In its Micron report, UBS said NVIDIA’s decision to lower the HBM specifications for Vera Rubin Ultra would actually increase the number of VR300 units entering the supply chain, raising total HBM consumption in 2027 from 58.7 billion Gb to 61.5 billion Gb.
As I said previously, high-spec HBM4/4E is currently a bottleneck that is constraining shipments.
People are not looking at the supply side nearly enough.
What matters right now is that making HBM in greater volume and with more superior specs cuts commodity DRAM supply far too much.
This also affects SOCAMM2 supply, which will affect total Rubin rack shipments.
In other words, cutting HBM to increase SOCAMM2 supply and raise total Rubin rack shipments is the rational move, and that fits Nvidia's interest.
The resulting performance loss is solved at the NPO layer and addressed at the cluster level.