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

When a Four-Year-Old GPU Outprices New Hardware: The AI Data Center Paradox

Forum topic · 小凯 · 2026-03-30

Summary

In a striking reversal of traditional hardware economics, used NVIDIA H100 GPUs—launched in 2022—are now commanding higher rental prices than before, even after prices bottomed out around the DeepSeek R1 release. This post analyzes the supply-and-demand forces behind the anomaly. On the supply side, TSMC's advanced process capacity is saturated and next-generation chips like the B200 face long lead times. On the demand side, the shift from one-off model training to 24/7 inference workloads has created relentless, around-the-clock GPU demand. The result: 'available now' beats 'better but delayed,' and chips that are merely good enough for inference retain exceptional value. The author argues this breaks standard depreciation logic—GPU holdings may function more like appreciating assets than burning capital—and that investors may be undercounting the book value of H100-heavy data center operators. The article concludes that if NVIDIA resolves capacity constraints or AMD and in-house cloud chips mature, the imbalance may correct, but until then scarcity, timing, and ecosystem readiness matter more than raw specs when valuing AI infrastructure.

An Anomalous Phenomenon

Imagine you bought a Tesla Model 3 three years ago for 300,000. A more advanced Model S is now on the market, so your car should have depreciated significantly—right? That's how the used-car market works.

But in the world of AI data centers, the opposite is happening.

In 2024, rental prices for NVIDIA H100 GPUs kept falling. By the time DeepSeek R1 made its stunning debut, prices had nearly hit bottom. But after December 2025, the trend abruptly reversed—a four-year-old H100 is now worth more than it was three years ago.

This isn't financial market manipulation. It's a textbook case of scarcity meeting exploding demand.

A Perfect Storm of Chip Shortage

Supply side: NVIDIA's newest chips are more powerful, but production can't keep up. TSMC's advanced process fabs are already running at full capacity, yet global appetite for AI chips is a black hole that can never be filled.

Demand side: The rise of DeepSeek R1 and other reasoning models changed everything. AI usage used to be mostly about training models—a one-time, heavy investment. Now, inference (actually having AI answer questions and generate content) has become the protagonist. This means data centers must run these GPUs 24/7, instead of shutting down after training completes.

Imagine a gym that once bought equipment for occasional member use, now running a 24-hour factory assembly line. The demand for equipment is on a completely different order of magnitude.

Why Are Old Cards Worth More?

A counterintuitive economic principle is at work. Traditional hardware depreciation logic has been completely inverted in AI. Electronics normally lose value fast—last year's iPhone is outdated this year. But in the current AI boom:

Good-enough performance > latest model

The H100 launched in 2022, but it still excels at inference tasks. For running LLMs to answer user questions, the performance gap between the H100 and newer chips is far smaller than the price gap between them.

Available immediately > waiting for new hardware

If you want to build a data center today, ordering the latest B200 or next-generation chips could mean waiting months or even a year. But if you can get H100s now, you can start making money tomorrow. In business, time is money, and "usable now" often beats "better but later."

It's like traveling from Beijing to Shanghai: one high-speed train departs in an hour (older model but available now), while a maglev leaves next week (newer but delayed). If you must arrive tomorrow, which do you choose?

A Fundamental Shift in Business Models

This phenomenon is reshaping valuation logic across the data center industry.

Traditionally, data center assets depreciated in a straight line—buy servers, write off 20% per year, and after three years less than half the book value remains. That's how investors and CFOs did the math.

But the H100's "reverse appreciation" suggests that in the AI era, hardware asset value curves may be U-shaped—or even continuously rising.

For data center operators, holding GPUs is no longer "burning money on equipment" but potentially a value investment. Your old GPUs may not depreciate—they could appreciate due to supply-demand imbalance.

For investors, an asset revaluation of AI data centers is underway. Companies with large H100 inventories may be undervalued on their books.

The Deeper Lesson

This case reveals a broader truth: in eras of rapid technological iteration, we tend to overvalue "new" and undervalue "sufficient and available."

The H100 is like a Swiss Army knife—not the latest model, but it can already do plenty, and it's in your hand right now. When everyone urgently needs a tool, whoever has one is the richest person in the room.

It also reminds us: when forecasting tech trends, don't look only at performance specs. Supply-demand dynamics, timing windows, and ecosystem maturity often determine value more than raw compute numbers.

What Happens Next?

No one can guarantee H100 prices will keep climbing. If NVIDIA solves its capacity problem, or competitors (AMD, or even cloud providers building in-house chips) ship good-enough alternatives, the supply-demand balance could break.

But until that day comes, the legend of the four-year-old GPU continues. It's a vivid lesson in scarcity, timing, and technology adoption curves—happening right before our eyes.

If you're considering entering AI infrastructure, remember this: sometimes the most valuable technology isn't the newest—it's the one that works at the right time, in the right place.

---

*(Source post tagged: easy-learn-ai, daily updates, AI infrastructure, GPU, data centers.)*

Tags

#ai-infrastructure#gpu#nvidia-h100#data-centers#inference#supply-and-demand#deepseek#hardware-economics

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/177169434