Pearl Blockchain’s 320,000 RTX 3090 GPUs Burn 112 MW on Zero Useful AI

A research study claims that the Pearl blockchain network is consuming massive amounts of GPU power with no useful AI output. The network's 320,000 RTX 3090-class GPUs burn an estimated 112 megawatts of electricity doing random matrix math, according to the preprint.

Study shows network performs no real AI computation

Pearl's consensus protocol, called cuPOW, verifies that a matrix multiplication was performed but does not check whether the computation comes from a real AI workload. Researcher Abhinaba Basu demonstrated this by mining with random matrices and receiving pool-accepted shares, proving that no actual AI work is required.

The study estimates Pearl's network runs at roughly 24 exahashes per second. That hashing power is equivalent to about 320,000 RTX 3090-class GPUs drawing 112 megawatts. The paper describes the network as performing "zero useful AI computation."

After Pearl's mining software went public in May, GPU rental prices on the marketplace vast.ai jumped 38%, and utilization climbed from 57% to 94%. Basu also mined the first Pearl shares on non-Nvidia hardware, driving an AMD Instinct MI300X at 10.6 million tiles per second, faster than the Nvidia miner on an RTX 3090. The study estimates that independent researchers face an extra $600,000 per year in rental costs due to Pearl mining, though the researcher cautions that figure depends on assumptions about price stability.

The protocol enables useful work in theory, but it does not require it in practice, the research preprint notes. Pearl's current price is about 0.76 USD.