Nvidia's latest Vera Rubin AI servers are commanding extraordinarily high prices, with Vera Rubin-based NVL72 VR200 systems currently quoted at USD 5 million to USD 7 million per unit, and future Vera Rubin Ultra configurations potentially reaching USD 7 million to USD 8.8 million. The astronomical price reflects both unprecedented compute density and a fundamental shift in how Nvidia structures its supply chain, one that is reshaping the economics for the companies that actually assemble these systems.
The issue for equipment manufacturers is straightforward: margins are shrinking across the board. Not only is it difficult to maintain a 10 percent margin on items that cost millions, Nvidia has also reduced the role of server makers and systems integrators in the final bill of materials. If a Rubin NVL72 costs around USD 3 million, a 10 percent margin yields only USD 300,000 in profit per rack, which hyperscalers have shown they're unwilling to pay.
What the data actually tells us is that Nvidia is progressively absorbing the assembly work that once belonged to equipment manufacturers. Under the reported shift, Nvidia would supply partners with fully assembled Level-10 compute trays complete with compute hardware, cooling systems, and interfaces, leaving major ODMs with little design or integration work. If adopted, partners would be left primarily with rack-level integration rather than full server design, since these compute trays likely represent around 90 percent of a server's cost. Partners would still construct the outer chassis, integrate power supplies, and handle final assembly and testing, but these tasks offer little hardware differentiation.
The commercial logic is clear from Nvidia's perspective. Vera Rubin has seen immense interest from hyperscalers, and almost all leading entities including OpenAI, Amazon, Microsoft, and others have placed orders. CEO Jensen Huang projected USD 1 trillion in revenue from Blackwell and Rubin between 2025 and 2027, indicating unprecedented demand. When demand vastly exceeds supply and customers are desperate to gain competitive advantage, the supplier's negotiating position becomes nearly absolute.
Yet the human cost of this consolidation deserves consideration. The changes within rack architectures, including modularity, liquid-cooling, and advanced power delivery systems, have forced server manufacturers to invest heavily in research and development, meaning ODMs are not in the best of positions despite Vera Rubin demand. Equipment makers are investing significantly in engineering capability at precisely the moment their value capture is narrowing. Future AI data centers will be complex engineering systems integrating power, cooling, and signal transmission, with their construction costs and operation and maintenance difficulties increasing significantly. After Nvidia switches to its Vera Rubin platform next year, there will be a new cycle driven by iterated designs.
The counterargument worth acknowledging is that acceleration benefits everyone. The purported change could accelerate the VR200 ramp, since partners would no longer need to complete all design work themselves, and production costs could decrease. Shorter timelines to deployment mean customers get access to cutting-edge hardware faster, and hyperscalers who are burning capital on the AI race gain systems sooner. There is genuine value in speed.
However, this dynamic reflects a broader pattern: the company with the irreplaceable technology sets the terms, and the companies that add value through systems integration become cost centres rather than partners. Nvidia has legitimate reason to consolidate control; vertical integration can improve reliability and reduce variability. But the squeeze on ODM margins raises a question about sustainability. When profit margins on high-complexity manufacturing are compressed to near-zero percentages, the incentive to innovate in system design evaporates.
Vera Rubin is already in full production as of Q1 2026, with partner availability still scheduled for the second half of 2026. By the time equipment makers are formally selling these systems, the commercial architecture may already be locked in. The question for the manufacturing ecosystem is whether pre-assembled components from Nvidia become the standard, or whether hyperscalers eventually demand customisation options valuable enough to justify margin recovery. That negotiation has not yet been publicly resolved.