NVIDIA's Rubin & Rubin Ultra servers are going to be super expensive, with a large portion of the costs going towards memory alone.
HBM & DRAM Costs Push NVIDIA's Next-Gen Rubin "Oberon" & Rubin Ultra "Kyber" Servers Way Beyond Blackwell Prices
Rising DRAM and equipment costs will push NVIDIA Rubin server prices up into the stratosphere. You must have thought the Blackwell servers were crazy expensive, but the Rubin generation will hit an historic high in per server costs.
According to BofA Global Research's estimates, the breakdown of the Rubin and Rubin Ultra racks shows a massive price bump, mainly due to the increased HBM costs.
Let's start with the basics: the NVIDIA NVL72 rack is called Oberon and makes use of 72 GPUs plus 36 CPUs. A single Vera Rubin tray houses 4 "Rubin" GPUs and 2 "Vera" CPUs. Two GPUs and a single CPU are housed on a motherboard, which is called Superchip. There are 36 Superchips on the NVL 72 rack. So that's a total of 72 GPUs and 36 CPUs.
Each Rubin GPU houses 288 GB of HBM4 memory, and each Vera CPU comes with 1.5 TB of LPDDR5X memory. For an NVL72 rack, that's 20.7 TB of HBM4 memory and 54 TB of LPDDR5X memory. There is a lot more that goes into Vera Rubin NVL72 racks, such as networking, cooling, power, interconnects, etc.
We previously discussed the full system breakdown in the estimated BoM shared by Morgan Stanley Research. Those were full unit costs covering the GPUs, CPUs, memory, and the rest of the infrastructure that goes into making the Rubin VR200 "Oberon" racks.
Blackwell B200 vs Rubin V200 Estimated Costs
As per the new estimates, the Rubin VR200 "Oberon" rock will comprise 72 Rubin V200 GPUs with a total of 288 GB HBM4 memory per chip and a total of 20,736 GB of HBM4 memory per rack. The HBM4 memory alone on these racks will amount to $382,000 US, and that's excluding the LPDDR5X costs. The cost per GB is $18.40.
Compared to Blackwell, the HBM costs for Rubin will be over twice as expensive, and the $/Gb cost will rise from $11.26 to $18.40. Each Rubin server will have an ASP of $6 million per rack, while Blackwell is around $3 million per rack. The HBM alone makes up 6.4% of the rack costs.
| Metric | B200 (Blackwell) | V200 (Rubin) |
|---|---|---|
| GPU | B200 | V200 |
| Architecture | Oberon | Oberon |
| Ramp | 2H24-2H25 | 2H26-2H27 |
| GPUs/rack | 72x | 72x |
| HBM/GPU | 192 GB | 288 GB |
| HBM GB/rack | 13,824 GB | 20,736 GB |
| HBM Type | HBM3e | HBM4 |
| $/GB | $11.26 | $18.40 |
| HBM $/rack | $156 | $382 |
| ASP $/rack | $3,000 | $6,000 |
| HBM % of rack cost | 5.2% | 6.4% |
Blackwell Ultra B300 vs Rubin Ultra V300 Estimated Costs
For Rubin Ultra, the price increase is even more staggering. Each Rubin Ultra V300 chip will pack 576 GB of HBM4e memory, and NVL144 "Kyber" racks will feature up to 82,944 GB of HBM4e DRAM.
While the $/GB is expected to remain similar to Rubin V200, around $18.49, the overall HBM costs per rack will swell to $1.534 million, a 4x increase over standard Rubin "Oberon" racks and almost a 5x increase over Blackwell Ultra. Each Rubin Ultra "Kyber" Rack is expected to have an ASP of $21 million, up from $4 million for Blackwell Ultra.
| Metric | B300 (Blackwell Ultra) | V300 (Rubin Ultra) |
|---|---|---|
| GPU | B300 | V300 |
| Architecture | Oberon | Kyber |
| Ramp | 2H25-2H26 | 2H27-2H28 |
| GPUs/rack | 72x | 144x |
| HBM/GPU | 288 GB | 576 GB |
| HBM GB/rack | 20,736 GB | 82,944 GB |
| HBM Type | HBM3e | HBM4e |
| $/GB | $15.27 | $18.49 |
| HBM $/rack | $317 | $1,534 |
| ASP $/rack | $4,000 | $21,000 |
| HBM % of rack cost | 7.9% | 7.3% |
As per Foxconn, NVIDIA's Vera Rubin AI datacenters will cost north of $47 billion per gigawatt of development, with the power bills hitting $ 1.3 billion annually. NVIDIA Rubin isn't just expensive; it'll also cost a lot to operate these servers. At the same time, Rubin is expected to offer the lowest TCO and token cost in the industry with unfathomable levels of AI performance.
Considering just how big the demand is for Rubin, it is expected to outpace Blackwell big time. Leading AI firms are already installing the first Vera Rubin systems, and with full-volume production ongoing on these AI powerhouses, the world will soon get to see why these AI firms are spending huge sums of money on large-scale AI data centers. NVIDIA is on track to deliver the leading AI capabilities across the globe.
News Source: @SemiconductorsX
About the author: A Software Engineer by training and a PC enthusiast by passion, Hassan Mujtaba serves as Wccftech's Senior Editor for hardware section. With years of experience in the industry, he specializes in deep-dive technical analysis of next-generation CPU and GPU architectures, motherboards, and cooling solutions. His work involves not only breaking news on upcoming technologies but also extensive hands-on reviews and benchmarking.
Follow Wccftech on Google to get more of our news coverage in your feeds.
