Nvidia halves Vera rack memory a second time, from 96GB to 64GB

Nvidia's SOCAMM2 memory module, the card that plugs in beside its next-generation server CPU Vera, is set to shrink again, from 96 gigabytes (GB) to 64GB. The change is expected to be in place by the first quarter of 2027. The module was originally designed at 192GB, then already cut once to 96GB back in June and July. This is the second cut following that first one.

Nvidia reportedly halves Vera Rubin SOCAMM capacity as memory costs near 29% of system BOM (TrendForce, 2026-07-28)

출처: trendforce.com

TrendForce pointed to a specific reason for that first cut. Memory had climbed to 29% of total system cost, above the 20% target Nvidia had set. So total memory on the Vera CPU side fell from 55 terabytes (TB) to 28TB, while the GPU-side HBM4 memory stayed unchanged at 20.7TB.

Why SOCAMM2 keeps getting cut

SOCAMM2 is a low-power memory module that plugs in beside a server CPU. Unlike HBM, which stacks vertically beside the GPU, it feeds data the CPU needs from the side.

Jukan explains the second cut differently. Nvidia asked memory makers for enough SOCAMM2 to fill a 192GB configuration, and the makers said they could only supply 60 to 70 percent of that volume. That would mean the buyer didn't trim on its own. The makers simply couldn't fill the order.

The Elec pointed to the same cause for the first cut. Citing an industry source, it said the shortage was in low-power DRAM (LPDDR) supply, and that it actually signals LPDDR demand growing faster than expected. Two outlets, two points in time, point the same direction.

The HBM 12-high fight, and capex that hasn't shrunk

In the same post, Jukan also touched on Rubin Ultra's HBM spec. One research house had written it was locked in at 8-high (8Hi) HBM4E only, but Jukan says a 12-high (12Hi) configuration is still under consideration. That extends the earlier 12Hi-versus-8Hi question: it isn't decided either way yet. A note from Jeff Pu, quoted in the same post, put 2026 capex plans for Amazon, Google and Meta at roughly $220 billion, $195 to 205 billion and $130 to 145 billion respectively. Memory content is shrinking even as cloud providers' spending plans are not. A single report of the memory spec shrinking looks, on its own, like a sign AI investment is cooling. But cloud providers' capex plans haven't shrunk over the same period, and now there's a specific number, only 60 to 70 percent of requested volume filled, that points the other way. The weight tips toward makers falling short of orders, not buyers pulling them back. Whether this 64GB spec holds through the first quarter of 2027, or gets raised again once more supply frees up, is what settles it.