Nvidia's Next AI Rack Costs Nearly Double the Last One, and Memory Is Why. What Infrastructure Buyers Should Budget For.

The single most expensive part of Nvidia's next-generation AI rack is no longer just the GPU. According to a Morgan Stanley bill-of-materials analysis, memory now accounts for roughly a quarter of the total, and the overall system price has nearly doubled generation over generation. For any organization budgeting AI infrastructure for the second half of 2026 and beyond, that shift changes how the numbers should be built.

The numbers

Morgan Stanley Research estimates that Nvidia's Vera Rubin VR200 NVL72 rack will cost hyperscale cloud providers around $7.8 million per unit, up from roughly $4 million for the current GB300 NVL72 generation. The GPUs themselves are not the driver of the increase. The bank estimates Nvidia will charge about $55,000 per Rubin GPU and $5,000 per Vera CPU when sold in volume.

The memory line is where the cost concentrates. Morgan Stanley puts memory content per rack at about $2 million, a 435% jump from the prior generation, driven by a threefold increase in LPDDR5X capacity to 54 terabytes per rack plus roughly $1 million or more in 3D NAND storage that was virtually absent from earlier systems. First shipments are scheduled for the third quarter of 2026, with volume ramping in the fourth.

As a share of the bill of materials, memory rose from 5 to 10 percent on the GB200 to 25 to 30 percent on the VR200, while the GPU's share fell from roughly 63 percent to 51 percent.

Why this reaches beyond hyperscalers

Only a handful of companies buy racks at this scale. But the force behind the price, constrained supply of DRAM, NAND, and high-bandwidth memory against AI-driven demand, is the same one now raising component costs across the market. Contract DDR5 pricing has climbed sharply, and device makers including Samsung and Apple are expected to pass higher memory costs into upcoming products. The AI buildout and the price of the memory in a laptop are now tied to the same supply squeeze.

For enterprise technology buyers, that means memory has become a strategic budget variable rather than a rounding error, whether the purchase is a data center commitment or a fleet refresh.

What infrastructure planners should account for

Three considerations follow directly from the Morgan Stanley breakdown.

• Model your AI spend on total system cost, not GPU count. The generation-over-generation story is that peripheral components, memory, printed circuit boards, networking, cooling, and power, are absorbing a larger share of every dollar. Budgets built around GPU headline prices will understate the real bill.
• Treat memory supply as a scheduling risk, not only a cost. With memory representing a quarter or more of rack value and supply tight, availability and lead times deserve the same planning attention as accelerator allocation.
• Revisit the build-versus-rent math. As rack prices climb toward $8 million and memory pricing stays volatile, the calculus between owning infrastructure and consuming cloud capacity shifts. Morgan Stanley notes that a hyperscaler self-sourcing certain memory modules could bring the rack price down to around $6.7 million, a reminder that procurement structure now materially moves the number.

The bottom line

Nvidia's Vera Rubin platform will deliver a step change in performance, and demand is not in question. What has changed is where the money goes. The AI infrastructure bill is increasingly a memory bill, and the organizations planning capacity for the next two years will be the ones that budget for the full system rather than the chip on the front of the box.