Idle GPUs are the new zombie servers

The classic zombie is a forgotten VM idling at 2% CPU. Its GPU-era cousin is louder: a single on-demand H100 runs around €2,100 a month, and three of them left running over a weekend is a four-figure surprise.

Why GPU clouds are harder to watch

Neoclouds like Lambda and Modal give you the same NVIDIA hardware at a fraction of hyperscaler prices — which is exactly why they attract experimentation, and exactly why spend creeps. Many of them expose an instance API but no utilization metrics, so you can't ask "is this GPU actually busy?" the way you'd ask a VM about its CPU.

When you can't measure utilization, run-rate is the signal. A running GPU with no owner is a question worth asking every day.

What to watch instead

Without a utilization number, the honest approach is inventory plus price: enumerate the running instances, price each at its real hourly rate projected to a month, and surface the total. "You have three running H100s at €6,300/mo — intentional?" is a more useful question than any dashboard average.

Pair that with change detection — a new instance that appeared since yesterday, a run-rate that jumped — and the sprawl that used to accumulate silently becomes a daily, answerable prompt instead of a monthly regret.