What happened
Bernstein published a research report on October 5 providing a detailed cost breakdown for building 1GW AI data centers across Nvidia, AMD, Google, and OpenAI architectures. The report estimates total capital expenditure ranges from $34.6 billion to $39.5 billion per GW, with depreciation identified as the primary financial burden.
Bernstein released a report on October 5 analyzing the capital expenditure required to build 1GW AI data centers. The firm estimates costs range from $34.6 billion for OpenAI's Jalapeno architecture to $39.5 billion for Nvidia's Vera Rubin architecture. The report also revised down the cost of a single Nvidia Rubin NVL72 rack from $9.1 million to $7.52 million, a 17% reduction driven by lower HBM pricing assumptions.
The analysis compares hardware costs across four major architectures: Nvidia Rubin, AMD Helios, Google TPU v7, and OpenAI Jalapeno. While Google's TPU v7 single-system cost is less than 40% of Rubin's, the total investment per GW differs by only about $500 million. This is because TPU systems have lower power density, requiring nearly three times as many units to achieve the same 1GW capacity, which offsets the lower per-unit cost.
Bernstein identifies depreciation as the most significant financial burden. For a 1GW data center using Rubin architecture with a $39.5 billion investment, annual depreciation over six years is approximately $6.6 billion. This far exceeds the estimated annual electricity cost of $1.3 billion, assuming electricity at $0.15 per kilowatt-hour. The report notes that depreciation policies vary by company, with Google using six years and Meta using four to five years for server equipment.
The report projects global AI equipment shipments to increase power capacity from 27GW in 2025 to 43.3GW in 2027. It notes that this global growth exceeds the projected addition of new data center capacity in the United States, suggesting a shift of AI infrastructure construction overseas due to U.S. power supply constraints and policy restrictions.
Source details: finance.biggo.com ↗
Why it matters
This analysis clarifies the true economic scale of AI infrastructure, revealing that while individual chip costs vary significantly, total data center costs are driven more by power density and hardware quantity than by chip price alone. It highlights that depreciation, not electricity, is the largest ongoing cost, which has significant implications for the long-term financial viability and investment returns of AI labs and cloud providers.
The report demonstrates that the choice of accelerator architecture has a smaller impact on total data center cost than previously assumed, as power density and the resulting quantity of hardware required are the dominant cost drivers. This challenges the notion that cheaper chips automatically lead to significantly cheaper infrastructure.
By highlighting that depreciation costs are several times higher than electricity costs, the report shifts the focus of AI infrastructure economics from energy efficiency to hardware lifespan and capital allocation. This has direct implications for how AI companies model their long-term profitability and investment returns.
The projection that global AI data center capacity will grow faster than U.S. capacity suggests a structural shift in the geography of AI infrastructure. This could impact supply chains, regulatory environments, and the competitive dynamics between domestic and international AI deployments.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
What should a useful AI forecast state?
What to watch next
Monitor how these cost estimates influence future capital expenditure decisions by major tech companies and whether the projected shift of AI data center construction overseas continues to accelerate due to U.S. power and policy constraints.
Track whether major AI labs and cloud providers adjust their capital expenditure plans based on these cost estimates, particularly regarding the balance between Nvidia GPUs and custom ASICs.
Observe the pace of AI data center construction in non-U.S. markets, such as Australia, Singapore, and Southeast Asia, to see if the projected shift in infrastructure location materializes.
Monitor changes in HBM and DRAM pricing, as Bernstein's cost models are sensitive to these component costs, and any further price fluctuations could alter the relative economics of different architectures.