A new preprint alleges that an AI cryptomining network called Pearl is burning 112 megawatts of power across 320,000 RTX 3090-class GPUs — yet performing no verified useful AI computation. The study claims the cards are executing random matrix math, not productive inference or training.
The network's GPU rental costs have reportedly jumped 38%, squeezing legitimate AI researchers and startups already grappling with hardware shortages. If verified, the finding would represent a massive misallocation of computing resources in a market where GPU access is increasingly scarce.
The preprint draws on network analysis and power consumption estimates, though its claims have not been peer-reviewed. The 112 MW figure would rival the energy draw of a small data center cluster, while the purported fleet of 320,000 GPUs dwarfs many enterprise AI deployments.
Pearl has not publicly responded to the study. The allegations, if substantiated, could intensify scrutiny of cryptomining operations that disguise themselves as AI ventures. Regulators and cloud providers may face pressure to distinguish genuine AI workloads from arbitrary computation.
Some industry observers caution that the study's methodology remains unverified and may conflate idle processing with deliberate obfuscation by the network.