If you believe Nvidia's fiscal 2028 number, you have to have a view on what moves the data between those chips. Arista did over $3B in Q2, up 37.7% year over year, trailing revenue up 32.6%. It trades at 62x trailing, 42x forward, at $195 with a 52-week range of $114 to $215. That is a much more expensive stock than Nvidia on forward earnings — 42x versus 18x — for materially slower growth. The bull case has to be that networking is a longer-duration position in the buildout than the accelerator itself: chips get replaced on a two-year cycle, the switching fabric around them doesn't. I'm not sure I believe that's worth double the multiple. Posting it because I want someone to tell me what I'm missing on the durability argument. $NVDA

Vistra is down to $136, more than a third off its 52-week high of $219.82, while the actual AI-power story hasn't changed — it's arguably gotten stronger. The stock trades at 23x earnings. Constellation, doing roughly the same trade, trades at 26x. Neither multiple looks like the market believes datacenter power demand is a multi-year structural story, which is the whole thesis everyone claims to hold on NVDA. The mechanism is straightforward: AI capex needs electrons before it needs chips, and Texas (ERCOT) is where a huge share of the new datacenter load is trying to interconnect. Vistra is a merchant generator sitting inside that market with no rate-case process slowing it down. If the AI buildout is real, this is one of the more direct ways to be long it without paying a chip multiple for it. The reason it's cheap isn't a mystery — it's the risk everyone's pricing in. $VST