Bernstein: Vera Rubin Rack Cost Rises to $9.1M; Capex per GW Data Center Reaches $47B

Overview

Published by Bernstein, this report aims to update market cost estimates for rack-level economics in NVIDIA's next-generation Vera Rubin (VR) NVL72 architecture AI data centers. Through interviews with industry experts and cross-validation with third-party data, the report indicates that the widely cited market figure of ~$8M/rack is based on outdated memory pricing, whereas actual costs have risen to ~$9.1M due to HBM 4 price increases. Consequently, the fully loaded capital expenditure (Capex) per GW of capacity is derived to be approximately $47B. Despite higher unit costs, the increased compute density of the Rubin architecture significantly accelerates performance per dollar. The report also details changes in cost composition across the supply chain and provides investment ratings for relevant stocks.

Core views

Core View 1: Vera Rubin rack costs are underestimated, with HBM being the primary source of variance. Using a bottom-up model, the report estimates Vera Rubin NVL72 rack costs at approximately $9.1M, significantly higher than the widely reported $8M. This difference mainly stems from High Bandwidth Memory (HBM) pricing assumptions. Media models largely adopt historical HBM 4 prices (~$16.6/GB), but the report expects HBM 4 prices to rise to $48/GB as Vera Rubin ships in volume in 2027. Adding NVIDIA's potential dynamic pricing mechanism and an approximate 10% markup, the final BOM price paid by customers could reach ~$53/GB. This would surge HBM's cost contribution per rack from $344k to $1.1M, making it the largest driver of cost overruns. Additionally, CPU DRAM (LPDDR5X) prices exceed mobile DRAM due to SOCAMM architecture premiums and potential shortage risks, further pushing total memory costs to $3.2M/rack (35% share), well above the $2M implied by historical prices. Core View 2: GPUs remain the largest cost item; networking and cooling shares remain stable. Even with sharply rising memory costs, GPUs (excluding HBM) remain the largest component of rack costs, accounting for approximately $4M (72 Rubin GPUs at ~$55k each). CPU costs are approx. $180k (36 units at ~$5k each). Networking accounts for roughly 13% of rack costs (~$1.2M), with Scale-up content like NVLink switches, cables, and backplanes taking a large share, while SpectrumX switches account for about half of the Scale-out fabric. Cooling and power delivery costs also increase significantly to approx. $150-160k/rack, a notable rise from the GB200 era. Core View 3: Capex per GW data center reaches $47B; compute efficiency accelerates. Based on a Vera Rubin rack rated power of 220kW and racks consuming ~80% of data center power, each GW can support approximately 3,557 racks, corresponding to total rack costs of ~$32.3B. Adding ~$1.5B/GW for physical infrastructure (land, buildings, MEP), the fully loaded AI data center Capex per GW is approximately $47B. Notably, despite higher absolute costs, Rubin NVL72 rack FP8 performance reaches 2,520 PLOPS, nearly 3.5x Blackwell's 720 PLOPS. This implies meaningful acceleration in compute capacity on both a per-GW and per-dollar basis, helping alleviate compute bottlenecks and driving further AI adoption. Core View 4: DRAM and power content share rises; TCO shifts toward hardware depreciation. The report notes that compared to the Blackwell cycle, Rubin cycle costs per GW are expected to increase by 9%. Power content share rises from 1.0% to 1.6%, while DRAM content (TB-scale) surges 320%, far exceeding the 50% growth for NAND and HBM. Since IT hardware (servers, networking) depreciation lifespans (typically 4-6 years) are shorter than MEP or buildings (10-40 years), and operating electricity (~$1.3B/GW/year) and labor costs are relatively low, the true economic cost (TCO) is more heavily weighted towards servers, storage, and networking hardware than cash Capex. Annual depreciation (~$7.2-7.9B/GW) is the dominant operating cost.

Analysis framework

The report employs an analytical approach combining bottom-up BOM (Bill of Materials) teardowns with macro capacity extrapolation. First, at the micro level, the report conducts detailed cost breakdowns for individual racks of the Vera Rubin NVL72 architecture. By referencing NVIDIA public specifications, datasheets from vendors like Super Micro, and channel checks, it estimates quantities and unit prices for GPUs, CPUs, HBM, DRAM, NAND storage, networking equipment (NVLink/SpectrumX), cooling, power, and other components. Specifically, the report introduces dynamic pricing assumptions, considering NVIDIA's motivation to protect margins by passing upstream component (e.g., HBM) price volatility to downstream customers, thereby correcting biases from static historical pricing. Second, at the macro level, using single-rack power density (220kW) and data center PUE assumptions, the report extrapolates the number of racks supported per GW of power capacity. Combined with previous estimates for top-tier facility physical infrastructure costs (~$15B/GW) and aggregated rack costs, it derives fully loaded Capex per GW. This 'single-rack cost × rack density + fixed infrastructure cost' method more accurately reflects capital expenditure pressure in large-scale deployments. Finally, by comparing performance metrics (PLOPS) and costs across generations (Hopper, Blackwell, Vera Rubin), the report calculates compute efficiency per dollar to assess the economics of technological iteration. Simultaneously, combining depreciation lifespan analysis distinguishes cash Capex from economic cost (TCO), revealing the impact of rapid hardware iteration on long-term operating cost structures.

Methodology notes

• Industry Analysis FrameworkPrice-Volume Decomposition

  BOM Cost Breakdown & Price Volume Analysis

  The report decomposes AI server racks into sub-components like GPUs, memory, and networking, estimating usage (volume) and unit price (price) separately. This method precisely identifies cost drivers, e.g., finding that while GPU volume remains constant, rising HBM unit prices significantly increase total memory costs.

• Industry Analysis FrameworkSupply Chain Pass-through

  Supply Chain Pass-through Mechanism

  The report assumes NVIDIA possesses dynamic pricing power to pass upstream HBM and DRAM cost increases to end customers rather than absorbing them to maintain gross margins. This reflects the bargaining power and cost pass-through logic of dominant brands in the supply chain.

• Company Fundamentals & Financial FrameworkFree cash flow analysis

  Capex to Depreciation & TCO Analysis

  Beyond initial Capex, the report compares shorter hardware depreciation lifespans with longer physical facility lives, noting that in true economic cost (TCO), server and networking depreciation shares far exceed cash expenditure shares. This aids understanding of long-term cash flow pressures for data center operators.

• Valuation MethodologyPE/PEG valuation

  Performance per Dollar Efficiency

  By calculating FP8 compute power (PLOPS) per GW or per dollar of Capex, the report evaluates the cost-effectiveness of tech iterations. This 'compute per unit cost' metric serves as a core non-financial valuation aid for measuring AI hardware ROI.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVIDIA (NVDA)

  Core beneficiary; transfers HBM costs via dynamic pricing, maintaining high margins

  Strengths

  Strong pricing power, significant Rubin architecture performance uplift, high ecosystem moat

  Weaknesses

  Reliance on upstream HBM supply stability

  Comparison

  Superior to other GPU makers due to complete system-level pricing power

  Risks

  Antitrust regulation, HBM supply shortages limiting shipments

• Delta Electronics (2308.TT)

  Beneficiary; significantly higher power and cooling content share

  Strengths

  Leading share in AI server power and liquid cooling, high technical barriers

  Weaknesses

  Facing share competition from Vertiv and others

  Comparison

  Superior to traditional power suppliers due to deep integration with NVIDIA reference designs

  Risks

  Intensified competition pressuring gross margins

• Unimicron (3037.TT)

  Beneficiary; ABF substrate demand grows with GPU/CPU complexity

  Strengths

  Core ABF substrate supplier, scarce capacity

  Weaknesses

  High cyclicality, heavy capital expenditure

  Comparison

  Duopoly with Ibiden, high certainty of benefit

  Risks

  Downstream demand slowdown leading to lower utilization rates

• CoreWeave (CRWV)

  Negatively impacted/Stressed; high Capex and depreciation pressure affects profitability

  Strengths

  Focused on AI compute leasing, strong customer stickiness

  Weaknesses

  Asset-heavy, high depreciation costs, significant financing pressure

  Comparison

  Disadvantaged in scale effects and financing costs vs. hyperscalers

  Risks

  Compute price wars, rising interest rates increasing financing costs

• Quanta (2382.TT)

  Stressed; thin assembly margins squeezed by upstream costs

  Strengths

  Large scale, stable customer relationships

  Weaknesses

  Low margins, sensitive to upstream component price volatility

  Comparison

  Share in high-end AI servers may be limited vs. Wistron/Foxconn

  Risks

  Order loss to other ODMs, further margin compression

Key data

• Vera Rubin NVL72 Single Rack Cost~$9.1MSignificantly above media expectation of ~$8M, primarily due to HBM price hikes
• Expected HBM 4 Price (2027)~$53/GBIncludes ~10% premium, far above current historical reference of ~$16.6/GB
• Fully Loaded Capex per GW AI Data Center~$47BIncludes ~$32B rack costs + ~$15B physical infrastructure costs
• Vera Rubin NVL72 FP8 Performance2,520 PLOPS~3.5x increase over Blackwell's 720 PLOPS
• Single Rack Memory/Storage Cost~$3.2M~35% of total rack cost, with HBM contributing ~$1.1M
• Single Rack GPU Cost (Excl. HBM)~$4M72 GPUs x $55k/unit, remains largest cost item
• Annual Electricity Cost per GW~$1.3BAssumes electricity price of $0.15/kWh

Impact & implications

For hardware suppliers, Vera Rubin's high-cost and high-power characteristics directly benefit high-value component suppliers. Delta Electronics, as a primary beneficiary of power and cooling solutions, sees increased content value; Unimicron and Ibiden benefit from growing ABF substrate demand; NVIDIA successfully transfers memory cost pressures via dynamic pricing, maintaining profitability. For data center operators (e.g., Digital Realty, Equinix), Capex as high as $47B per GW implies massive capital barriers, potentially increasing industry concentration as leaders consolidate positions via financing advantages and economies of scale. For cloud compute lessors (e.g., CoreWeave), high hardware depreciation and power costs pose severe profitability challenges, especially if they cannot effectively pass on costs, which explains the 'Underperform' rating. Overall, while accelerated compute supply helps unlock more AI applications, increased capital intensity in infrastructure build-out may cause near-term capacity bottlenecks and cash flow strain.

Risks

• Severe volatility in HBM and DRAM prices; if prices fall faster than expected, inventory impairments or pricing strategy failures may occur
• If NVIDIA's dynamic pricing mechanism is not accepted by the market, its market share or margins could be impacted
• Shortages of LPDDR5X or other key components could limit Vera Rubin shipments
• Data center construction constrained by power supply or physical infrastructure bottlenecks, preventing full demand realization
• Macroeconomic downturn leading to cloud provider Capex cuts, impacting demand across the AI hardware supply chain

What to watch

• Spot and contract price trends for HBM 4 and LPDDR5X, especially changes from H2 2026 to 2027
• Actual shipment timeline and customer adoption of NVIDIA Vera Rubin
• Whether major hyperscalers raise 2027 Capex guidance to match increased per-GW costs
• Gross margin trends for key component suppliers like Delta Electronics and Unimicron to verify cost pass-through
• Difficulty in securing data center power and policy changes, potentially becoming new bottlenecks for expansion