Pearl AI network accused of faking workloads on 320,000 RTX 3090-class GPUs

A research report alleges that Pearl, a major GPU rental network, is running fake workloads across 320,000 RTX 3090-equivalent cards. The study claims these GPUs perform "zero useful AI computation" — just random matrix operations to look busy while burning 112 megawatts. Tom's Hardware broke the story.

Pearl's fleet represents roughly 6% of all high-end GPUs currently rented for AI training. If the allegations hold, the operation is inflating demand metrics without delivering real compute. GPU rental prices have jumped 38% in the past quarter, a spike the study attributes in part to Pearl's artificial scarcity.

This matters if you're hunting used RTX 3090s or 4090s. A network this size dumping cards (or getting shut down) could flood the market. It also explains why rental costs for local fine-tuning have gotten brutal — you're competing with phantom workloads.

If you're gaming on a 3090 or 4090, check your settings to make sure you're getting every frame those cards can push. No sense leaving performance on the table while the rental market goes sideways.