Rubin Ultra May Cut 192GB HBM to Save Supply-NVIDIA's 20% Cost Trim Changes the AI Trade

Rubin Ultra's proposed HBM cut looks like a supply-management move

Reports that Rubin Ultra could move from 288GB of memory to 192GB matter less as a chip rumor than as a supply signal. They suggest NVIDIA may be willing to trim memory capacity to protect shipment volume in a market already strained by Rubin's demand for HBM4 memory, DRAM, multilayer ceramic capacitors and power-management silicon.

The bill-of-materials trade-off is large

If the report is right, the economic case is straightforward: HBM's share of total cost would fall from nearly 40% to 28%, the per-rack bill of materials would drop from roughly $8 million to $6.4 million, and interconnect would take on more strategic importance. In that setup, a smaller HBM configuration is not just a spec reduction; it is a way to keep systems buildable when memory availability and cost are the bottleneck.

Why the timing matters now

The debate is splitting into two camps. Bulls see optimization: NVIDIA could keep peak compute intact while shifting emphasis toward system scale and interconnect. Bears see weaker HBM demand, which could temper pricing power next year.

The more useful frame is simpler. TrendForce says HBM specs are adjustable, with final specs tied to 2H26 verification. Until that review is complete, the near-term question is not whether Rubin Ultra will look weaker on paper; it is how much memory NVIDIA is prepared to trade away to secure supply and control system cost.

Investor read: memory faces configuration risk, interconnect could gain

The market's first reaction was blunt. After reports of Rubin Ultra moving from 288GB down to 192GB, SK Hynix and Samsung shares both plunged around 8%. That suggests investors are treating the story as configuration risk rather than just another AI-demand headline.

With TrendForce pointing to 2H26 verification as the decision point, HBM producers are increasingly trading as an event around supply certainty and final specs. At the same time, NVIDIA appears to be putting more weight on large-scale interconnect and the NVL576 architecture, up to 576 GPUs. If that path wins out, system-level exposure such as interconnect, switchgear, boards, laminates, power delivery, and rack integration could become relatively more important.

What to watch next