Source
- Source: easy-learn-ai daily AI news digest, 2026-03-28
- For founders: compute costs may stay high for a long time. If your business depends on heavy AI computation, re-evaluate your cost structure.
- For investors: this explains NVIDIA's lofty valuation and why governments worldwide are competing for advanced-node capacity.
- For users: it explains why ChatGPT subscription prices will not fall soon—the underlying compute costs are real and rising.
When Old GPUs Outlast New Cars: The Compute Economics Behind the H100 Rental Price Rebound
Have you ever imagined that a four-year-old electronic product could be worth more today than when it was new?
This is not a quirk of the antiques market—it is happening right now in AI compute.
A Counterintuitive Phenomenon
Ordinary electronics depreciate steeply: a laptop bought for a large sum is worth a fraction of that four years later. But the NVIDIA H100 GPU, the "hard currency" of the AI era, is breaking that rule.
According to recent market observations, H100 rental prices fell rapidly through 2024, especially after the release of DeepSeek R1—a Chinese model that achieved top-tier performance with remarkably little compute, leading many to believe compute demand had peaked. The story did not follow that script: starting in December 2025, H100 rental prices rebounded sharply. More surprisingly, cards that have been in service for four years—old enough for a museum—are now worth more than they were three years ago.
This is not some mysterious force. It is a vivid economics lesson.
The Supply-Demand Seesaw
To understand why the H100 matters, imagine running a factory that needs a special machine only one company in the world can build—NVIDIA. Worse, that company is politically barred from selling to some of its largest potential customers.
On one side, demand is exploding. From GPT-3 to GPT-4, model parameters grew dozens of times, and required compute grew by orders of magnitude more. Every tech company is scrambling for H100s like gold miners grabbing shovels.
On the other side, supply is choked. TSMC's advanced process capacity is limited, and CoWoS packaging capacity—the packaging technology the H100 relies on—is even tighter. Export controls mean China, one of the largest AI markets, can only obtain chips through gray channels, further inflating global prices.
When demand surges while supply is capped, price acts like a compressed spring.
DeepSeek's "False Alarm"
In early 2024, DeepSeek R1's release shook the market. A Chinese team demonstrated that clever algorithms and engineering could reach top results with far less compute. Talk of "peak compute demand" spread, H100 rental prices fell, and some began to worry about a bubble in NVIDIA's valuation.
But the market quickly recognized a problem: efficiency gains do not equal demand reduction.
Historically, every improvement in computing efficiency has ultimately produced *more* computing demand, not less—the famous Jevons Paradox: when coal became more efficient to burn, people burned more coal. DeepSeek made training cheaper, meaning far more teams could afford to train large models. A game once reserved for tech giants opened up to mid-sized companies. Result: compute demand rose instead of falling.
More crucially, DeepSeek did not solve inference. Training is one-time; inference is ongoing. Every ChatGPT query requires H100-class inference compute, and as AI applications explode, inference demand is growing at a staggering rate. So after a brief dip, H100 prices turned upward—and climbed higher than before.
Why Old Cards Are Worth More
Why would a four-year-old H100 be more valuable than three years ago?
In a market of extreme scarcity, availability—not age—sets the price. For an AI company, lacking H100s means being unable to train models or serve customers. In this hunger game, an old card is still a meal when the restaurant has only one dish on the menu.
Moreover, for AI training and inference specifically, the performance gap between old and new H100s is smaller than expected: the core architecture and specifications are the same as when it launched. As long as the card runs, it creates value.
Rebuilding the Asset Model
This is changing the business logic of data centers and cloud computing. Traditionally, GPUs were booked under IT depreciation models assuming 20–30% annual value loss and near-zero residual value after five years. The H100's value curve looks nothing like that.
Some shrewd operators now see the H100 as "digital gold" rather than an electronic device—its value derives from scarcity, not freshness. Until the supply shortage resolves (which may take years), its real value may hold steady or even appreciate.
This creates an arbitrage opportunity: buy H100s today, rent them out, and resell years later—the rental income plus the resale price may exceed your initial outlay. Such "negative depreciation" assets are extremely rare in traditional industries.
What It Means for Ordinary Users
Why should this matter if you don't run a data center? Because it reveals a底层 logic of the AI wave: compute is the new oil. Just as the industrial revolution ran on coal and the information revolution on chips, the intelligence revolution runs on compute—and whoever controls compute controls leverage.
What Comes Next?
Will H100 prices rise forever? Of course not. The imbalance will eventually ease: NVIDIA's next-generation Blackwell architecture is ramping to volume production, AMD's MI series is catching up, and cloud giants are designing their own chips. In two to three years, the compute market may reach a new equilibrium.
Until then, the H100 story continues. It reminds us that during periods of violent technological change, traditional business logic can be upended, with new opportunities and new risks coexisting. That four-year-old graphics card is telling a classic story of scarcity, demand, and value through its price curve—and we are living through its most dramatic chapter.