When a 4-Year-Old GPU Outprices a New One: The AI Compute Market's Counterintuitive Story

> Source commit: 0a830d5 > Original analysis: H100 rental price trends

A Counterintuitive Phenomenon

Imagine you bought a GPU for 100,000 RMB three years ago. Normally, selling it secondhand would recoup maybe 30,000. Instead, buyers are now offering 120,000—more than the original price.

This isn't collectibles speculation. It's what actually happened to the NVIDIA H100 between late 2025 and early 2026.

Why Is an Old Card Worth More?

Variable 1: DeepSeek R1's Shock—and Rebound

In early 2025, DeepSeek's R1 model stunned the industry: a Chinese team achieved reasoning capability rivaling OpenAI's o1 at extremely low training cost. The market panicked—"maybe we don't need that much compute after all?"—and H100 rental prices bottomed out. Companies stockpiling GPUs feared their assets would become scrap.

But the market soon realized R1 didn't reduce compute demand—it increased it. R1 proved powerful reasoning models are viable, and reasoning models require longer inference chains—meaning each conversation consumes more compute.

Variable 2: The Chip Shortage Didn't Ease

Geopolitical and supply-chain constraints persisted. Although Blackwell (B200/B100) has been announced, production ramp takes time. Demand rose while supply lagged, making the 2022-era H100 a "scarce resource" in 2026.

Variable 3: Reasoning Models' Appetite for Compute

A traditional language model is like a bright but impatient student: you ask, it answers immediately. A reasoning model is more like a careful researcher: it decomposes the problem, derives step by step, verifies intermediate conclusions, then answers.

This "thinking process" generates massive intermediate tokens. Where a traditional model might need 100 tokens to answer, a reasoning model may need 1,000 or more. Result: the stronger the reasoning, the higher the compute consumption.

Reshaping the Business Model

From Depreciating to Appreciating Assets

GPUs have traditionally been seen as rapidly depreciating hardware. But now a 4-year-old H100 is worth more than 3 years ago—unheard of in the hardware industry. For data centers and cloud providers, this means:

• Depreciation models need rewriting: the traditional 3–5 year cycle no longer applies
• Inventory strategy changes: stockpiling GPUs is no longer foolish—it may be a good investment
• Rental pricing logic shifts: long-term lease pricing benchmarks need recalibration

A Structural Supply-Demand Mismatch

The deeper signal: the mismatch between AI compute supply and demand is structural, not cyclical. Even if Blackwell ships at scale, new models' compute demands may keep outpacing supply growth—a kind of "AI compute law" where AI models' compute appetite doubles every two years, a twist on Moore's Law.

Implications for Founders and Developers

1. Compute cost uncertainty has increased. We used to assume compute would get cheaper. That assumption no longer holds—product planning should buffer for upward cost risk.

2. Local deployment is being revalued. If cloud compute prices rise, on-premise or private clusters look more economical, potentially accelerating edge computing and localized deployment.

3. Efficiency optimization becomes more valuable. When compute is cheap, efficiency matters little. When compute prices rise, an optimization that cuts consumption by 50% becomes a hard-dollar competitive advantage.

What Does This Mean Long-Term?

The H100's "reverse depreciation" may be a snapshot of the AI industry entering a new phase: shifting from technology-driven competition (who trains the better model) to resource-driven competition (who controls compute and energy supply). Technology still matters, but the window of technical advantage is shrinking while the moat of resource advantage deepens.

Google's willingness to fund Anthropic's data centers isn't just about Anthropic's technology—it's about ensuring a key ecosystem partner doesn't fall behind due to compute bottlenecks. In this sense, compute is becoming the oil of the AI era—both a production input and a strategic resource.

Final Thoughts

A four-year-old GPU outpricing a new one sounds like a joke, but it reflects the AI industry's most serious structural change. When planning your AI project, ask: if compute gets more expensive rather than cheaper, does your business model still hold? For those with warehouses full of H100s, this may be their most unexpected "investment"—even though they originally bought the cards just to train models.