Memory now makes up 26% of a Rubin rack's cost. It was 9% on Blackwell.

Memory now makes up 26% of the total bill-of-materials cost in Nvidia's next-gen AI server rack, Vera Rubin, according to Morgan Stanley. On the prior generation, Grace Blackwell, that share was just 9%. Chip analyst Jukan relayed the shift, citing checks from Taiwan's Fubon Securities and other research desks. Per Fubon, Micron ($MU) channel checks suggest Nvidia's HBM4 cost will rise to $31-32 per gigabyte, from $17-18 for HBM3E in 2026. That's also pushing estimates for a single Rubin GPU's price toward $78,000-$80,000.

Why did memory get this expensive

Morgan Stanley pegs the bill-of-materials for one rack, 72 GPUs across 36 superchips, at about $7.8 million. GPUs account for roughly $4 million of that (up 57% from Blackwell), and memory (HBM4 plus LPDDR5X) comes to about $2 million. Memory cost just $373,900 on Grace Blackwell, so it's jumped more than fivefold.

Bernstein estimates that without changes, memory would eat up 29% of bill-of-materials cost, versus a 20% target, and says Nvidia needs to trim the memory configuration itself to get there. GF Securities' checks suggest Nvidia is looking at halving CPU-side SOCAMM capacity, from 192GB to 96GB, and cutting Vera CPU memory from 54-55TB to 28TB. GPU-attached HBM4 capacity, 20.7TB per rack, stays untouched.

Why Nvidia isn't backing off anyway

Per the checks Jukan relayed, Nvidia plans to keep the rack's basic layout unchanged (72 GPUs, cable for scale-up, optical for scale-out) despite the cost jump. The reason is token economics. Blackwell processes about 80,000 tokens per second at 150 megawatts; Rubin does roughly 800,000 tokens per second in the same power envelope, about 10 times more. For cloud providers, that means even pricier GPUs can still pencil out if cost per token keeps falling.

Other chipmakers hit the same wall

The checks Jukan relayed also covered Google and Intel. Google reportedly plans to deploy 12-15 million of its own AI chips (TPUs) by 2028, when it moves to a ninth-generation TPU with four compute dies, more than doubling throughput. Intel's advanced packaging (EMIB) capacity is expected to double, from 10,000-12,000 units by the end of this year to 24,000-25,000 by the end of next year. The memory and packaging bottleneck isn't unique to Nvidia.

What actually settles this

It comes down to one thing. How much of the memory cost increase gets offset by falling cost per token. Here's the test. If cloud providers take delivery of Rubin on schedule, shipments in Q3 and volume ramp in Q4, the token-economics logic wins. If deployment gets delayed or scaled back, memory costs were the real drag.