Why 4-Year-Old H100 GPUs Are Worth More Than New Cars: The Economics Behind the GPU Rental Price Rebound
You wouldn't expect a four-year-old electronic device to be worth more today than when it was new. Yet this is exactly what is happening in the AI compute market with NVIDIA's H100 GPU.
A Counterintuitive Phenomenon
Normal consumer electronics follow a steep depreciation curve—buy a laptop for $1,000, sell it four years later for $300. The H100, however, has broken this rule. After a sharp drop in 2024 rental prices—triggered partly by the release of DeepSeek R1, a Chinese model that achieved top performance with far less compute—market sentiment swung the other way. From December 2025, H100 rental prices rebounded sharply. Even more strikingly, units that have been running for four years are now worth more than they were three years ago.
The Supply-Demand Seesaw
Training large models requires massive compute, and demand continues to grow exponentially. From GPT-3 to GPT-4, parameter counts jumped several dozen times, while compute requirements rose even faster. Nearly every major tech company is competing for H100 capacity—the modern equivalent of gold-rush miners scrambling for shovels.
On the supply side, constraints are severe. TSMC's advanced-node capacity is limited, and the CoWoS advanced packaging used in H100 is even more constrained. US export restrictions prevent direct sales to China, one of the largest AI markets, pushing demand into grey channels and further inflating global prices. When demand explodes and supply is capped, prices behave like a compressed spring.
The DeepSeek 'False Alarm'
DeepSeek R1 showed that clever algorithms and engineering optimization could deliver top-tier performance with far less compute. Headlines proclaimed 'peak compute demand.' H100 rental prices dropped. Some questioned NVIDIA's valuation.
The market quickly realized that efficiency gains do not reduce demand. This is Jevons paradox in action: when coal-burning efficiency improved historically, people used more coal, not less. Cheaper training expanded the addressable market—mid-sized companies that previously couldn't afford to train frontier models can now participate. Compute demand rose rather than fell. Crucially, DeepSeek did not solve the inference problem. Training is one-time; inference is continuous. Every ChatGPT query needs GPU compute. As AI applications proliferate, inference demand grows rapidly.
Why Are Older Cards Worth More?
Electronics usually depreciate because newer models offer higher performance and lower power. H100 is different because it sits in a severely supply-constrained market. For AI companies, no H100 means no model training and no service. In this 'hunger game,' availability—not age—determines price.
In addition, old and new H100 units differ little in AI training and inference performance. The architecture has been AI-optimized since launch, and core specifications have remained consistent across production years. As long as a unit runs, it creates value.
Reimagining Asset Models
Data centers traditionally depreciate GPU equipment at 20–30% per year, assuming near-zero residual value after five years. The H100 value curve inverts this model. Some operators now view H100 as 'digital gold' rather than electronic equipment—its value is driven by scarcity, not newness. Until supply constraints ease (potentially years away), H100 real value may hold or appreciate.
This creates an arbitrage opportunity: purchase H100 today, rent it out, and resell later. Rental income plus resale price can exceed the initial outlay. Such 'negative depreciation' is extremely rare in traditional industries.
Implications for End Users
For entrepreneurs, compute costs may stay elevated long-term. Business models relying on heavy AI inference should reassess cost structures. For investors, this explains NVIDIA's premium valuation and why governments compete for advanced-node capacity. For end users, it explains why ChatGPT subscription prices are unlikely to drop soon—underlying compute costs are real and rising.
The Road Ahead
H100 prices will not rise forever. Supply-demand imbalances will eventually ease. NVIDIA's Blackwell architecture is ramping, AMD's MI series is catching up, and hyperscalers are developing custom silicon. Within two to three years, the compute market may reach a new equilibrium. Until then, the H100 story continues—a classic tale of scarcity, demand, and value, unfolding in real time.
---
Source: Commit 0a830d5, easy-learn-ai 2026-03-28 AI News Daily.