A statement made by Clay Magouyrk during $ORCL’s latest earnings call directly addresses one of the most common misconceptions in the debate surrounding AI infrastructure:

“All GPU capacity that came up for renewal in Q1 was either renewed or resold at a 20% premium compared to previous contracts. Most of these GPUs are four years old or older.”

A common market narrative was that a GPU has an economic lifespan of four years, after which it becomes junk. This conflates technical obsolescence with economic lifespan.

Technical lifespan, technological lifespan, and economic lifespan are not the same thing.

The fact that a GPU was manufactured four years ago does not mean it is useless today. It is necessary to distinguish between three concepts:

-- Technical lifespan: Is the chip physically operational?
-- Technological lifespan: How does its performance and energy efficiency compare to the new generation?
-- Economic lifespan: Does this capacity still generate profit for the customer?

When a new generation is released, the old GPU does not simply stop working. It merely becomes less preferred for certain tasks. This does not mean its value drops to zero. The data provided by Oracle demonstrates exactly this.

The key signal: renewed or resold—and at a 20% premium.

The critical part of the sentence is “renewed or resold.” All capacity reaching its renewal date either remained with the existing customer or was sold to another. Moreover, most of these GPUs are four years old or older.

If the “dies in four years” thesis were true, Oracle would have accumulated unsellable, outdated capacity. The opposite occurred. The initial contract ends, but the GPU’s economic life does not.

An even stronger indicator is the pricing. This capacity was renewed or resold at a price 20% higher than the previous contracts. In other words, the point isn't just that it “still works.” Customers are willing to pay a premium for capacity that is over four years old.

GPU utilization in the same quarter stands at 97.9%. This points more to price appreciation in a tight market than to an idle, outdated fleet.

New GPUs don't wipe out the old; they take on the most demanding workloads.

Not all GPUs perform the same tasks. The latest generation offers significant advantages for frontier training and workloads requiring peak performance. However, not every task in a data center involves training the world's largest model.

Inference, fine-tuning, embedding, enterprise applications, smaller models, batch processing, various HPC workloads—for a significant portion of these, there is no requirement for the newest, most expensive GPU.

This creates a tiered structure within the fleet:

-- New GPU → handles the most demanding workload.
-- Previous generation → shifts down a tier.
-- Older units → move to lower-intensity tasks or other customers.

The new generation doesn't automatically render older capacity obsolete;
it shifts its use case. Customers don't always need maximum performance. Sometimes, the need is simply finding sufficient compute for the job at a reasonable price. A four-year-old GPU might not be ideal for frontier training, but it can still generate revenue through inference or enterprise applications.

This is where the criticism regarding AI CAPEX falls short.

The most serious objection to AI CAPEX is that investments are massive, GPUs become obsolete quickly, and companies cannot monetize the assets for long enough.
Technological obsolescence is real; generations arrive rapidly, and energy efficiency gaps are significant. However, concluding that "economic value drops to zero after four years" is incorrect.

Oracle isn't discussing a theoretical model; it is referring to actual capacity that is due for renewal. Much of it is over four years old. Yet, it is being renewed, resold, and priced at a premium of approximately 20%. That is why calculating the return on investment based solely on the initial contract is flawed. A GPU can continue generating revenue through a chain like this:

initial contract → renewal → second customer → different workload

For a company like Oracle that leases capacity, the real question isn't "How many years will this chip run?" but rather, "For how many years will this asset generate revenue?" This quarter provides a strong real-world answer to that question.

The situation is similar for NVIDIA: if older GPUs aren't rendered obsolete by the new generation, then global AI compute capacity doesn't simply reset with every new generation. New capacity is added on top, while older capacity shifts down the stack. The story isn't about scrapping the fleet every 3–4 years; it’s about adding new layers to a growing total compute pool.

We shouldn't overstate this.

You can't conclude from this data that a GPU retains the same economic value for 10 years. The performance and performance-per-watt advantages of newer generations eventually put older cards at a disadvantage, particularly in frontier model training. Energy costs can also shorten their economic lifespan. Moreover, this reflects a single quarter’s refresh cycle; we lack details on the product mix, contract structures, and power costs.

That isn't the point of the debate, though. The real question is: does a GPU’s value drop to zero once it turns four years old? Current data shows the answer is no.

The real takeaway isn't about GPU sentimentality; it's about demand.

I don't interpret this simply as "old GPUs are actually pretty good." The more important message is this:

Demand for AI compute remains so tight that, even with the new generation on the market, capacity that is four-plus years old is changing hands at a premium rather than a discount.

If there were an oversupply, we would expect prices to fall. Instead, Oracle is seeing a 20% premium. This serves as real-world evidence that both the economic lifespan of a GPU may be longer than we thought and the capacity gap has not yet closed.

Four years from now is not the date the GPU dies; it may be the period when its use case begins to shift in the face of the next generation. Magouyrk’s statement is significant because it moves the debate regarding CapEx and depreciation from the realm of theory and anchors it to actual contracts.