NVIDIA Data Center Costs Surge: Is the Investment Still Worth It for Tech Giants?

The cost of AI data centers equipped with NVIDIA's latest Vera Rubin architecture is far exceeding previous market expectations. According to the latest calculations by Bernstein Research, the cost of a single Vera Rubin NVL72 rack is as high as approximately $9.1 million, and overall data center Capex has climbed to about $47 billion per gigawatt— yet at the same time, the continuous improvement in cost-performance ratio of computing power makes this huge investment still economically reasonable for tech giants.

According to Zhuifeng Trading Desk, Bernstein analysts including Stacy A. Rasgon pointed out in a research report released on June 8 that the previously widely circulated quote of "about $8 million per rack" was based on outdated memory prices and seriously underestimated the actual cost. The core divergence lies in High Bandwidth Memory (HBM): the current price of HBM 4 is about $16.6 per GB, but it is expected to rise to about $53 per GB by the time Vera Rubin ships in large volumes in 2027, and NVIDIA is likely to pass these costs on to end customers through dynamic pricing mechanisms.

Rising costs have not shaken Bernstein's bullish stance on NVIDIA. The firm maintains its "Outperform" rating on NVIDIA with a target price of $315. The report also points out that the FP8 computing power of the Vera Rubin NVL72 rack reaches 2,520 petaflops (PFLOPS) per second, a significant jump from the 720 PFLOPS of the previous generation Blackwell, significantly enhancing the computing power obtained per gigawatt and per dollar, which is expected to further promote the adoption of AI applications.

Rack Cost: The Underlying Logic of the $9.1 Million Figure

Bernstein used a bottom-up approach to break down the Vera Rubin NVL72 rack item by item, ultimately arriving at a cost estimate of approximately $9.1 million.

GPUs remain the largest single cost item. The report shows that the Rubin GPU sells for about $55,000 each, with 72 GPUs per rack, meaning the GPU cost alone reaches $3.96 million, accounting for nearly half of the total rack cost. In addition, the 36 Vera CPUs per rack contribute about $180,000.

The significant increase in memory and storage costs is the main source of divergence between this estimate and market expectations. Bernstein expects this cost to be about $3.2 million, far higher than the approximately $2 million calculated using historical prices. Of this, HBM 4 contributes about $1.09 million, CPU DRAM (LPDDR5X) about $800,000, and direct-attached storage about $1.28 million. The report specifically warns that memory and storage prices are highly volatile—NAND prices have increased 11.3-fold from the low point in April 2023 to May 2026, with an annualized increase of 115%, so investors need to continuously track price changes to maintain forecast accuracy.

Networking, cooling, and power supply collectively contribute about $2 million. Of this, networking costs are about $1.27 million, including about $250,000 for NVLink switches, about $240,000 for cables, about $380,000 for backplanes and other scale-out components, and about $200,000 for SpectrumX switches; cooling costs are about $160,000, and power supply costs are about $150,000.

$47 Billion Per Gigawatt: The Real Bill for Full-Stack Data Centers

When extrapolating from single-rack costs to overall data center Capex, the figures are even more staggering.

The Vera Rubin NVL72 rack has a rated power consumption of 220 kilowatts. Bernstein estimates that rack power consumption accounts for about 78% of the total electricity usage in a data center, implying that each gigawatt can accommodate about 3,557 racks, corresponding to a rack cost of about $32.3 billion. Adding about $15 billion in physical infrastructure costs per gigawatt (including mechanical and electrical equipment and land/buildings), the Capex for a full-stack AI data center is about $47.3 billion per gigawatt, an increase of about 17% compared to the approximately $40.5 billion in the previous Blackwell cycle.

The operating cost structure is also worth noting. The report points out that even calculating at a relatively high electricity price of $0.15 per kilowatt-hour, the annual electricity bill for operating a one-gigawatt data center is about $1.3 billion; personnel costs are almost negligible, with the largest data centers requiring only 8 to 10 employees. In contrast, calculated on a 6-year depreciation cycle, the annual depreciation cost reaches about $7.9 billion, constituting the main component of operating costs. Since the depreciation period for IT hardware (servers, network equipment) (4 to 6 years) is much shorter than that for mechanical and electrical equipment and land/buildings, the weight of servers and networks in the true economic cost is actually higher than the proportion presented by cash Capex.

Computing Power Leap: The Hedging Logic Against Rising Costs

Although Capex per gigawatt continues to climb, the improvement in computing power cost-performance provides an economic basis for this investment.

Bernstein data shows that the FP8 computing power of the Vera Rubin NVL72 rack is 2,520 petaflops per second, which is 3.5 times that of Blackwell (720 PFLOPS). Converted to the per-gigawatt dimension, Vera Rubin can provide about 8,960 EFLOPS of FP16 sparse computing power, more than double Blackwell's 4,269 EFLOPS; the computing power corresponding to every $1 billion in Capex has also increased from 105.5 EFLOPS to 189.3 EFLOPS.

The report also points out that in an environment of tight computing power, data center operators tend to extend the lifespan of GPUs as much as possible and prioritize deploying new-generation GPUs with newly built capacity. If new construction is not possible due to power or physical infrastructure constraints, operators may need to consider decommissioning old GPUs to free up space for new chips.

Acceleration on the AI demand side also supports continued investment. Citing data, the report states that Anthropic's annualized revenue has soared from $9 billion at the end of 2025 to $47 billion in May 2026, and the company stated that it had to actively give up some customers and revenue due to computing power constraints.

Supply Chain Impact: Who Wins, Who Bears the Pressure?

Changes in the cost structure are reshaping the beneficiary landscape of the AI supply chain.

Memory is the biggest structural beneficiary. CPU DRAM specifications have increased by 320% (in TB terms) compared to the previous generation, far exceeding the approximately 50% increase in HBM and NAND. Bernstein also mentioned that the application of CXL memory for KV caching is increasing, and if supply allows, DRAM is expected to benefit disproportionately.

Demand for power supply components continues to expand. The report shows that the proportion of power supply content in rack costs has increased from about 1.0% in the previous generation to about 1.6%, and the early adoption of 800VDC solutions has further promoted this trend. Bernstein maintains its "Outperform" rating on Delta Electronics with a new target price of NT$2,620, viewing it as one of the main beneficiaries of the growth in power supply content.

Regarding substrates, Bernstein believes that demand for ABF substrates will continue to grow, holding a positive view on Ibiden and Unimicron, with the latter having a target price of NT$990.

In contrast, Bernstein maintains an "Underperform" rating on CoreWeave with a target price of $67; and an "Underperform" rating on Quanta with a target price of NT$250.

Costs Will Continue to Rise, Power Demand Lags Behind Capex

Looking ahead, Bernstein expects costs per gigawatt to continue rising, but the growth in power demand will lag behind the expansion pace of Capex by hyperscale cloud providers.

The report points out that the cost increase per gigawatt in the Rubin cycle is about 9%, slightly higher than the 8% in the Blackwell cycle. Market consensus expectations show that Capex by hyperscale cloud providers and emerging cloud service providers will increase by 69% year-on-year in 2026, with the growth rate slowing to about 13% in 2027, meaning that relatively stable additions in power capacity can support continued growth. Bernstein believes that this apparent contradiction may precisely indicate that there is still upside potential in market expectations for hyperscale Capex.

Notably, potential supply shortages of LPDDR memory may constrain the shipment of Vera Rubin and standalone Vera CPUs. NVIDIA may choose to ship with lower default memory configurations, allowing customers to expand capacity later, so that customers can flexibly decide on the optimal configuration based on the latest memory prices.