The anticipated launch of Nvidia's Vera Rubin AI server rack has sent ripples through the tech industry, with its price projected to reach $7.8 million. This figure represents a dramatic increase from the current Blackwell GB300 rack, which retails for under $4 million. The nearly doubled cost reflects a broader trend in the semiconductor market, particularly affecting high-performance AI servers.
Rising Component Costs
Market analysis shows that memory components are not the only factors driving rising costs. A report by Morgan Stanley analyst Howard Kao indicates that the share of memory in the bill of materials for the Vera Rubin rack has surged from just 5-10% to 25-30%. Consequently, the cost of memory has skyrocketed by as much as 435%. In contrast, the share of GPUs in the overall rack cost has decreased from 65% to about 51%.
These findings illustrate how significant shifts in component pricing can alter the economics of server manufacturing. Other components, such as Printed Circuit Boards (PCBs) and Multilayer Ceramic Capacitors (MLCCs), are also experiencing considerable price increases. For instance, the cost of PCBs has risen from around $35,000 for the Blackwell rack to approximately $117,000 for the Vera Rubin—a 233% increase. Similarly, the price of MLCCs has surged to about $4,300 per rack, up from $1,500.
Market Dynamics and Future Implications
Procurement dynamics are changing as well. Hyperscale cloud providers like Microsoft and Google may choose to buy memory modules directly, potentially reducing the rack price to around $6.7 million by avoiding Nvidia's markup. This could significantly impact Nvidia's revenue from memory sales, making it a critical area for investors to watch.
Meanwhile, Original Design Manufacturers (ODMs) are expected to see their value-added per rack rise by 35-40%. Despite a decline in gross margins—from about 2.7% for the Blackwell to approximately 1.9% for the Vera Rubin—absolute profits are projected to improve. Kao emphasized that “investors should focus on the growth of absolute dollar profits, not just the margin percentage.”
The Emergence of Consignment Models
An emerging trend in the industry involves the adoption of consignment business models, where cloud providers source core components directly, leaving ODMs responsible for assembly. This shift could ease working capital pressures for ODMs but may also constrain their revenue growth. As the market evolves, it remains uncertain how widespread this model will become, and analysts warn of potential long-term implications if anticipated profit growth does not materialize.
Upgrades in Power and Cooling
The Vera Rubin platform is expected to standardize on a 110kW power module, with some cloud providers already implementing independent HVDC power racks. Analysts anticipate the introduction of 800V DC systems by 2027, with initial deployments planned for 2026. The Vera Rubin racks will feature a fully liquid-cooled design, enhancing thermal performance while driving up the total value of liquid cooling components to around $72,080 per rack.
As Nvidia prepares for the rollout of the Vera Rubin, the combination of soaring component prices and changing procurement strategies is set to reshape AI server production. Investors and industry stakeholders should closely monitor these developments as they unfold.