NVIDIA's next-generation AI server rack prices are experiencing rapid escalation. According to a report published on May 22, the upcoming Vera Rubin (VR200) rack, procured from ODM manufacturers, is priced at approximately $7.8 million. In contrast, the current GB300 Blackwell rack costs less than $4 million, indicating a near doubling of the price per rack within a single product generation. It is noteworthy that following the company's earnings release on that day, NVIDIA's stock closed down nearly 2%, while memory-related stocks surged 6% to 10%. This divergence highlights that the beneficiaries of these price increases extend beyond NVIDIA itself.
Memory has transitioned from a supporting role to a primary driver of cost inflation. Analysts point out that since the introduction of the GB200 NVL72, memory prices have risen substantially. Under the previous cost structure, memory accounted for only 5% to 10% of the GB200 NVL72 rack's bill of materials (BOM). However, for the VR200, this share has skyrocketed to 25% to 30%, representing a cost increase of 435%. This surge in memory's proportion directly reduces the GPU's share of the total rack cost, which has declined from approximately 65% in the GB200 to about 51% in the VR200.
The price increases are not limited to memory; they are widespread across nearly all components. Supply chain investigations indicate comprehensive inflation: PCB (Printed Circuit Board) costs have risen by 233%, jumping from about $35,000 for the GB300 to roughly $117,000 for the VR200. This increase is attributed to new modules introduced in Rubin (such as ConnectX modules and midplane PCBs), alongside upgrades in circuit board layer counts and material grades. For instance, the compute board has been upgraded from a 22-layer HDI PCB in the GB300 to a 26-layer board, with material grade improving from M7 to M8. The switch tray PCB has increased from 24 to 32 layers. Additionally, a new 44-layer midplane PCB has been added to the compute tray, a component absent in the GB300. MLCC (Multi-Layer Ceramic Capacitor) costs have increased by 182%, with content per VR200 rack reaching about $4,300 compared to $1,500 for the GB300. Newly introduced BlueField and ConnectX modules also contribute to additional MLCC demand. Analysts believe this is a direct reason for the currently robust demand for high-end AI server MLCCs and ODM manufacturers actively stockpiling, as the Rubin rack is expected to enter mass production in the second half of 2026. ABF substrate costs have risen by 82%. The VR200 incorporates twice the number of NVLink and ConnectX chips compared to the Blackwell system, leading to increased substrate usage. The ABF substrate unit price for the Rubin GPU is approximately $200 per chip, a 100% increase from Blackwell's roughly $100. Power costs have increased by 32%, and cooling materials by 12%.
A key variable identified is the procurement method for SOCAMM (Small Outline Compression Attached Memory Module). If hyperscale cloud vendors directly purchase memory modules, the rack price could drop to approximately $6.7 million. In the baseline scenario, where NVIDIA procures and resells SOCAMM at a 70% gross margin, the rack price is around $7.8 million. However, if hyperscale cloud vendors (such as Microsoft, Google, Amazon.com) choose to source SOCAMM directly, bypassing NVIDIA's markup, the rack price would decrease to about $6.7 million. This variable directly impacts NVIDIA's potential revenue from memory sales and is a critical point for the market to monitor.
Regarding ODM manufacturers, while market expectations suggested their value-add might shrink due to more "standardized" compute tray designs in the Rubin rack, the analysis presents a contrary conclusion. The report estimates that the ODM value-add portion will increase by 35% to 40%, rising from about $108,000 per rack for the GB300 to approximately $149,600 for the VR200. This assessment aligns with statements from Wistron's management during their Q4 2025 earnings call, where they explicitly stated that the dollar value-add for Rubin racks would increase. However, gross margins are indeed declining: the ODM gross margin for the GB300 was about 2.7%, dropping to around 1.9% for the VR200. Analysts suggest that "investors should focus on the growth in absolute dollar profit, rather than the decline in the profit margin."
A concerning trend highlighted is the quiet spread of the consignment business model among ODM manufacturers. Hon Hai was the first to mention this model during their Q4 2025 earnings call. Subsequently, Quanta indicated during their Q1 2026 earnings call that they expect some projects to transition to a consignment model in the second half of 2026. The consignment model involves the customer (cloud vendor) procuring core components themselves, with the ODM responsible only for assembly. This alleviates the ODM's working capital pressure but also compresses their revenue scale. Analysts note that it remains unclear what proportion of projects will shift to consignment, but if the aforementioned "absolute profit growth" does not materialize, this trend poses long-term risks.
Looking at power and cooling, the next upgrade direction, supply chain checks show the Vera Rubin platform comes standard with a 110kW power module. However, at least one US cloud service provider has adopted an HVDC (High Voltage Direct Current) independent power rack for the Vera Rubin platform. Analysts anticipate that 800V DC power will see widespread adoption in the Rubin Ultra platform expected in the second half of 2027. Delta has partnered with at least three US cloud service providers to advance HVDC platform deployment, with initial rollouts expected to begin in the second half of 2026. For cooling, the Vera Rubin rack will feature a fully liquid-cooled design (fanless), with the total value of liquid cooling components per rack estimated at approximately $72,080.
Comments