@Damnang's latest: Vera Rubin's DRAM cut in half — is the memory cycle really over?

On June 4, SemiAnalysis reported NVIDIA lowering the Vera Rubin SOCAMM standard build from 192GB to 96GB modules, halving CPU-side memory per rack from ~54TB to ~28TB. Memory and storage stocks fell the next day.

The piece starts by separating two things the market merged:

- NVIDIA announced nothing — the official spec still lists up to 54TB

- What SemiAnalysis reported is a supply-chain account that the shipping build drops toward 96GB

- The math checks out (8 SOCAMM per Vera CPU x 36 CPUs per rack), but a shipping account and an official spec are different things

It's also where three earlier threads converge: the bottleneck moving outside the GPU, who actually gets paid when a bottleneck forms, and how much demand truly has to stay in HBM.

The through-line: what the market priced in and what the report actually said were not the same thing. Worth reading for anyone watching AI memory.