NVIDIA Rubin Ultra rack costs could reach $21 million as HBM4e memory drives AI server prices higher

NVIDIA’s next generation Rubin and Rubin Ultra AI servers are expected to push data center hardware costs far above the current Blackwell generation, with memory becoming one of the biggest cost drivers. New estimates suggest that a Rubin Ultra rack could reach an average selling price of around $21 million, while the HBM4e memory alone may cost about $1.5 million per rack.

The figures show how expensive AI infrastructure is becoming as demand for high bandwidth memory, advanced CPUs, GPUs, networking, cooling, and power delivery continues to rise. Blackwell systems were already seen as extremely costly, but Rubin appears set to move the market into a higher tier.

The standard Rubin rack, known as Oberon, is expected to use 72 Rubin GPUs and 36 Vera CPUs. Each Rubin GPU reportedly carries 288GB of HBM4 memory, while each Vera CPU is paired with 1.5TB of LPDDR5X. That gives one Rubin NVL72 rack about 20.7TB of HBM4 and 54TB of LPDDR5X memory.

Rubin Ultra raises the memory load sharply with 144 GPUs per rack

Rubin Ultra, also referred to as Kyber, increases the scale even further. The V300 Rubin Ultra rack is expected to include 144 GPUs, with each GPU carrying 576GB of HBM4e. That brings total HBM4e memory per rack to 82,944GB, or nearly 83TB.

The memory cost is where the jump becomes clear. A standard Rubin V200 rack is estimated to carry about $382,000 worth of HBM4. Rubin Ultra pushes that figure to about $1.534 million for HBM4e alone. That is roughly four times higher than the standard Rubin rack and far above current Blackwell Ultra estimates.

The higher memory cost does not come only from capacity. HBM4 and HBM4e are complex products that require advanced packaging and tight integration with AI accelerators. As AI companies demand more compute and more memory bandwidth, memory suppliers have more pricing power. That pressure is already affecting consumer DRAM and server memory markets, and AI racks sit at the highest value end of the supply chain.

The cost of operating these systems is also expected to be enormous. Foxconn has estimated that Vera Rubin AI data center deployments could cost more than $47 billion per gigawatt of development, with yearly power bills reaching about $1.3 billion. That shows why AI infrastructure is no longer only a chip story. It is also a power, cooling, memory, and construction story.

Still, large AI firms are expected to keep buying these systems because the performance gains can lower the cost per token and improve total cost of ownership at scale. In simple terms, the upfront price is huge, but companies training and running large AI models may still see value if Rubin delivers enough performance per watt and per rack.

The Rubin generation shows where the AI server market is heading. Racks are becoming larger, memory capacity is rising quickly, and system prices are moving into territory that only the biggest cloud and AI companies can absorb. For NVIDIA, that demand keeps its data center roadmap highly valuable. For the rest of the market, it signals that memory supply and power availability will remain major pressure points for years.