The $9.1 Million AI Rack | Generated by AI

Home 2026.06

Question: Why will NVIDIA’s Rubin AI rack (NVL72) be so expensive, with estimates reaching up to $9.1 million per cabinet?

Answer: Recent 2026 financial and supply chain reports from investment firms like Bernstein Research and Morgan Stanley have broken down the staggering cost of NVIDIA’s next-generation Vera Rubin NVL72 server rack (referred to as a “柜” or cabinet). While Morgan Stanley initially estimated the cost at around $7.8 million, Bernstein’s updated data projects it could reach $9.1 million per rack.

The price is driving a massive spike compared to the previous Blackwell generation (which costs sub-$4 million) due to several severe hardware and supply chain constraints:

1. Massive Surges in Memory and Storage Costs (~$3.2 Million)

Memory is the single largest factor driving up the price, seeing a staggering 435% to 485% increase in cost compared to older architectures.

2. High Density of Premium Processors (~$4.14 Million)

3. Sophisticated Networking and Infrastructure (~$1.27 Million)

To connect 72 GPUs and keep data flowing without bottlenecks, the architectural infrastructure has become exponentially more complex:

4. Advanced Materials and Engineering Limits (~$500,000)

While a single cabinet costs nearly $9.1 million, cloud giants continue to buy them because the sheer compute efficiency and bandwidth per dollar still heavily outperform previous architectures for training and deploying frontier AI models.

References:


Back Donate