admin
The debate surrounding the appropriate depreciation rate for NVIDIA’s high-performance graphics processing units (GPUs) has emerged as a central point of contention for analysts, investors, and cloud service providers navigating the ongoing artificial intelligence build-out. For much of the past two years, the market has been divided between proponents of a longer shelf life for these specialized chips and detractors who argue that a three-year useful life estimate is the only prudent accounting measure given the rapid pace of silicon innovation. However, recent data tracking hourly rental rates for NVIDIA’s flagship hardware suggests that the "obsolescence" argument may be premature. Current market figures indicate that rather than seeing their value eroded by newer iterations, older generations of GPUs are commanding higher premiums today than they did at the start of 2026, signaling a significant shift in the economic lifecycle of AI infrastructure.
The Counter-Intuitive Rise in GPU Rental Premiums
In a typical hardware cycle, the introduction of next-generation technology leads to a precipitous drop in the market value and rental cost of predecessor models. In the semiconductor industry, this is often driven by Moore’s Law and the relentless pursuit of flops-per-watt efficiency. Yet, the current AI infrastructure landscape is defying these traditional economic gravity wells. According to the latest figures from Silicon Data, NVIDIA’s H100 GPUs—the workhorses of the 2024 and 2025 LLM training boom—were commanding an hourly rental rate just below $2.00 in January 2026. By mid-year, however, those rates have climbed decidedly north of the $2.00 threshold and are currently trending toward the $3.00 mark.
This upward trajectory is not limited to older silicon. The newer B200 GPUs, based on NVIDIA’s Blackwell architecture, have seen a similar appreciation in value. In January 2026, B200 units were available at an hourly rate hovering just below $5.00. Recent market data shows these rates have now stabilized between $5.50 and $5.80 per hour. The fact that rental prices are rising across almost every generation of hardware, even as new capacity is brought online by hyperscalers and specialized GPU clouds, serves as a robust rebuttal to claims that the AI compute market has been overbuilt.
A Chronology of the GPU Supply-Demand Imbalance
To understand why depreciation models are being rewritten, one must look at the timeline of the AI infrastructure rollout. The surge began in late 2022 with the public release of generative AI models, leading to a desperate scramble for NVIDIA’s A100 and H100 chips throughout 2023 and 2024. During this period, the primary concern for hyperscalers like Microsoft, Amazon Web Services (AWS), and Google Cloud was simple availability.
By 2025, as supply chains normalized and NVIDIA began shipping its Blackwell (B100/B200) architecture in volume, many analysts predicted a "GPU glut." The theory suggested that as the most advanced models migrated to Blackwell, the older Hopper (H100/H200) units would be relegated to less intensive tasks, driving their rental prices down toward commodity levels.
Instead, 2026 has revealed a different reality. The demand for inference—the process of running a trained model for end-users—has scaled more rapidly than the demand for training. Because H100s remain highly capable inference engines, they have not been "retired" by the arrival of the B200. Furthermore, the sheer volume of enterprises now integrating AI into their daily workflows has created a floor for demand that exceeds the total available supply of both old and new silicon.
Software Optimization and the Jevons Paradox
The unexpected longevity of NVIDIA hardware is partly attributed to the rapid advancement of software optimization. In the early days of the AI boom, hardware was often utilized inefficiently. However, the development of new kernels, quantization techniques (such as moving from FP16 to INT8 or FP4 precision), and more efficient attention mechanisms has allowed developers to serve more tokens through the same physical footprint.
This phenomenon aligns with the Jevons Paradox, an economic theory stating that as a resource becomes more efficient to use, the total consumption of that resource actually increases rather than decreases. As it becomes cheaper and more efficient to run a large language model on an H100, more companies find it economically viable to deploy AI products. This increased viability leads to higher total demand for H100 clusters, which in turn keeps rental prices high and prevents the "collapsing" price scenario that skeptics predicted.
Furthermore, optimizations in the CUDA software stack have allowed older GPUs to remain compatible with the latest model architectures. Unlike previous cycles in consumer graphics where a three-year-old card might struggle with new software, the enterprise AI ecosystem is designed for backward compatibility and long-term deployment, effectively stretching the "useful life" of the silicon beyond the initial three-year accounting window.
Strategic Shifts in Hyperscaler Accounting
The financial implications of this trend are substantial, particularly for the "Big Three" cloud providers. In a recent disclosure, Microsoft’s Chief Financial Officer, Amy Hood, signaled a significant change in the company’s accounting treatment of its infrastructure. Effective at the start of the 2027 fiscal year, Microsoft is extending the estimated useful life of its data centers and office buildings from 15 to 25 years. While this specific change applies to physical structures, it reflects a broader corporate philosophy that is increasingly being applied to the server racks and GPUs within those buildings.
If a company like Microsoft or Alphabet determines that its tens of billions of dollars worth of NVIDIA GPUs can be depreciated over five or six years rather than three, the impact on their income statements is immediate and massive.
The Mechanics of Depreciation and Net Income
When a hyperscaler purchases $10 billion worth of GPUs, that cost is not expensed all at once. Instead, it is spread out over the "useful life" of the asset. Under a three-year depreciation schedule, the company would recognize roughly $3.33 billion in depreciation expense annually. If the useful life is extended to five years, that annual expense drops to $2 billion.
The $1.33 billion difference flows directly to the "bottom line," boosting net income and earnings per share (EPS) without requiring a single dollar of additional revenue. By demonstrating that older GPUs like the H100 are still generating significant rental revenue and commanding higher prices in mid-2026 than they did in early 2026, hyperscalers have the factual basis required by auditors to justify extending these depreciation timelines. This accounting shift is likely to provide a significant tailwind for the profitability of AI-focused mega-cap tech stocks in the coming quarters.
Market Reactions and Industry Implications
The resilience of GPU pricing has also impacted the broader ecosystem of "GPU Clouds" such as CoreWeave, Lambda Labs, and Northern Data. These companies, which often use their GPU clusters as collateral for debt financing, benefit immensely from the high residual value of their hardware. If an H100 maintains 80% of its rental value after two years, the risk profile for lenders changes dramatically, allowing these specialized providers to secure more favorable terms and further expand their capacity.
Industry analysts have noted that the current pricing power of older chips suggests that the "compute-as-a-commodity" era is still far off. "The chart is one of the clearest rebuttals to the claim that AI compute is already being overbuilt," noted Ricky Ho, a prominent technology analyst. "GPU rental prices are not collapsing as new capacity comes online; they are rising. This indicates that we are still in the supply-constrained phase of the AI revolution."
Future Outlook: The Rubin Architecture and Beyond
Looking ahead, the market is bracing for the arrival of NVIDIA’s "Rubin" architecture, slated for late 2026 and 2027. While Rubin promises another exponential leap in performance, the current trend suggests it will not render Blackwell or Hopper obsolete. Instead, the market appears to be bifurcating: the latest, most expensive chips (Rubin and Blackwell) will be reserved for training the next generation of "frontier" models, while the previous generations (Hopper and Ampere) will handle the massive and growing load of global inference.
This "tiered" utility model supports the argument for a five-to-seven-year useful life for enterprise-grade GPUs. As long as the revenue generated per node continues to rise—driven by both high rental rates and increased token-serving efficiency—the economic value of the hardware remains intact.
In conclusion, the data from the first half of 2026 has provided a vital data point in the ongoing valuation of the AI sector. The rise in rental rates for the H100 and B200 suggests that the "useful life" of AI infrastructure is significantly longer than many initial estimates suggested. For the tech giants investing hundreds of billions into these systems, this shift from a three-year to a five-plus-year horizon represents a fundamental improvement in the unit economics of artificial intelligence, potentially ushering in a period of sustained margin expansion even as capital expenditures remain at record highs.
