TCO than GB300 NVL72 at even "ultrafast" speeds of 325 tok/s/user and up to 9.11x better perf per TCO at 200 tok/s/user. (1/2)🧵

Very positive. My opinion is that the optical sector underperformance recently from $AAOI to $SIVE has been absurd. The demand visibility... is just way too stupidly high. AOI: “Even combined AOI, Coherent program altogether, is still very tough to meet the customer demand in the next 3 years” (2029) Elazr GM: "entire optical supply chain was facing major shortages." "The shortage will continue to the next few years" Sivers CEO said the same thing about InP laser demand imbalances expected for the next 3-5 years. Can go on and on about $LITE, $MTSI and other comments. We haven't even hit the inflection point with 1.6T, NPO, CPO scale out/up, and optics with memory (as seen with SK Hynix). Yet the industry is already bottlenecked by EML/CW and all your other upstream components from PDs/TIA/DSPs, transceivers, and soon FAU + others when CPO scales… I'm personally extremely comfortable watching this all play out, but just a little confused that markets don't know how to math a year or two ahead.