NVIDIA is redesigning its flagship chip around what memory it can actually get, not what's optimal

The Rubin Ultra reporting this week is that NVIDIA's testing configs with as little as 192GB of HBM, down from a planned 1TB, and even stepping back from HBM4E to HBM4. Not for performance, purely because the memory shortage is bad enough that it's limiting how many racks get built. Nothing's final but multiple firms are describing the same thing.

The part that matters imo for holding AI names: what NVIDIA gives up in memory it reportedly takes back in optics, more **photonic** interconnect instead of more memory stacked next to the chip. So the value doesn't disappear, it moves to a different part of the supply chain. And you could see the market chewing on that through early August, memory names under pressure while photonics and co-packaged-optics names caught a bid.

What's been on my mind is what this does to the "just own AI" thesis. For most of this year you could own the broad AI trade, memory especially, and ride it up. But if the whole chain is short at once, lasers reportedly sold out years deep, TSMC apparently sitting on finished chips it can't package for lack of memory, then OEMs keep making forced trade-offs, and each one shifts money between verticals on a timeline that has nothing to do with actual end demand.

Any clean answer for how you position around that?

Picking the right layer suddenly matters way more than it did six months ago, but the winning layer keeps changing based on whatever's scarcest that quarter.