English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

When Old GPUs Outvalue New Cars: The Compute Economics Behind the H100 Rental Price Rebound

Forum topic · 小凯 · 2026-03-28

Summary

This Chinese tech forum post analyzes a striking anomaly in the AI compute market: NVIDIA H100 GPUs, despite being four years old, are now worth more than when they launched, with rental prices rebounding sharply since December 2025. The article traces the market dynamics: H100 rental prices fell after DeepSeek R1's launch in early 2024, when efficient low-compute training sparked fears that AI compute demand had peaked. However, the Jevons paradox kicked in—cheaper training lowered barriers, expanding the pool of teams training large models, while explosive inference demand (which DeepSeek's efficiency gains did not address) kept pushing prices up. Supply constraints from TSMC's limited CoWoS packaging capacity and US export controls to China further tightened the market. The author argues H100 behaves like 'digital gold' with negative depreciation, upending traditional IT asset amortization models and creating rental arbitrage opportunities. Implications include sustained high compute costs for startups, justification for NVIDIA's valuation, and sticky AI subscription prices. The post concludes that Blackwell, AMD MI-series chips, and hyperscaler in-house silicon may rebalance the market within two to three years.

A Counterintuitive Phenomenon

Electronics normally depreciate fast: a laptop bought for ¥10,000 might fetch ¥2,000–3,000 four years later. But the NVIDIA H100—hard currency of the AI era—is breaking that rule. After rental prices dropped sharply in 2024, especially following the release of DeepSeek R1 (a Chinese model achieving top-tier performance with minimal compute), the market assumed compute demand had peaked. Yet starting December 2025, H100 rental prices rebounded significantly—and four-year-old cards are now worth more than they were three years ago.

The Supply-Demand Seesaw

H100 sits in a market with only one supplier—NVIDIA—facing exponentially growing demand:

  • Demand: Training large models requires massive compute. From GPT-3 to GPT-4, parameters grew tens of times and required compute grew over a hundred times. Every tech company is buying H100s like prospectors buying shovels in a gold rush.
  • Supply: TSMC's advanced process capacity is limited, CoWoS packaging capacity even more so, and export controls bar China—one of the largest AI markets—from direct purchases, pushing gray-market prices higher.
  • DeepSeek's False Alarm

    DeepSeek R1's early-2024 release showed that clever algorithms and engineering could achieve top results with far less compute, triggering a narrative of peaking demand and falling H100 rental prices. But the market soon recognized: efficiency gains don't reduce demand—this is the classic Jevons paradox. Cheaper training meant more teams could afford to build large models, expanding participation beyond tech giants.

    More critically, DeepSeek didn't solve inference. Training is one-time; inference is continuous. Every ChatGPT query requires inference compute, and as AI applications explode, inference demand is growing at a startling pace. H100 prices reversed upward after the brief dip, rising higher than before.

    Why Are Old Cards Worth More?

    In a severely supply-constrained market, availability matters more than age. For AI companies, lacking H100s means being unable to train or serve models. Moreover, performance differences between old and new H100s are smaller than expected—the core specifications of a four-year-old H100 and one built today are identical. As long as it runs, it creates value.

    Restructuring the Asset Model

    Data centers traditionally depreciate GPUs 20–30% per year, assuming near-zero residual value after five years. But some operators now see H100 as more like digital gold than electronics—its value derives from scarcity, not newness. This creates an arbitrage opportunity: buy H100s, rent them out, and years later sell them for possibly more than the initial purchase price, with rental income on top. Such negatively depreciating assets are extremely rare in traditional industries.

    Implications

  • Founders: Compute costs may stay elevated long-term; business models relying on heavy AI computation need cost-structure reassessment.
  • Investors: This explains NVIDIA's high valuation and the global race for advanced process capacity.
  • Users: AI subscription prices won't fall soon—underlying compute costs are real and rising.

What's Next?

Prices won't rise forever. NVIDIA's Blackwell architecture is entering mass 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 four-year-old GPU's price curve is telling a classic story of scarcity, demand, and value—and we're in its most exciting chapter.

---

*Source: easy-learn-ai daily AI news digest, 2026-03-28 (commit 0a830d5).*

Tags

#nvidia-h100#gpu-rental-prices#ai-compute-economics#deepseek#jevons-paradox#export-controls#ai-inference-demand#data-center-assets

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169381