The cost of building AI data centers equipped with NVIDIA's next-generation Vera Rubin architecture is far exceeding previous market expectations. According to the latest calculations from Bernstein Research, a single Vera Rubin NVL72 rack costs approximately $9.1 million, pushing the overall data center capital expenditure to about $47 billion per gigawatt. However, the continuous improvement in computing performance per dollar means this massive investment still holds economic rationality for major technology companies.

According to a report from Bernstein analysts including Stacy A. Rasgon, dated June 8, the widely circulated media estimate of "about $8 million per rack" was based on outdated memory prices and significantly underestimated the actual cost. The core discrepancy lies in High Bandwidth Memory (HBM): while current HBM 4 prices are around $16.6 per GB, they are expected to rise to about $53 per GB by 2027 when Vera Rubin is expected to ship in volume. Furthermore, NVIDIA is likely to pass these costs on to end customers through dynamic pricing mechanisms.

The cost increase has not shaken Bernstein's bullish stance on NVIDIA. The firm maintains an "Outperform" rating with a $315 price target. The report also notes that the Vera Rubin NVL72 rack delivers FP8 compute performance of 2,520 petaFLOPS, a significant leap from the previous generation Blackwell's 720 PFLOPS. This translates to substantial gains in both compute per gigawatt and compute per dollar, which is expected to further drive AI application adoption.

Breaking Down the Rack Cost

Bernstein employed a bottom-up approach to deconstruct the Vera Rubin NVL72 rack, arriving at the ~$9.1 million cost estimate.

GPUs remain the largest single cost item. The report indicates a Rubin GPU price of about $55,000 each. With 72 GPUs per rack, the GPU component alone costs $3.96 million, accounting for nearly half the rack's total cost. Additionally, 36 Vera CPUs per rack contribute approximately $180,000 in total.

The significant increase in memory and storage costs is the primary source of the divergence from market expectations. Bernstein estimates this cost at around $3.2 million, far above the ~$2 million calculated using historical prices. This includes ~$1.09 million for HBM 4, ~$800,000 for CPU DRAM (LPDDR5X), and ~$1.28 million for direct-attached storage. The report specifically cautions that memory and storage prices are highly volatile—NAND prices have surged 11.3x from the April 2023 low to May 2026, with an annualized increase of 115%. Investors need to continuously monitor price changes to maintain forecast accuracy.

Networking, cooling, and power supply contribute a combined ~$2 million. This includes ~$1.27 million for networking (NVLink switches ~$250k, cables ~$240k, backplanes/scale-up components ~$380k, SpectrumX switches ~$200k), ~$160k for cooling, and ~$150k for power supply.

The Full-Stack Data Center Bill

Scaling from a single rack to the overall data center capital expenditure reveals an even more staggering figure.

The Vera Rubin NVL72 rack has a rated power consumption of 220 kilowatts. Bernstein estimates that rack power accounts for about 78% of a data center's total electricity usage, implying that one gigawatt can accommodate approximately 3,557 racks. This corresponds to a rack cost of about $32.3 billion. Adding approximately $15 billion per gigawatt for physical infrastructure (including mechanical/electrical equipment and land/building), the full-stack AI data center capital expenditure reaches about $47.3 billion per gigawatt. This represents a further increase of about 17% from the ~$40.5 billion per gigawatt in the previous Blackwell cycle.

The operational cost structure is also noteworthy. The report points out that even at a relatively high electricity price of $0.15 per kilowatt-hour, the annual power cost to run a one-gigawatt data center is about $1.3 billion. Personnel costs are almost negligible, with even the largest data centers requiring only 8 to 10 employees. In contrast, assuming a 6-year depreciation cycle, the annual depreciation cost is approximately $7.9 billion, constituting the main component of operating costs. Since IT hardware (servers, networking) has a much shorter depreciation life (4-6 years) compared to mechanical/electrical equipment and buildings, the economic weight of servers and networking in the true cost is actually higher than their proportion in cash capital expenditure suggests.

Performance Leap as a Counterbalance

Despite the rising capital expenditure per gigawatt, the improvement in compute performance per dollar provides the economic justification for this investment.

Bernstein's data shows the Vera Rubin NVL72 rack delivers 2,520 PFLOPS of FP8 compute, which is 3.5 times that of Blackwell (720 PFLOPS). On a per-gigawatt basis, Vera Rubin can provide about 8,960 exaFLOPS of FP16 sparse compute, more than doubling Blackwell's 4,269 EFLOPS. The compute per billion dollars of capital expenditure also increases from 105.5 EFLOPS to 189.3 EFLOPS.

The report also notes that in a compute-constrained environment, data center operators tend to extend GPU lifespans as much as possible and prioritize using new capacity for deploying next-generation GPUs. If unable to build new capacity due to power or infrastructure constraints, operators may need to consider decommissioning older GPUs to free up space for new chips.

Accelerating demand from the AI sector also supports continued investment. The report cites data showing Anthropic's annualized revenue surged from $9 billion at the end of 2025 to $47 billion by May 2026, with the company stating it has had to turn away some customers and revenue due to compute limitations.

Supply Chain Implications

The changing cost structure is reshaping the AI supply chain's beneficiary landscape.

Memory is the biggest structural beneficiary. CPU DRAM capacity has increased by 320% (in TB) compared to the previous generation, far exceeding the ~50% increases for HBM and NAND. Bernstein also mentions growing applications for CXL memory in KV caching, suggesting DRAM could see disproportionate benefits if supply allows.

Demand for power supply components continues to expand. The report shows the power supply's share of rack cost has increased from about 1.0% in the previous generation to about 1.6%, a trend further driven by early adoption of 800VDC solutions. Bernstein maintains an "Outperform" rating on Delta Electronics with a TWD 2,620 price target, viewing it as a key beneficiary of this power content growth.

Regarding substrates, Bernstein expects continued growth in ABF substrate demand and holds a positive view on Ibiden and Unimicron, with a TWD 990 price target for the latter.

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

Future Cost and Power Trends

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

The report indicates the per-gigawatt cost increase in the Rubin cycle is about 9%, slightly higher than the 8% in the Blackwell cycle. Market consensus suggests capital expenditure from hyperscale cloud providers and emerging cloud services will grow 69% year-over-year in 2026, slowing to about 13% in 2027. This implies that relatively modest additions to power capacity could support continued growth. Bernstein suggests this apparent contradiction might indicate that market expectations for hyperscale capital expenditure still have room to move higher.

It's worth noting potential supply shortages for LPDDR memory could constrain shipments of Vera Rubin and standalone Vera CPUs. NVIDIA may opt to ship with lower default memory configurations, allowing customers to expand later, enabling them to decide the optimal configuration based on prevailing memory prices.