Why a Four-Year-Old GPU Is More Valuable Than a New Car: The Compute Economics Behind H100 Rental Prices

Forum topic · 小凯 · 2026-04-01

Summary

This post analyzes the surprising price resilience of NVIDIA's H100 GPU, released in 2022, whose rental rates are rising despite its age. The author traces the shift to the January 2025 release of DeepSeek R1, which triggered a 17% single-day drop in NVIDIA's stock yet ultimately increased demand for compute. The key argument: reasoning models generate heavy, ongoing inference workloads—projected to exceed 70% of datacenter AI compute demand by 2026—shifting demand from training to inference. Since inference depends heavily on memory capacity (80GB VRAM and KV Cache storage) rather than raw compute speed, older H100s remain competitive. Supply constraints on newer Blackwell GPUs (e.g., Microsoft Azure pausing bookings due to insufficient capacity) plus a market-wide pivot from owning to renting GPUs have tightened rental supply. The post concludes that compute has become a form of power, leasing will dominate, and memory bandwidth will be the key bottleneck, making the H100 a fortunate beneficiary of its timing in AI history.

Source: easy-learn-ai / commit 0a830d5

🚗💨 Why a Four-Year-Old GPU Holds Value Better Than a New Car — The Compute Economics Behind H100 Rental Prices

> "Imagine: you bought a car four years ago for 30,000 yuan, drove it for four years, and now someone offers to rent it for a month at 40,000. Absurd? In the AI world, this is becoming reality."

🚙 A Strange Analogy: Cars vs. GPUs

Everyone knows the common sense: things get old and depreciate. A new car loses 20% of its value the moment you drive it home, and after four years may be worth less than half its original price.

But the NVIDIA H100, released in 2022 and now four years old, defies this logic. Electronics should depreciate even faster than cars, yet its rental prices are heading upward. A four-year-old GPU whose rent is rising, not falling.

Why? Because this is not an ordinary electronic product — it's compute. And in the AI era, compute is power.

🌊 The Shock That Changed Everything: DeepSeek R1

In early 2025, China's DeepSeek released the R1 model, achieving performance close to OpenAI's top reasoning models at extremely low cost. The market reacted violently: NVIDIA's stock plunged 17% in a single day, wiping out $589 billion in market value. People asked: if good models can be built with cheaper chips, do we still need so many expensive GPUs?

Many assumed this was the end of compute demand. They were wrong — the opposite happened.

⚡ The Rise of Inference: From "Top Student" to "Thinker"

Earlier AI models were like a super student: feed it a book and it memorizes everything instantly, answering from sheer memory. This capability relies on pretraining, which demanded massive GPU clusters — hence the earlier GPU demand explosion.

DeepSeek R1 represents a new paradigm: reasoning models. Instead of answering immediately, they stop, think step by step, write out their reasoning, and then answer — like a thoughtful scholar working through a draft rather than a student reciting from memory.

This "thinking" process is called inference.

🔄 The Demand Shift: From Training to Inference

DeepSeek R1 proved that reasoning models can rival massive models built by brute-force compute. But there's a crucial side effect: reasoning models need far more computation. Using one is like hiring a professor who thinks, derives, and drafts — this takes time and compute.

Industry projections suggest that by 2026, inference workloads will account for over 70% of datacenter AI compute demand. GPU demand didn't shrink after DeepSeek — it changed direction.

Moreover, inference is a continuous demand: training is one-time, but as long as people use AI, compute is needed to "think."

💎 Why Are Old Cards Becoming More Valuable?

The answer lies in shifting supply and demand:

Supply side: New cards are unavailable

NVIDIA's Blackwell GPUs (like the B200) are stronger, but severely undersupplied. Months ago, Microsoft Azure paused bookings for some new GPUs, citing "insufficient capacity." Like an iPhone with a 3-month shipping delay — buyers fall back on older models, and if those are also scarce, prices rise.

Demand side: Inference needs memory

Reasoning models require large amounts of VRAM for the "KV Cache" — essentially the AI's scratchpad while thinking. More thinking means a bigger scratchpad. The H100's 80GB of VRAM makes it ideal for large-model inference. For inference tasks, memory capacity and bandwidth often matter more than raw compute speed, so old cards aren't far behind.

The rental market dynamic

Companies increasingly rent compute rather than buy it, because AI technology changes too fast and leasing avoids depreciation risk. But when everyone wants to rent, rental supply becomes extremely tight. As industry analysts put it: "AI datacenters are shifting from a purchasing model to a leasing model, and supply can't keep up."

🏭 The Datacenter Business Revolution

Traditional datacenters were like rented warehouses — priced by rack, bandwidth, and power. In the AI era, datacenters have become "intelligence factories": customers ask how much compute you can provide, how large a model you can run, how fast you respond. Value no longer comes from hardware, but from what you can produce.

When compute becomes the scarcest resource, those who hold it hold the power. Amazon, Microsoft, Google, and Meta are building datacenters frantically — not because they love buying GPUs, but because compute is power. Whoever controls compute can train the best models, offer the cheapest inference, and stand undefeated in the AI wave. Application-layer companies must bow to compute owners. That's the new power structure.

🕰️ Old Cards as Antiques: A Metaphor for an Era

The H100 witnessed AI's explosive growth — once the mainstay for training GPT-4, now a pillar of inference services. Its rising price is this era's most absurd yet truthful footnote.

Some antiques appreciate not because their function is advanced, but because they embody an unrepeatable historical moment and scarcity. The H100 is becoming a digital-age antique — not because its technology is the most advanced, but because it happened to stand at the right historical inflection point: exactly when inference demand exploded, new card supply fell short, and everyone was scrambling for compute.

🔮 What Comes Next?

Markets always seek balance. NVIDIA is ramping up Blackwell production, and more capacity will come online in coming quarters, which may push H100 prices back down.

But some things have permanently changed:

1. Inference will dominate compute demand — a structural shift, not a cyclical fluctuation. 2. Leasing will dominate the market — flexibility beats ownership. 3. Memory will be the key bottleneck — for inference, "can fit it" matters more than "computes fast." 4. Compute is power — this will define the tech landscape for the next decade.

📌 Conclusion: The Story of One GPU

So next time someone asks why a four-year-old GPU holds value better than a new car, tell them: this isn't the story of a GPU — it's a microcosm of an era.

As AI moves from "memorizing" to "thinking," as compute turns from "resource" into "power," and as the entire market goes mad for "chips that can reason," an old card born at just the right moment becomes the scarcest hard currency.

In this fast-changing era, what holds value isn't newness or oldness — it's standing at exactly the right point in time. The H100 is that lucky "old-timer."

---

References: 1. CoreWeave Market Update - AI Infrastructure Demand Analysis, 2025 2. NVIDIA H100 vs B200 Supply Chain Report, Q1 2026 3. DeepSeek R1 Technical Report & Market Impact Analysis 4. "Inference vs Training: The Shift in AI Compute Demand" - SemiAnalysis, 2025 5. "The New Economics of AI Datacenters" - Morgan Stanley Research, 2026