NVIDIA Vera Rubin’s Memory Bill Hits $2M, More Than Doubling Grace Blackwell’s Cost

NVIDIA’s next-generation Vera Rubin AI rack is shaping up to be one of the priciest pieces of computing hardware ever assembled, and memory is the biggest reason why. According to a Morgan Stanley Research bill-of-materials estimate that has now been corroborated by multiple outlets, a single Vera Rubin “VR200” NVL72 rack carries a projected total cost of roughly $7.8 million, up 95% from the current Grace Blackwell generation, driven largely by a memory bill that has more than quintupled.

According to a Morgan Stanley research report, the bill for a single VR200 NVL72 rack is estimated to add up to $7,803,148, with $2,001,600 of those costs attributed to memory, a 435% increase in memory cost over the GB300. That memory line covers combined HBM4 and LPDDR5X, and it jumps from $373,939 in Grace Blackwell to over $2 million on the Vera Rubin platform.

HBM4 and LPDDR5X Push Memory to Roughly a Quarter of Rubin’s Bill

The share of the total system cost eaten up by memory has swelled dramatically between generations. NVIDIA’s BoM for its upcoming Vera Rubin “NVL72” rack shows a massive surge in memory prices that now make up 26% of the total system cost, up from roughly 9% on Grace Blackwell: a shift Fudzilla summed up bluntly, noting memory now accounts for more than a quarter of the whole system cost, showing how HBM4 and LPDDR5X supply pressure is warping the AI buildout.

Even with that spike, memory still isn’t the single priciest line item. The Rubin GPUs are expected to cost almost $4 million, making them the single-most expensive cost within the NVL72 VR200 rack: a 57% bump over the Blackwell NVL72 B300 rack, which had the GPU price around $2.5 million, or about $55,000 per GPU. Printed circuit boards saw an even steeper percentage jump than memory: PCBs see the second biggest jump, rising 233% from $35,100 in Blackwell to $116,730 on Rubin.

20.7TB of HBM4 and 54TB of LPDDR5X Packed Into One NVL72 Rack

The eye-watering price tag comes down to sheer volume. Each Rubin GPU houses 288GB of HBM4 memory, and each Vera CPU comes with 1.5TB of LPDDR5X memory, adding up to 20.7TB of HBM4 memory and 54TB of LPDDR5X memory for an NVL72 rack, which is built around 36 Superchips for a total of 72 GPUs and 36 CPUs.

Put another way, that’s around 74.7TB of combined memory sitting inside a single cabinet: enough, by Shane the Gamer’s own rough back-of-envelope math against a typical 16GB smartphone memory config, to roughly match the combined RAM of well over four thousand handsets. It’s a useful sense-check on just how much silicon NVIDIA is now packing per rack compared with Grace Blackwell, where each Rubin GPU carries 288GB of HBM4 memory while each Vera CPU comes with 1.5TB of LPDDR5X memory had already been a notable step up.

NVIDIA Reportedly Trims SOCAMM Memory to Keep Costs Under 20%

The cost pressure is apparently steep enough that NVIDIA is now scaling back planned memory configurations rather than let costs run away. TrendForce reported that without adjustments, memory could account for around 29% of the Vera Rubin VR200 system’s estimated $2.1 million bill of materials, well above NVIDIA’s preferred level of 20%, citing analysis that NVIDIA is reducing the memory capacity of its rack-scale system to cope with elevated prices and supply shortages.

The scale of the trim is significant: NVIDIA is reportedly moving to 96GB SOCAMM modules for each Vera CPU in its Vera Rubin NVL72 racks, down from the previously planned 192GB, a change that would see total Vera CPU memory decline from about 55TB to 28TB per rack. TrendForce also noted a supply squeeze feeding into the decision, pointing to preliminary allocation plans from Samsung, SK hynix, and Micron showing NVIDIA is expected to receive enough LPDRAM to meet only about 60% of its estimated demand. Some estimates put the risk even higher without those cuts, one Bernstein-sourced figure cited by TrendForce suggested a single Vera Rubin NVL72 rack could cost as much as $9.1 million before mitigation.

Jensen Huang Bets Vera Rubin Beats Blackwell Despite the Price Tag

Despite the cost blowout, NVIDIA is confident demand for the platform will outstrip even Grace Blackwell’s run. Responding to an investor question, CEO Jensen Huang said “Every single frontier model company will jump on Vera Rubin from the get go, and that was not true before on Blackwell,” adding that Vera Rubin was already off to a strong start. NVIDIA’s broader financial position backs up that confidence, with the company’s latest earnings call reporting record revenue during the first quarter of 2026, to the tune of $81.6 billion, the bulk of it from AI infrastructure and data-centre sales. Shipments of the new hardware are already locked in, with NVIDIA planning to begin first Vera Rubin shipments in the third quarter of 2026, followed by a broader production ramp in Q4.

For gamers and PC builders in New Zealand and Australia, the story isn’t just an abstract enterprise footnote. Hyperscalers and cloud providers hoovering up HBM4 and LPDDR5X supply at this scale is the same dynamic already pushing up consumer DDR5 and GDDR pricing worldwide, meaning the next round of graphics card and memory kit price hikes reaching local retailers can be traced directly back to racks like this one competing for the same fabs and the same wafers.