Nvidia Cut the Memory in Each Rubin Ultra Chip. UBS Says Total Memory Demand Goes Up Anyway

Nvidia has reduced the amount of high-bandwidth memory it plans to put in each Rubin Ultra GPU, according to an August 11, 2026 UBS report — not because the chips need less memory, but because the world cannot make enough of it. HBM, the stacked memory that sits next to an AI processor and feeds it data, comes from only three producers: SK Hynix, Samsung, and Micron. UBS says supply is tight enough that Nvidia cut the planned memory content of each VR300-series Rubin Ultra to 512GB of HBM4E from an originally planned 768GB.

The middle row is the one that looks like a typo and isn't. Cutting the memory in each chip raised the bank's estimate of total memory demand. The logic: memory, not processors, is the binding constraint on how many AI chips Nvidia can ship. Put a third less memory in each GPU and the same constrained memory supply supports meaningfully more GPUs — and UBS judges the resulting increase in VR300 shipments large enough to lift total 2027 memory consumption to 61.5 billion gigabits from its prior 58.7 billion estimate.

For the memory makers, every part of this is good news. Their product is scarce enough to force the world's most valuable chipmaker to redesign its flagship, demand keeps growing anyway, and pricing power keeps compounding — UBS now models blended HBM average selling prices rising about 79% year over year in 2027, up from the 67% it previously forecast. A separate industry analysis this week made the same point from the cost side: memory has become the single largest cost component inside Nvidia's Vera Rubin systems, ahead of the processor silicon itself.

For Nvidia the calculation is more defensive. A 512GB Rubin Ultra is still a generational leap, and shipping more units at lower memory content protects volume and revenue while the shortage lasts. But it hands a visible lever to its suppliers: the scarcer HBM gets, the more of the AI computing bill flows to the memory industry rather than to the company whose name is on the chip. We covered the underlying shortage when it first crossed on August 6 — this report is the first hard number on what it costs the product.