Megawatt return calculations: Nvidia rewrites the rules for computing hardware depreciation

Nvidia places its hand on the most financially sensitive nerve in the computing infrastructure sector, as the cost of building an AI factory reaches roughly sixty million dollars per megawatt. Data-center operators do not invest at this scale without a clear view of the financial return, which is an equation based on three interlinked pillars: the annual earnings potential if the facility sells every token it produces, the actual productive lifespan of the hardware, and the market demand for those tokens.Strength on one side does not compensate for weakness on another, high production capacity loses its value if the operator fails to market the full capacity, and massive demand is useless if the servers stop operating at full power after one year.

Electric power is the governing constraint on computing facilities, making the token production rate in tokens per second per megawatt the actual revenue metric. Data from Semi Analysis, a research firm, via its Egent X index show that the Vera Rubin NVL 72 systems achieve productivity more than thirty times per megawatt compared with the GB 300 NVL 72 systems, while reducing the cost of a million tokens by up to forty-five times when running the DeepSec V4 Pro model. This sharp drop in the cost per token does not reduce overall demand for computing power, but expands it, because the lower cost opens economic opportunities for new use cases that together consume a volume of tokens exceeding what efficiency gains have delivered.

The second pillar of the equation rests on extending the economic lifespan of hardware and postponing its retirement as technology generations succeed each other. Computational workloads differ in complexity, allowing earlier generations to keep generating returns, as is the case with the A100 processors launched in 2020 that remain in commercial service after six years, with CoreWave, a chipmaker, extending its reservation contracts for those units until 2029.Data from a study carried out by Sprout, a research organization, in September 2026 show that major operators have continued to extend the accounting depreciation schedules for their servers to reflect that they actually remain in service for longer periods.And reality indicators confirm that; Microsoft’s fleet of “F100” chips has occupied about 8.4 years versus a nominal lifespan of only six years, Parker’s analyses estimate the productive life at five to six years for H100 systems and nine to ten years for GP300 systems, while Silicon Data notes that the A100 chip retains a quarter of its original value after six years, and Orn Data records the market paying 80 percent of the lease price in contracts extending five years compared with monthly contracts.

The third pillar is hardware’s suitability for diverse workloads, as general-purpose architectures avoid the risk of confining investment to single-task chips that may become obsolete quickly. The CUDA platform and more than a thousand specialized libraries within the CUDA X suite enable over ten million developers to run deep-learning, generative AI, agents and embodied AI workloads, as well as scientific simulation, graphics and data-processing tasks that all converge on the same parallel mathematical operations. In practice this is manifested by Illumina’s operation of genome models and molecular design on a cluster containing 10^16 processors, by Pinterest’s reliance on 14,000 processors spanning Blackwell, Hopper and older generations, by Revolut’s use of data-acceleration libraries to process billions of records, and by the University of Texas at Austin running its supercomputer with utilization rates ranging from 95 to 98 percent across twenty-seven research tasks serving seven institutions.

This shift in return calculations directly affects technology infrastructure managers and computing project leaders in the Gulf, Egypt and the Arab East. When building national data centers or negotiating cloud-infrastructure deals, it becomes clear that adopting short-term depreciation schedules of three to four years burdens budgets with imprecise accounting entries and makes demonstrating the economic viability of projects more difficult. The decisive metric lies in calculating the cost per available megawatt of capacity and the ability of purchased systems to handle inference tasks and serve agents over the long term without having to write off assets or replace them entirely with each announcement of a new chip generation at industry events such as the upcoming GTC Berlin conference (Nvidia's GPU Technology Conference).