NVIDIA's Vera Rubin AI racks are becoming the next 'prized possessions', but at the same time, pricing is expected to increase significantly from Blackwell.
NVIDIA's Vera Rubin Is Reported to Cost Up To $7 Million Per Rack, yet the Demand Consistently Grows
When we talk about the complexity of NVIDIA's AI portfolio, it is without a doubt that Vera Rubin is one of the company's most sophisticated offerings, featuring seven newly-designed chips that target almost every element of the rack infrastructure. You are looking at the Vera CPU, the Rubin GPU, an enhanced NVLink 6 chip, and much more, which implies that the BOM for the Rubin rack would likely be much higher than Blackwell Ultra's. A report by DigiTimes has disclosed that a single Rubin rack could cost between $3 million and $7 million.
Now everyone's competing to see who has the most cash, because if a company doesn't keep up with this technological awakening, it might disappear like Yahoo! during the dot-com bubble. This is war, and the winner takes all.
- DigiTimes
The analogy with Yahoo is surely interesting, and we'll come to that in a bit. One bit mentioned in the report to look at is that, for server manufacturers like Foxconn and Quanta, profit margins have diluted since rack-scale deployment became prevalent, mainly because it is claimed customers aren't willing to pay standard margins as rack costs rise. If we assume that Rubin NVL72 costs around $3 million, this means $300,000 in 10% margins for server ODMs, which hyperscalers aren't willing to pay at all. This has resulted in manufacturers' profit margins per rack decreasing with each generation.
At the same time, the changes within rack architectures, whether it is modularity, liquid-cooling, advanced power delivery systems, and several others, have forced server manufacturers to invest heavily into R&D measures as well, which means that ultimately, ODMs aren't in the best of positions with Vera Rubin demand. But given that customer prospects with Vera Rubin are rising as well, the only option for ODMs is to stick with manufacturing, likely finding other methods to maintain their margins.
NVIDIA's Vera Rubin NVL72 has seen immense interest from hyperscalers in recent months, and almost all leading entities, including OpenAI, Amazon, Microsoft, and others, have placed orders. CEO Jensen Huang projected $1 trillion in revenue from Blackwell and Rubin between 2025 and 2027, and this doesn't include streams like networking and CPU-only deployments, indicating unprecedented demand. At the same time, hyperscaler CapEX commitments have reached a whopping $660 billion this year as well, driven by the fact that compute is the major bottleneck for them.
Relating the situation to what happened with Yahoo during the dot-com era is a pretty interesting way of saying what's happening in today's AI industry. Yahoo dominated early in the internet boom, driven by its suite of tools, including search, webmail, and more. Yet, as demand for internet products grew, the company failed to sustain the momentum and even passed on the opportunity to buy Google, simply because the firm believed its moat was strong enough. Compare this to spending on compute, and you'll realize that hyperscalers are worried about becoming another Yahoo.
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
