NVIDIA Vera Rubin AI Racks Could Cost $9.1 Million as HBM4 Memory Prices Surge, Bernstein Says
Memory pricing may become one of the biggest cost pressures in the next phase of the AI hardware boom. Investment bank Bernstein believes prices for high bandwidth memory, especially next-generation HBM4, could climb sharply as NVIDIA’s Vera Rubin AI platform begins shipping in large volumes.
According to Bernstein, HBM4 memory could reach around $53 per gigabyte in 2027. That forecast is far higher than some earlier assumptions and suggests that the cost of building advanced AI server racks may rise significantly as demand for AI accelerators continues to outpace supply.
The firm’s estimate centers on NVIDIA’s upcoming Vera Rubin NVL72 rack, a next-generation AI system expected to succeed current Blackwell-based infrastructure. Earlier estimates from Morgan Stanley suggested that a single Vera Rubin NVL72 rack could cost roughly $7.8 million, with memory and storage accounting for about $2 million of that total.
Bernstein, however, believes that figure may be too low. The firm argues that the previous estimate appears to rely on outdated memory pricing and does not fully reflect where HBM4 costs are heading. Using its updated assumptions, Bernstein estimates that a Vera Rubin NVL72 rack could cost closer to $9.1 million.
The key difference is the price of HBM4. Bernstein says older calculations appear to assume HBM4 pricing of around $16.6 per gigabyte, a level the firm no longer views as realistic for 2027. Instead, it expects pricing to rise to about $53 per gigabyte when Vera Rubin systems are shipping at scale.
That shift would have a major impact on the total bill of materials for NVIDIA’s AI racks. Bernstein estimates that memory and storage alone could cost around $3.2 million per rack, compared with Morgan Stanley’s earlier estimate of roughly $2 million.
The reason behind the expected jump is simple: AI systems are becoming increasingly memory-hungry. As large language models and advanced AI workloads grow, GPUs require faster and larger memory pools to keep performance high. High bandwidth memory is essential because it allows AI accelerators to move massive amounts of data quickly, reducing bottlenecks during training and inference.
For NVIDIA, HBM has become just as critical as the GPU itself. The company’s most advanced AI platforms depend not only on cutting-edge chips but also on complex packaging and high-capacity memory stacks. As demand rises, memory supply can become a major constraint, pushing prices higher.
Bernstein also believes NVIDIA is likely to pass these increased memory costs on to customers rather than absorbing them. That means cloud providers, hyperscalers, AI startups, and enterprise customers could face higher infrastructure costs when deploying Vera Rubin-based systems.
The broader implication is that the AI hardware market may remain extremely expensive through 2027. Even as companies race to build bigger AI clusters, supply chain pressures around HBM4 memory, advanced packaging, and GPU production could keep prices elevated.
NVIDIA’s Vera Rubin generation is expected to play a major role in the next wave of AI data center expansion. If Bernstein’s forecast proves accurate, the cost of deploying top-tier AI compute could climb even higher, reinforcing NVIDIA’s pricing power while also increasing capital spending requirements for companies competing in artificial intelligence.
In short, Bernstein sees memory as a major driver of future AI system costs. With HBM4 demand expected to surge and NVIDIA’s Vera Rubin racks likely to command premium pricing, the next generation of AI infrastructure may become more powerful, but also far more expensive.
