The new NVIDIA Vera Rubin server platform looks like a tech blockbuster: raw compute, enormous memory—and a sticker shock to match. One NVL72 system runs into the millions; most of that tab comes from memory modules, plain and simple.
The NVL72 (codename Oberon) packs 72 Rubin GPUs and 36 Vera processors. Memory-wise: each accelerator gets 288 GB of HBM4, and each CPU can be fitted with up to 1.5 TB LPDDR5X—i.e., a single rack ends up with roughly 20.7 TB of HBM4 plus about 54 TB of LPDDR5X. That doesn’t even count the rest of the kit: heavy-duty cooling, chunky PSUs, networking gear and high-speed interconnects. (FYI: those extras are neither cheap nor lightweight.)
UBS did a line-by-line cost breakdown. For one Superchip, the total is quoted at about $38,902, with roughly $24,297 tied to HBM4 and SOCAMM2. The Vera processor number is striking: ~$20,059 per chip, but nearly all of that—about $19,355—is memory. The bare CPU, stripped of its memory, would be roughly $704. So yes, memory dominates the bill: ~62% of a Superchip’s cost, and it only balloons when you scale to a full rack.
Buying one of these means paying a premium for capacity. If you’re into monstrous memory pools (eg. training giant models or massive in-memory datasets), it makes sense on paper; if not, expect sticker shock. Sources: wccftech