We spent $50 to measure Pearl's "AI mining" – 320K GPUs produce zero AI (arxiv.org)

A recent study examined Pearl's Proof-of-Useful-Work (PoUW) protocol, highlighting significant shortcomings in its purported capabilities. Researchers measured the performance of Pearl's 24 EH/s network, which comprises about 320,000 GPUs, and found that it produced no substantial AI computation despite high-profile endorsements. Their analysis revealed that while the network utilized powerful inference-capable hardware, the dominant mining software lacked any inference functionality, allowing for random matrix acceptance that undermined verification. This has raised concerns as GPU rental prices surged by 38%, displacing legitimate research workloads in favor of unproductive mining activities. This discovery sheds light on the broader implications of the verifiability-usefulness tension in PoUW systems, revealing a stark mismatch between the network's resource consumption and its actual output. As miners faced negative returns on investment due to low token prices, the findings challenge the legitimacy of claims surrounding PoUW's dual purpose of securing blockchain and performing AI tasks. These empirical measurements provide critical insights into the economic consequences of current mining models, making it an essential study for the AI/ML community that reflects on the viability of integrating computational efficiency and blockchain technology.