When Old GPUs Outvalue New Cars: The Compute Economics Behind H100 Rental Price Rebound
*Translated from a Chinese tech forum post (source: easy-learn- AI daily news digest, commit 0a830d5, dated 2026-03-28).*
Have you ever considered that a four-year-old electronic product could be worth more today than when it was brand new? That's not an antique-market curiosity—it's happening right now in AI compute.
An Anomalous Phenomenon
Normal electronics depreciate steeply: a laptop bought for the equivalent of $1,400 might fetch only a few hundred dollars four years later. But the NVIDIA H100—the "hard currency" of the AI era—is breaking that rule. Rental prices fell rapidly in 2024, especially after DeepSeek R1's release, when a Chinese model achieving top-tier performance with minimal compute led many to believe compute demand had peaked. But starting December 2025, H100 rental prices rebounded sharply. Cards that have been in service for four years—cards that "should be in a museum"—are now worth more than they were three years ago.
The Supply-Demand Seesaw
Training large models requires massive, exponentially growing compute—going from GPT-3 to GPT-4 multiplied parameters dozens of times and computational needs over a hundredfold. Every tech company is buying H100s like gold-rush miners buying shovels. Meanwhile, supply is choked: TSMC's advanced process capacity is limited, CoWoS packaging capacity even more so, and export controls force one of the largest AI markets (China) into gray-market channels, further pushing up global prices. When demand surges while supply is capped, price acts like a compressed spring.
DeepSeek's "False Alarm"
DeepSeek R1 proved that clever algorithms and optimized engineering can reach top results with far less compute, triggering "peak compute demand" narratives and a price drop. But the market soon remembered a lesson from history: efficiency gains don't reduce demand—they increase it. This is the Jevons paradox: when coal burning became more efficient, people used *more* coal. Cheaper training meant more teams could afford to build large models, expanding participation beyond tech giants. Crucially, DeepSeek didn't solve inference: training is one-time, but inference is continuous—every ChatGPT query requires GPU compute. As AI applications explode, inference demand is growing at an astonishing pace, so H100 prices rebounded higher than before.
Why Are Old Cards Worth More?
In an extremely undersupplied market, availability matters more than age. Like a starving diner accepting a dish that's been sitting out, AI companies will pay a premium for any usable H100—without one, they cannot train models or serve customers. Moreover, for AI training and inference specifically, the performance gap between a four-year-old H100 and a freshly made one is smaller than assumed: the core specifications are identical, and the architecture was purpose-built for AI. As long as the card runs, it creates value.
Restructuring the Asset Model
Data centers traditionally account for GPUs under IT depreciation models—20–30% value loss per year, near-zero residual value after five years. But H100's value curve looks nothing like that. Some operators now see the H100 as "digital gold" rather than electronics: its value derives from scarcity, not newness. Until supply catches up (possibly years away), its value may hold steady or even appreciate. This creates an arbitrage opportunity: buy H100s, rent them out, then resell years later—rental income plus resale value could exceed the original purchase price. Such "negative depreciation" assets are extraordinarily rare in traditional industries.
What It Means for Ordinary Users
This reveals a core logic of the AI wave: compute is the new oil. For entrepreneurs, AI compute costs may stay high for the long term, requiring a rethink of cost structures. For investors, it explains NVIDIA's high valuation and why governments worldwide are competing for advanced-node capacity. For consumers, it explains why subscription prices for services like ChatGPT won't fall soon—the underlying compute costs are real and rising.
What Happens Next?
H100 prices won't rise forever. NVIDIA's next-generation Blackwell architecture is ramping to volume production, AMD's MI series is catching up, and cloud giants are developing in-house chips. In two to three years, the compute market may reach a new equilibrium. Until then, the H100's price curve tells a classic story of scarcity, demand, and value—and we are living through its most exciting chapter.