AI cryptomining network’s 320,000 RTX 3090-class GPUs allegedly burn 112 megawatts of power on ‘zero useful AI computation’ — GPU rental costs jump 38%, but Pearl’s cards are doing random matrix math, study claims
Publish Date: 2026-06-14 07:30:00
Source Domain: www.tomshardware.com
Here is a concise summary of the key points from the article using an unordered list:
– A new research preprint suggests Pearl’s blockchain system is performing no actual AI computation, despite attracting a GPU mining rush recently.
– The study estimates Pearl’s network hash rate at around 24 EH/s, equivalent of drawing power equivalent to 320,000 RTX 3090 GPUs, yet produces “zero useful AI computation.”
– Researchers found a 38% increase in budget GPU rental prices on vast.ai after Pearl mining software was released in May 2023.
– Instead of real AI inference or training, Pearl seems to be performing random matrix multiplication that resembles AI workload.
– Pearl’s proof-of-work scheme (cuPOW) checks multiplication correctness but does not authenticate whether it’s connected to real AI work.
– Abhinaba Basu created a miner feeding the network random matrices to show the gap between claimed useful work and what is actually happening.
– The study found that the majority of miners ran hardware capable of AI inference but no identifiable machine-learning framework code.
– The increase in GPU usage reduced availability, thus making computing resources less accessible for independent researchers.
– Although Pearl’s claim is not disproven by the study, its protocol seems ineffective at enforcing true useful computation.
