Source: easy-learn-ai / commit 0a830d5

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

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

A Strange Analogy: Cars vs. GPUs

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

But there's a GPU — the NVIDIA H100, released in 2022, now four years old. Electronics should depreciate even faster than cars. Yet the reality is that its rental prices are rising. A four-year-old GPU whose rent is going up, not down.

Why? Because this isn't 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, Chinese company DeepSeek released R1, a model that achieved performance close to OpenAI's top reasoning models at extremely low cost. The market shook: NVIDIA's stock plunged 17% in a single day, erasing $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: When AI Goes from Memorizer to Thinker

Earlier AI models were like a super student with perfect memory: trained on massive data ("pre-training"), they could answer instantly from what they had absorbed. That's why GPU demand exploded in recent years — everyone was training bigger models.

DeepSeek R1 represents a new paradigm: reasoning models. Instead of recalling answers, they stop, think step by step, write out their reasoning, and then answer — like a professor with a scratch pad rather than a student reciting from memory. This "thinking" process is called inference.

The Demand Shift: From Training to Inference

R1 proved that reasoning models can rival massive compute-trained giants. But there's a key side effect: reasoning models require far more computation per use. Training is a one-time cost; inference is a continuous one — as long as anyone uses AI, compute is needed to "think."

Industry forecasts 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.

Why Is an Old GPU Becoming More Valuable?

Supply Side: New Cards Are Unavailable

NVIDIA's latest Blackwell architecture GPUs (like the B200) are more powerful but severely undersupplied. Months ago, Microsoft Azure paused reservations for some new GPUs, citing insufficient capacity. If you can't wait for the new model — and the old model is also scarce — prices rise.

Demand Side: Inference Needs Memory

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

The Rise of the Rental Market

More companies now rent compute rather than buy it, because AI hardware evolves too fast — today's purchase may be obsolete next year. But when everyone wants to rent, rental supply tightens. As one industry analyst put it: "AI datacenters are shifting from a purchasing model to a rental model, and supply can't keep up."

The Datacenter Business Revolution

Traditional datacenters rented out warehouse-like server space billed by rack, bandwidth, and power. In the AI era, datacenters are becoming "intelligence factories": customers care about how much compute you can provide, how large a model you can run, and how fast you respond. Value comes from what you can produce, not the hardware itself.

When compute is the scarcest resource, whoever holds it holds the power. Amazon, Microsoft, Google, and Meta are racing to build datacenters not because they love buying GPUs, but because compute determines who trains the best models and offers the cheapest inference. Application-layer companies must bow to compute owners — that's the new power structure.

Old Cards Becoming Antiques

The H100 witnessed AI's explosive growth: once the workhorse for training GPT-4, now a pillar of inference services. Like antiques, its value comes not from being the most advanced, but from representing an unrepeatable historical moment — it happened to be the "just good enough" choice when inference demand exploded, new cards were scarce, and everyone was scrambling for compute.

What Happens Next?

Markets seek equilibrium. NVIDIA is ramping Blackwell production, and H100 prices may fall once new supply arrives. But some things have permanently changed:

1. Inference will dominate compute demand — a structural shift, not a cyclical fluctuation. 2. Rental models will lead the market — flexibility beats ownership. 3. VRAM will be the key bottleneck — for inference, "fits in memory" matters more than "computes fast." 4. Compute is power — this will define the tech landscape for the next decade.

Conclusion

So next time someone asks why a four-year-old GPU holds its value better than a new car, tell them: this isn't a story about a GPU — it's a microcosm of an era. As AI moves from memory to reasoning, and compute from resource to power, an old card that arrived at the right moment becomes the hardest currency. What holds value isn't newness, but 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