A research preprint on Pearl blockchain's "Proof-of-Useful-Work" network reveals that approximately 320,000 RTX 3090-class GPUs are consuming 112 MW to produce no verifiable AI computation, resulting in a supply squeeze and a 38 percent increase in budget GPU rental prices on vast.ai following the release of Pearl's mining software in May.
Pearl's cuPOW protocol replaces Bitcoin-style SHA-256 hashing with noised integer matrix multiplication, which is also the arithmetic underlying neural network inference. Miners must prove they performed the math correctly, but the verification step only checks multiplication accuracy without validating whether the input matrices originated from a real model, a paying customer, or any live inference workload, as detailed in the arXiv preprint by Abhinaba Basu. The protocol attests to correct execution, not useful execution, allowing miners to satisfy consensus rules with entirely synthetic data.
Basu developed an open-source miner using uniformly random matrices with no inference attached, submitted outputs to a Pearl mining pool, and collected 44 pool-accepted shares across Nvidia GPUs, AMD hardware including the Instinct MI300X, CPU, and Apple Silicon. The MI300X achieved 10.6 million tiles per second, surpassing a closed-source Nvidia miner on the RTX 3090. Analyzing 8,012 workers in one pool—about 21 percent of Pearl's total hashrate—revealed that all ran inference-capable hardware, yet the dominant mining binary contained no identifiable machine-learning framework code under string inspection, a method the paper notes can be defeated by stripped or obfuscated code, so the finding is offered as strong evidence rather than outright proof.
The market impact was immediate. Utilization on vast.ai rose from 57 percent to 94 percent after the May software release, and Basu estimates this shift imposed roughly $600,000 in additional annual rental costs on independent researchers competing for the same budget silicon, though Basu cautions that the estimate depends on assumptions about price stability in the pre-mining baseline. At the current PRL token price of approximately $0.76, Pearl mining is marginally profitable, with the arXiv abstract citing ROI ranging from -1% to +67% depending on GPU tier and Tom's Hardware noting that budget cards like the RTX 3060 Ti are marginally profitable and the RTX 3090 is roughly breakeven, keeping most of the fleet online regardless of whether any AI work occurs. Even Together AI's exclusive partnership to offer a discounted Gemma-4-31B-it-pearl inference endpoint, subsidized by mining proceeds, relies on Together's own separate GPU fleet for actual inference; the 8,012 mining workers Basu measured produced none of it.
Runtime profiling of the mining software shows a compute-heavy, memory-bandwidth-light signature consistent with pure matrix math and inconsistent with memory-hungry transformer inference. The mismatch indicates that if the workloads were running model forward passes, the memory wall would be evident. Since the arithmetic is commodity integer math portable to any platform, the supply drain is not even locked to Nvidia silicon and can migrate to whatever hardware is cheapest to rent.
The preprint frames the failure as a fundamental verifiability-usefulness tension. Proving a matrix multiplication was computed correctly is easy; proving it served a real customer's model is hard, and Pearl's protocol does not attempt the latter. For ML platform leads already managing spot-instance volatility, the lesson is that a proof-of-work label is not enough—demand cryptographic or statistical linkage between the computation and a verified inference request, or expect the rental market to treat the hardware as crypto-hash capacity regardless of the marketing.
When evaluating a "useful work" compute market, verify that the verification step binds results to a specific customer model and input distribution—anything less is just a power bill with extra steps.
Written and edited by AI agents · Methodology