Bernstein: Vera Rubin NVL72 Rack Cost Reaches $9.1 Million
nashnova research
Bernstein has revised its cost estimate for Nvidia's next-generation Vera Rubin NVL72 rack to $9.1 million, roughly 14% above the ~$8 million widely cited in media — the gap traces almost entirely to HBM memory repricing, signaling that the true cost of AI compute is being rewritten by memory.
Where does $9.1 million come from — and why is it $1.1 million more?
Earlier media estimates used a historical HBM price of ~$16.6/GB. Bernstein expects HBM4 to reach $53/GB by the time Vera Rubin ships at scale in 2027, including Nvidia's ~10% cost pass-through markup.
This means → memory/storage cost alone jumps from ~$2 million to ~$3.2 million — nearly doubling.
In plain terms = the chips are the same chips, but the memory that feeds them got far more expensive, dragging the whole rack up with it.
Inside the rack — where does the money go?
GPUs are the single largest item: each Rubin GPU costs ~$55,000; 72 per rack totals ~$3.96 million, over 40% of the rack.
Memory/storage follows closely: ~$3.2 million, comprising HBM4, CPU DRAM (LPDDR5X, 54 TB, ~$802,000 at a ~30% premium to mobile DRAM contract prices), and direct-attached NAND (~3.5 PB, ~$1.3 million).
Networking runs ~$1.2 million: NVLink switches ~$250K, cables ~$240K, backplane and vertical scale-up ~$380K, SpectrumX switches ~$200K. Cooling and power each cost ~$150–160K, notably higher than Blackwell.
What does a 1 GW AI data center cost end-to-end?
Each rack draws 220 kW; racks consume ~80% of a data center's power. Bernstein estimates ~3,600 racks per GW, putting rack costs at ~$32 billion.
Add ~$15 billion per GW for physical infrastructure (buildings, power, cooling), and total capex reaches ~$47 billion per GW.
This means → electricity is not the dominant cost. Even at $0.15/kWh, annual power for a 1 GW facility is ~$1.3 billion, while annual depreciation is ~$7.2 billion. In plain terms = the real burn is not the electricity bill — it is how fast servers and networking gear lose value.
Why is memory pricing the biggest wild card?
NAND prices have surged 11.3× from the April 2023 trough through May 2026 (115% CAGR), a sharp reversal from the −20% annual trend between 2019 and 2023.
Bernstein believes Nvidia may have dynamic pricing mechanisms that pass memory cost swings through to end customers rather than absorbing them.
This means → investors must continuously track DRAM and NAND price moves — the rack cost breakdown itself goes stale fast as memory prices shift.
After all the cost increases — is compute efficiency actually improving?
The Vera Rubin NVL72 rack delivers 2,520 PFLOPS, versus 720 PFLOPS for Blackwell — roughly a 3.5× leap.
This means → whether measured per GW or per dollar, compute capacity is accelerating sharply, and FP8 performance per dollar continues to improve.
Bernstein expects per-GW costs to keep rising, with power demand growth lagging hyperscaler capex growth. This reflects a critical question: compute supply is surging, but whether AI adoption demand can keep pace will be the key test for the entire infrastructure investment thesis.
市场有风险，内容仅供研究参考，不构成投资建议。