To The Moon

⚡ Old GPUs Keep Printing Money While New Ones Must Be Replaced? The Economics of $NVDA Data Centers Are Quietly Changing

When people talk about AI infrastructure, the discussion usually focuses on one simple idea:

More compute → more powerful GPUs → higher prices.

But the real shift in the industry isn’t just about performance.

It’s about hardware lifecycle economics.

Take a classic example: the $NVDA V100.

This GPU was released nearly a decade ago. When it first entered data centers, the typical depreciation cycle was about three years.

By traditional accounting logic, those machines should already be retired.

But reality looks very different.

Across many hyperscale data centers, V100 GPUs are still running at full capacity.

The reason is simple.

From an accounting perspective, these GPUs are already fully depreciated. Their book value is effectively zero.

Yet they continue to generate revenue.

Operators only need to pay NVIDIA for extended support and maintenance, and the hardware can continue running.

In other words:

The capital cost has already been written off,

but the compute power keeps producing income.

This became one of the most attractive economic characteristics of GPU-based data centers over the past decade.

However, the next generation of AI chips is operating under very different conditions.

With systems like $NVDA H100 and GB200, the operating environment has fundamentally changed.

These chips run:

24/7 at extremely high utilization

Under massive power draw

At the edge of cooling limits

With very limited repairability

In that kind of environment, when something fails, replacement is often the only realistic option.

That’s one reason NVIDIA now provides hyperscale customers with extended warranty and replacement guarantees.

In practice, this means NVIDIA is partially absorbing reliability risk.

If a chip fails, the replacement cost may fall on NVIDIA rather than the customer.

This creates a different operational logic for data centers.

Industry experts point out that continuing to run GPUs after warranty expiration can actually become more expensive.

Why?

Because once the warranty expires:

Replacement protection disappears

Failure risk rises

Service stability becomes harder to guarantee

As a result, many operators treat out-of-warranty GPUs as buffer capacity rather than primary production compute.

In other words:

Mission-critical compute must remain inside the warranty ecosystem.

Now consider the scale of modern AI data center construction.

Building a gigawatt-scale AI data center from scratch can cost roughly $30B–$35B.

The cost breakdown typically looks like this:

About 80% allocated to IT infrastructure

(GPUs, CPUs, memory, networking)

About 20% allocated to facilities

(power, cooling, buildings)

Looking only at GPU hardware, the numbers escalate quickly.

A GB200 node can cost around $60,000.

A rack with roughly 10 nodes approaches $600,000.

And some hyperscale deployments push density even higher.

Meanwhile, supply chains remain tight.

Because of that, some industry estimates suggest total AI data center capex could still rise by another ~5%.

When people talk about AI, they often focus on model intelligence.

But the real constraint shaping the industry may be something else entirely:

capital intensity.

AI is not a light-asset business.

In fact, it may be one of the most capital-intensive technology cycles of the past several decades.

Within that system, GPUs are no longer just pieces of hardware.

They have effectively become consumable infrastructure.

And as long as global demand for compute keeps rising, the economic model around GPUs will continue expanding with it.