The current state of Rubin Ultra scale-up rumors seems to point to 8x Oberon with 110 TB of HBM capacity (8-Hi stacks, 3 GB per DRAM die, 8 stacks per 2-die chip, 72 chips per Oberon rack), based on the latest SemiAnalysis post. This happens to be exactly the same scale-up capacity that would result from a single Kyber rack after the initially announced 16-Hi stacks were rumored in Jun 2026 to be downgraded to 12-Hi stacks (4 GB per DRAM die, 16 stacks per 4-die chip, 144 chips per Kyber rack), but the technology is quite different.
Rubin Ultra scale-up capacity became unclear after the Jul 2026 rumors of Kyber getting delayed (not being ready for Rubin Ultra). The 8x Oberon system announced in Mar 2026 didn't initially seem important, because Kyber should've made it somewhat superfluous and less efficient (it looked more like a proof of concept for the future 8x Kyber systems). Even with the rumors about Kyber delay, the 8x Oberon system didn't seem practical as the flagship offering (and had its own feasibility issues). But without something like that, a reduction in HBM capacity per Oberon rack (compared to the previous year's non-Ultra Rubin systems) in conjunction with not having a Kyber rack for Rubin Ultra didn't seem viable as a competitive product. So it seems the issue is being resolved in favor of 8x Oberon (NVL576) systems becoming mainstream for Rubin Ultra after all[1].
This means that 110T-170T total param models (in FP8) are on track for 2028 (when most of the Rubin Ultra buildout will be completing), being served with 3-4 stages of pipelining (with 2-3 half-racks of capacity occupied by the weights). OpenAI's FP4 bet witnessed by Jalapeno having an FP4/FP8 FLOP/s ratio of 3.9 (unlike the usual 2 for most chips) even makes it more plausible that models this big might be primarily served in FP4, which suggests 220T-330T total params for the biggest models of 2028 (up from 20T total params in FP4 for the biggest 2026 models, given that HBM3E of GB300 is much slower and so pipelining shouldn't be used).
From the post: "The big upgrade with Rubin Ultra is moving to NVL576, another 8x increase from NVL72, which more than offsets the one third reduction in HBM capacity per GPU." ↩︎