Nvidia’s data center costs are soaring—does this investment still make sense for tech giants?

AI data centers equipped with Nvidia's latest generation Vera Rubin architecture are costing far more than previously expected by the market. According to Bernstein Research's latest estimates, the cost of a single Vera Rubin NVL72 rack is as high as approximately $9.1 million, with total data center capital expenditures rising to about $4.7 billion per gigawatt—yet at the same time, the continued improvement in cost-to-performance means that this massive investment remains economically rational for tech giants.

According to Chasewind Trading Desk, Bernstein analysts Stacy A. Rasgon and others pointed out in a research report published on June 8, the previously widespread media quote of “about $8 million per rack” was based on outdated memory prices, seriously underestimating the actual cost. The core discrepancy lies in high-bandwidth memory (HBM): the current price of HBM4 is about $16.6 per GB, but it is expected to rise to about $53 per GB by 2027 when Vera Rubin is shipped in volume, and Nvidia will likely pass on this cost to end customers through dynamic pricing mechanisms.

The rising costs have not shaken Bernstein’s bullish stance towards Nvidia. The firm maintains an “outperform” rating for Nvidia, with a target price of $315. The report also points out that the FP8 computing power of a Vera Rubin NVL72 rack reaches 2,520 PFLOPS (petaflops) per second, a significant jump from the previous generation Blackwell’s 720 PFLOPS, with notable improvements in both computing power per gigawatt and per dollar, which is expected to further boost the adoption of AI applications.

Rack Cost: The Underlying Logic Behind $9.1 Million

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

GPUs remain the largest single cost item. According to the report, the price of a Rubin GPU is about $55,000 each, with 72 GPUs per rack, the GPU alone accounts for $3.96 million, nearly half the total rack cost. Additionally, each rack contains 36 Vera CPUs totaling about $180,000.

A significant increase in memory and storage costs is the main source of discrepancy between this estimate and market expectations. Bernstein expects these costs to be about $3.2 million, far higher than the roughly $2 million calculated using historical prices. Of this, HBM4 contributes about $1.09 million, CPU DRAM (LPDDR5X) about $800,000, and direct-attached storage about $1.28 million. The report especially notes that memory and storage prices are highly volatile—NAND prices have risen 11.3x from the low in April 2023 to May 2026, an annualized increase of 115%. Investors need to continuously monitor price changes to keep predictions accurate.

Networking, cooling, and power supply together contribute about $2 million. Network costs are about $1.27 million, including $250,000 for NVLink switches, $240,000 for cables, $380,000 for backplanes and other scale-expansion components, and $200,000 for SpectrumX switches; cooling is about $160,000, and power supply about $150,000.

$4.7 Billion per Gigawatt: The Real Bill of a Full-Stack Data Center

Extrapolating single rack costs to total data center capital expenditure yields even more staggering numbers.

The rated power consumption of a Vera Rubin NVL72 rack is 220 kW. Bernstein estimates that rack power usage accounts for about 78% of total data center electricity consumption, leading to an estimate of about 3,557 racks per gigawatt, corresponding to rack costs of about $3.23 billion. When adding about $1.5 billion per gigawatt for physical infrastructure (including mechanical/electrical equipment and land/buildings), full-stack AI data center capital expenditure is about $4.73 billion per gigawatt, an increase of about 17% over the previous Blackwell cycle’s approximately $4.05 billion.

Operational cost structure is also noteworthy. The report says that even with a high electricity price of $0.15 per kWh, the annual power bill for a 1 GW data center is about $1.3 billion; labor costs are almost negligible, with only 8-10 employees needed for the largest data centers. In contrast, calculated on a six-year depreciation cycle, annual depreciation costs are about $7.9 billion, making it the main component of operating costs. Because IT hardware (servers, network equipment) depreciates over 4-6 years, much shorter than mechanical/electrical equipment and real estate, the true economic weight of servers and networking exceeds their share of cash capital expenditure.

Performance Leap: Hedging Logic Against Rising Costs

Although capital expenditure per gigawatt continues to climb, improvements in the cost-to-performance ratio justify the investment economically.

Bernstein data shows that the FP8 computing power of Vera Rubin NVL72 racks is 2,520 PFLOPS per second, 3.5 times Blackwell (720 PFLOPS). Converted to gigawatt scale, Vera Rubin can provide about 8,960 EFLOPS of FP16 sparse computing, more than double Blackwell’s 4,269 EFLOPS; computing power per $1 billion capital expenditure also rises from 105.5 EFLOPS to 189.3 EFLOPS.

The report also notes that in a highly tight compute environment, data center operators tend to extend GPU lifespans as much as possible, and prioritize new capacity for deploying the latest generation GPUs. If unable to build new capacity due to power or physical infrastructure constraints, they may need to retire old GPUs to deploy new chips.

Accelerating demand on the AI side also supports continued investment. The report cites data showing that Anthropic’s annualized revenue has soared from $9 billion at the end of 2025 to $47 billion by May 2026, and the company says it is constrained by compute, having to voluntarily turn away some customers and revenue.

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

Changes in cost structure are reshaping the AI supply chain's benefit landscape.

Memory is the biggest structural beneficiary. CPU DRAM specs rose 320% over the previous generation (in TB), far exceeding the roughly 50% increase in HBM and NAND. Bernstein also mentions that applications of CXL memory for KV caching are increasing, and if supply allows, DRAM may gain disproportionately.

Demand for power supply components continues to expand. The report shows that power supply content as a percentage of rack cost has increased from about 1.0% to 1.6% with the early adoption of 800VDC schemes further promoting this trend. Bernstein maintains an “outperform” rating for Delta Electronics, with a target price of NT$2,620, regarding it as a major beneficiary of power supply content growth.

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

In contrast, Bernstein maintains an “underperform” rating for CoreWeave, with a target price of $67; and “underperform” for Quanta, with a target price of NT$250.

Costs Will Continue Rising, Power Demand Lags Behind Capital Expenditure

Looking ahead, Bernstein expects cost per gigawatt to keep growing, but the growth in power demand will lag the expansion pace of hyperscale cloud firms’ capital expenditure.

The report notes that the increase in cost per gigawatt during the Rubin cycle is about 9%, slightly higher than the Blackwell cycle’s 8%. Consensus forecasts show hyperscale cloud firms and emerging cloud service providers’ capital expenditure will grow 69% year-on-year in 2026, slowing to about 13% in 2027, meaning steady increases in power capacity can support continued growth. Bernstein believes this apparent contradiction may just indicate that there is still upside in market expectations for hyperscale capital expenditure.

It is worth noting that potential shortages in LPDDR memory may limit shipments of Vera Rubin and standalone Vera CPUs. Nvidia may choose to ship with lower default memory configurations, allowing customers to expand later, so customers can flexibly determine optimal configs based on the latest memory prices.

~~~~~~~~~~~~~~~~~~~~~~~~

The above brilliant content is from Chasewind Trading Desk.

For more detailed interpretations, including real-time analysis and frontline research, please join 【Chasewind Trading Desk · Annual Membership】

Risk Warning and DisclaimerThe market has risks, investments require caution. This article does not constitute personal investment advice and does not take into account the individual investment goals, financial situation, or needs of any particular user. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investing accordingly is at your own risk.