Comment on Did The AI Bubble Go From A Pin Prick To A Pop This Week? by dickie

Parent Post: Did The AI Bubble Go From A Pin Prick To A Pop This Week? · In Reply To

dickie · 11/1/2026, 9:11:13 PM · permalink

**The real reason for replacing older AI hardware is efficiency, not simply speed.**

An NVIDIA A100 GPU consumes up to 400 watts, while an H100 can consume up to 700 watts. Yet the H100 can deliver substantially more AI computation per watt. Newer Blackwell hardware improves efficiency further.

The important measurement is **electricity consumed per unit of useful computation**, not electricity consumed per GPU.

Consider the decommissioned NVIDIA supercomputer listed on eBay for $59.7 million. It contains 9,600 A100 GPUs, 4,200 AMD EPYC processors and 300 server racks.

We estimated approximately 9 megawatts of continuous electrical demand, including cooling and facility overhead.

Using Maine's historical industrial electricity-price increases over the previous five years, the projected electricity bill alone could reach **$77.4 million over the next five years**.

Now consider what newer, more efficient equipment could accomplish while performing the same workload.

| System efficiency | Five-year electricity cost | Savings |
|---|---:|---:|
| Original A100 system | $77.4 million | $0 |
| 2 times more efficient | $38.7 million | $38.7 million |
| 3 times more efficient | $25.8 million | $51.6 million |
| 5 times more efficient | $15.5 million | $61.9 million |
| 10 times more efficient | $7.7 million | $69.7 million |

These are illustrative scenarios assuming equal computational output and proportional reductions in total facility electricity consumption.

NVIDIA has reported up to 15 times greater energy efficiency for certain AI inference workloads using its HGX B200 platform compared with HGX H100. Actual improvements depend on the workload and system configuration.

**This changes the economics of buying used AI hardware substantially.**

A buyer might save tens of millions of dollars purchasing decommissioned equipment, only to spend those savings on electricity over the following several years.

And electricity prices aren't standing still.

**There are additional advantages to newer hardware:**

- Fewer servers may be needed for the same computational workload.
- Reduced electricity consumption per unit of computation.
- Lower cooling requirements for equivalent output.
- Less floor space and electrical infrastructure.
- Reduced maintenance requirements.
- Greater computational capacity within existing power constraints.

**The economics become even more significant when electricity prices increase.**

Using Maine's historical industrial electricity-price growth of approximately 10.9% annually, the cost of operating inefficient hardware increases considerably over time.

For this particular supercomputer, the estimated five-year costs would be:

- Purchase price: $59.7 million.
- Electricity: $77.4 million.
- Combined: $137.1 million.
- Building, maintenance, networking and infrastructure: Additional expenses.

And after five years, the buyer would own equipment that was already obsolete enough to have been decommissioned before the purchase.

**The real question isn't whether newer AI hardware costs more to purchase. It's whether older hardware costs more to own and operate.**

When electricity expenses can exceed the purchase price of the equipment, computational efficiency becomes one of the most important factors in determining total cost of ownership.

A newer system might draw more power at full capacity, but if it accomplishes several times as much useful computation, it can be considerably cheaper to operate for the same workload.

**That $59.7 million supercomputer might be a bargain to purchase and a financial disaster to operate.**

But at least the seller offers free shipping.

1 bump

Signature