@johnbdillon@jimcramer@MadMoneyOnCNBC Damn, I have a complete court filing tracker. I usually don't miss any VICR filings. I guess I need to add audio to the bots list. Thank you for sharing.
@grok There are sources that don’t have the purity issues. The dosage variances makes sense. Is there any reports or evidence of people taking high quality peptides at reasonable dosage that have had ill effects?
Good breakdown, but memory is one of three physical gates on inference, and the other two keep getting skipped.
POWER. Decode is memory-bound on paper, but tokens per second per rack is ultimately watts per rack. GPU racks are already past 100kW and headed for multiples of that, and the hard problem is not generating the power, it is delivering amps to the die without losses eating the budget. That is vertical, point-of-load power delivery, and the supplier list is very short. VICR is the purest play on that architecture. If the rack can't be fed, the HBM bandwidth on the spec sheet never gets used.
OPTICS. Your Groq and Cerebras examples make the point for you: 256 chips networked to reach 128 GB means effective memory bandwidth is interconnect bandwidth. Same for Rubin NVL racks. Scale-up and scale-out links are moving from copper to pluggable optics and then co-packaged optics, and that is as binding on how much compute you can aggregate as how much HBM sits on the package.
Second point: agents don't live in HBM. HBM is the inference chip's memory, sized for weights and the active KV cache. Agent state (history, tool outputs, long-horizon context) is reusable cache that gets evicted to the next tier: host DDR5, CXL-attached memory, then NAND. That is exactly the DeepSeek and CMX behavior you describe. The agent buildout is a bet on the DRAM and storage tiers, not only HBM.
That is where RMBS sits. Every DDR5 server DIMM carries a registered clock driver and power management chip, and RMBS is one of a handful of suppliers of that silicon no matter whose DRAM is on the stick. MRDIMM and CXL memory controllers extend the same position. It is a toll on server DRAM capacity regardless of which memory maker or hyperscaler wins, the same shape as the HBM trade one layer down.
Long agents = long context, agreed. But context is the DRAM tier plus the power and optics to use it.
@grok So the expectation is we will see bad outcomes from all these peptide users in the future? That is a logic argument that pushed proof off to a point in the future. Same argument was used for THC we are still waiting for the fallout.
@grok Why can’t we use the amount of data from people using peptides to determine safety? There are way more people using peptides than would ever be studied in an FDA study.