Pearl blockchain's 320,000 GPUs burn 112 megawatts on random matrix math, not useful AI, study finds

At a glance:

• Pearl blockchain's mining network consumes an estimated 112 megawatts across 320,000 RTX 3090-class GPUs while performing only random matrix multiplications, not useful AI work, according to a new research preprint.
• Budget GPU rental prices on vast.ai jumped roughly 38% and utilization climbed from 57% to 94% after Pearl's mining software went public in May, costing independent researchers an estimated $600,000 per year in additional rental costs.
• The study demonstrates that Pearl's cuPOW verification only confirms correct matrix multiplication, not whether inputs come from real AI workloads, and shows mining works on AMD Instinct MI300X, Apple M2, and CPUs with no vendor lock-in.

What the study found

The research preprint titled "The Usefulness Gap in Proof-of-Useful-Work" by Abhinaba Basu provides the first empirical measurement of a deployed Proof-of-Useful-Work system. Basu estimates Pearl's network runs at roughly 24 exahashes per second, the equivalent of about 320,000 Nvidia RTX 3090-class GPUs drawing an estimated 112 megawatts of power. Despite this massive compute footprint, the paper concludes the network produces "zero useful AI computation" because miners are grinding through random matrix multiplications that merely take the shape of AI math without any actual inference or training workloads attached.

To demonstrate the gap, Basu built a custom miner that feeds the network uniformly random matrices with no inference attached and submitted the output to a mining pool. The paper reports 44 pool-accepted shares on Nvidia and AMD hardware, with the same miner also benchmarked on a server CPU and Apple Silicon through Metal compute shaders, plus an on-chain payout earned by running the standard mining software unmodified. If random numbers collect rewards as readily as genuine AI work, the argument runs, the network cannot tell the two apart, and miners have every incentive to skip the AI part entirely.

How Pearl's cuPOW works and the verification gap

Pearl swaps Bitcoin's SHA-256 hashing for a scheme it calls cuPOW, which asks miners to compute noised integer matrix multiplications and prove they did so correctly. That operation is the same arithmetic that underpins neural network inference and training, which is the foundation of Pearl's pitch that mining and AI compute can be one and the same job. The problem, according to the study, is that the protocol's verification step only confirms that the multiplication was performed correctly. It never checks whether the input matrices came from a real model, a paying customer, or any AI workload at all.

Basu also analyzed 8,012 workers in a single pool, representing about 21% of Pearl's hashrate, and found that all of them ran hardware capable of AI inference. Yet the dominant mining binary contained no identifiable code for any machine-learning framework. That binary analysis relies on string inspection, which the paper notes can be defeated by stripped or obfuscated code, so the finding is offered as strong evidence rather than outright proof. Runtime profiling pointed the same way, with the miners showing heavy compute use and light memory-bandwidth use, a signature consistent with pure matrix math and inconsistent with the memory-hungry behavior of transformer inference.

Impact on GPU rental market and researchers

For GPU buyers, the resource angle is the part that stings. The study attributes a roughly 38% jump in budget GPU rental prices on the marketplace vast.ai to Pearl mining, with utilization climbing from 57% to 94% after the mining software went public in May. Using a difference-in-differences comparison against pricier datacenter cards, Basu estimates around $600,000 per year in additional rental costs borne by independent researchers who compete for the same cheap hardware. However, he cautions that the figure depends on assumptions about how stable prices were beforehand.

At PRL's recent price near $0.76, the paper calculates that mining is marginally profitable on budget cards such as the RTX 3060 Ti and roughly breakeven on an RTX 3090. The rental price surge affects a range of budget GPUs commonly used by academic labs and small startups, potentially crowding out genuine AI research workloads in favor of cryptomining that mimics AI computation without delivering its utility.

Cross-hardware mining and vendor lock-in

The work also chips away at the assumption that Pearl mining is an Nvidia-only affair. Basu reports the first Pearl shares ever mined on non-Nvidia hardware, driving an AMD Instinct MI300X at 10.6 million tiles per second, faster than the closed-source Nvidia miner managed on an RTX 3090. The same workload was also benchmarked on a server CPU and on an Apple M2 through Metal compute shaders. Because the computation is commodity integer arithmetic, the paper argues there is no vendor lock-in and no technical reason for the work to remain on any one company's silicon.

This cross-platform capability undermines the narrative that Pearl's mining is inherently tied to Nvidia's CUDA ecosystem. The researcher's miner implementation demonstrates that the cuPOW algorithm can run efficiently on AMD's CDNA architecture, Apple's Metal framework, and standard x86 CPUs, suggesting the network's hardware diversity could expand rapidly if miners optimize for different platforms.

Together AI partnership and financial arbitrage claim

Pearl may have a ready response, but the study takes it on directly. Together AI, which announced an exclusive partnership in May, framed the deal as letting "every GPU cycle powering AI training and inference" also mint the PRL token, and it now offers a discounted Gemma-4-31B-it-pearl inference endpoint subsidized by mining proceeds. Basu counters that this is financial arbitrage rather than useful mining, because Together AI's own GPUs perform that inference separately from the mining network, with PRL revenue used to trim the endpoint's price.

The 8,012 mining workers Basu measured, he says, produced none of that inference themselves. The partnership appears to create a subsidy loop where mining rewards fund cheaper API access, but the mining computation itself remains disconnected from actual AI workloads. This distinction is critical for evaluating whether Proof-of-Useful-Work systems can genuinely merge blockchain security with productive compute.

What this means for Proof-of-Useful-Work

The study's conclusion is not that Proof-of-Useful-Work is impossible, but that Pearl's current design leaves a major enforcement gap. The protocol enables useful work in theory, but it does not require it in practice. That leaves Pearl in an uncomfortable middle ground — it performs real computation, but according to the paper, the network currently has no way to prove that the computation is useful AI work rather than cryptomining with AI-shaped math.

Future PoUW designs would need to bind the verification step to externally verifiable AI workloads, such as cryptographic commitments to model inputs and outputs or zero-knowledge proofs of correct inference execution. Until then, the economic incentive for miners to substitute random matrices for real AI work will likely persist, and the GPU rental market will continue to feel the pressure from mining demand that masquerades as AI compute.