When Old GPUs Hold Value Like New Cars: The Compute Economics Behind H100 Rental Price Rebound

Forum topic · 小凯 · 2026-03-28

Summary

NVIDIA H100 rental prices, after dipping in 2024 following DeepSeek R1's release, have rebounded sharply since December 2025—four-year-old units now fetch more than they did three years ago. This article explains the economics: AI training demand is growing exponentially while supply is constrained by TSMC's advanced packaging (CoWoS) capacity and export controls limiting China's access. The DeepSeek R1-driven price dip proved to be a false alarm: efficiency gains triggered the Jevons paradox, lowering barriers to entry and expanding total compute demand, while inference workloads continue to scale with AI application adoption. Because availability, not age, drives pricing in a shortage market, H100s resist typical IT depreciation, with some operators treating them as appreciating 'digital gold'—enabling rental-income-plus-resale arbitrage. The piece expects rebalancing as Blackwell, AMD MI-series, and hyperscaler in-house chips arrive within a few years, but argues compute remains the strategic resource of the AI era.

Source

Commit: 0a830d5
Source: easy-learn-ai daily AI news digest, 2026-03-28

When Old GPUs Hold Value Like New Cars: The Compute Economics Behind the H100 Rental Price Rebound

Have you ever seen a four-year-old consumer electronic worth more today than when it was new? That's not an antique-collecting quirk—it's the real story playing out in AI compute.

A Counterintuitive Phenomenon

Consumer electronics normally depreciate like a slide: a ¥10,000 laptop might fetch ¥2,000–3,000 after four years. But the NVIDIA H100 GPU, the "hard currency" of the AI era, is breaking that rule.

In 2024, H100 rental prices fell quickly, especially after DeepSeek R1 launched—the Chinese model that achieved top-tier performance with remarkably little compute led many to believe compute demand had peaked. But the story didn't follow the script. From December 2025, H100 rental prices rebounded sharply, and four-year-old cards are now worth more than they were three years ago.

This is not mystical—it's a live economics lesson.

The Supply–Demand Seesaw

Training large models demands massive, exponentially growing compute—parameters multiplied dozens of times from GPT-3 to GPT-4, with compute requirements up a hundredfold. Every tech company is buying H100s like prospectors buying shovels in a gold rush.

Supply, however, is constrained: TSMC's advanced process capacity is limited, CoWoS packaging capacity even more so, and export controls mean China—one of the largest AI markets—can only source chips through gray channels, pushing global prices higher. When demand surges against constrained supply, price acts like a compressed spring.

DeepSeek's "False Alarm"

DeepSeek R1 showed that clever algorithms and engineering can reach top results with far less compute, and "compute demand has peaked" narratives spread—H100 rental prices fell accordingly, with some fearing an NVIDIA valuation bubble.

But the market quickly realized: efficiency gains don't reduce demand. Historically, every improvement in computing efficiency has ultimately increased total compute consumed—the classic Jevons paradox. Cheaper training means more teams can afford to train large models; a game once reserved for tech giants opened to mid-sized companies. Demand rose instead of falling.

More importantly, DeepSeek didn't solve inference. Training is one-off; inference is continuous. Every ChatGPT query requires GPU inference, and as AI applications explode, inference demand is growing at a staggering pace. So H100 prices turned upward, higher than before.

Why Are Older Cards Worth More?

Normally, electronics depreciate as newer products offer more performance at lower power. But H100s sit in an extremely supply-starved market. Like a hungry diner with one dish on the menu, AI companies will pay a premium even for aging hardware—without H100s they can't train or serve models. In this "hunger game," availability, not age, determines price.

Moreover, H100 is an AI-optimized architecture: the core specifications of a four-year-old H100 and a freshly produced one are identical. As long as it runs, it creates value.

A Reconstructed Asset Model

This is rewriting cloud and datacenter business logic. Traditional IT depreciation assumes 20–30% annual value loss and near-zero residual value after five years. H100's value curve looks nothing like that. Some operators now see H100 as "digital gold" rather than electronics: its value depends on scarcity, not newness, and may hold steady or even appreciate until supply catches up—possibly years away.

That creates an arbitrage: buy an H100 today, rent it out, resell it later—rental income plus resale price may exceed your initial outlay. Such "negative depreciation" assets are extraordinarily rare in traditional industries.

What It Means for Regular Users

The underlying logic: compute is the new oil. For founders, AI compute costs may stay elevated—reassess your cost structure if your business depends on heavy AI workloads. For investors, it explains NVIDIA's high valuation and governments' scramble for advanced process capacity. For everyday users, it explains why ChatGPT subscription prices won't drop soon—underlying compute costs are real and rising.

What Comes Next?

H100 prices won't rise forever. NVIDIA's Blackwell-generation products are ramping, 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 story continues—reminding us that during radical technological shifts, traditional business logic can be overturned, and new opportunities and risks coexist.

That four-year-old GPU is telling a classic story of scarcity, demand, and value through its price curve—and we're living through its most exciting chapter.