Pearl's AI mining pitch faces a 112 MW usefulness test

Why it matters

Pearl is trying to make proof-of-work useful by tying mining to AI matrix math. The new preprint challenges whether its live network verifies usefulness at all.

Omri Weinstein, the theoretical computer scientist behind Pearl Research Labs, has been trying to solve one of crypto's oldest credibility problems: how to make proof-of-work useful. A June 3 preprint now argues Pearl has built the hard part of that system - a live GPU-secured blockchain - without proving the work is useful AI computation.

Tom's Hardware reported Sunday on a preprint by Abhinaba Basu that estimates Pearl's network is running at roughly 24 EH/s, equal to about 320,000 RTX 3090-class GPU equivalents and 112 MW of power draw, while producing what the paper calls "zero useful AI computation." The paper also claims budget GPU rental prices rose 38% and utilization jumped from 57% to 94% after Pearl mining software became public.

Those are Basu's estimates, and the paper is a preprint, not a peer-reviewed finding. But the critique goes directly at Pearl's founding premise. Pearl does not pitch itself as another GPU marketplace or another token reward layer. In its whitepaper, Pearl says it replaces Bitcoin-style artificial hashing with matrix multiplication, the core operation behind modern AI training and inference, so GPU cycles can both secure a blockchain and subsidize AI workloads.

Weinstein is not a crypto mascot bolted onto a token launch. His academic page lists him as an associate professor in theoretical computer science at the Hebrew University of Jerusalem, on leave from Columbia University, with a Princeton PhD and research interests spanning data structures, information theory, optimization and high-dimensional search. His CV also lists a stint as chief scientist at VAST Data. That background is why Pearl's claim has attracted attention: it is a technically serious answer to a market question that has mostly been answered with slogans.

The new paper says the implementation leaves a gap large enough for miners to drive through.

The useful-work claim turns on verification

Pearl's technical move is called cuPOW. The protocol asks miners to perform noised integer matrix multiplications and prove they did the computation correctly. Pearl's argument is that matrix multiplication is not arbitrary busywork: it is the same arithmetic that dominates neural network workloads. If an AI inference or training job can be wrapped in the proof-of-work process, Pearl miners can earn PRL while doing work a customer already wanted done.

Basu's critique is narrower and more damaging than a general attack on proof-of-useful-work. The paper argues that Pearl can verify that matrix multiplication happened, but not that the matrices came from a real model, a paying customer, or any inference or training job. In Basu's test, a miner fed the network random matrices with no AI workload attached and still produced 44 pool-accepted shares across NVIDIA and AMD hardware, according to the preprint.

That matters because miners follow incentives, not theses. If random matrices earn the same reward as customer inference, the rational miner has no reason to take on the coordination, data handling, latency and reliability costs of serving real AI work. Pearl's own whitepaper anticipates a version of this problem, saying day-one token prices can depend on future expectations while many holders of useful work are not yet set up to mine Pearl during their workloads. Basu's paper argues the live system has not closed that gap.

The evidence is empirical but not absolute. Tom's Hardware notes that Basu analyzed 8,012 workers in a single pool, representing about 21% of Pearl's hashrate, and found that the workers had inference-capable hardware while the dominant mining binary showed no identifiable machine-learning framework code.

The Together AI partnership is proof of demand, not proof of the whole network

Pearl's strongest commercial validation so far is its May 15 partnership with Together AI. Together said it was introducing a Gemma-4-31B-it-pearl inference endpoint discounted by more than 25%, with the lower price offset by the future value of PRL emissions. In the announcement, Weinstein said Pearl lets GPU cycles powering AI training and inference also produce a proof-of-work digital asset.

That partnership matters. It shows Pearl is not merely publishing a paper and waiting for miners to appear. Together is a credible AI infrastructure company, and an inference endpoint is the kind of integration Pearl needs if PRL emissions are going to reduce price-per-token instead of just funding speculative mining.

But it does not answer Basu's broader claim. A discounted endpoint can demonstrate that one partner is experimenting with the subsidy model. It does not prove the wider Pearl mining network is routing its work into real customer jobs. The distinction is central: Pearl's protocol can be valuable if useful AI workloads are the primary source of mining work, and much less defensible if the network mostly pays miners to generate matrices for their own sake.

The GPU-market spillover is the immediate operator problem

The preprint's most concrete market claim is not the headline power figure. It is the alleged effect on cheap GPU supply. Basu attributes a 38% increase in budget GPU rental prices on Vast.ai, and a utilization rise from 57% to 94%, to the public release of Pearl mining software. Tom's Hardware reports that Basu used a difference-in-differences comparison against pricier datacenter GPUs and estimated roughly $600,000 per year in added rental costs for independent researchers competing for the same low-cost hardware.

That number is small next to hyperscaler capex, but meaningful in the long tail of AI research, where rented RTX-class cards are often the difference between running an experiment and shelving it. It also exposes the tension in decentralized compute networks: if token incentives are high enough, they can mobilize idle hardware quickly. If the incentives are not tied tightly to customer work, they can also bid hardware away from researchers and developers without adding useful supply.

Pearl is not alone in trying to connect crypto incentives to AI infrastructure. Akash, Render, io.net, Aethir, Gensyn, Bittensor and others are all trying to turn distributed compute, models or ML work into networked markets. Pearl's sharper claim is that chain security itself can be bound to matrix multiplication, with token emissions making inference cheaper. Basu's paper attacks that claim at the seam between cryptographic proof and economic usefulness.

Pearl has built a real network. Now it has to prove what the network is doing.

Pearl's public footprint is already more substantial than a landing page. Its GitHub monorepo contains the full node, wallet, SPV light client, zero-knowledge proof-of-work components and supporting tools. Pearl also operates a public block explorer for its mainnet. The whitepaper lists a 2.1 billion PRL coin supply and a 194-second block interval.

That makes the critique more important, not less. Pearl has moved from theory into production enough that the practical behavior of miners now matters. The original April 2025 paper by Ilan Komargodski and Weinstein framed proof-of-useful-work from arbitrary matrix multiplication as a way to replace wasteful hashing with AI-relevant computation. Basu's preprint says the deployed system has not yet verified the part users actually care about: whether the work was useful before it became a proof.

For Weinstein, the path forward is not to prove that matrix multiplication is central to AI. That case is already obvious. The harder founder problem is product and mechanism design: make useful AI work the dominant, verifiable and economically preferred way to mine Pearl, rather than a story layered on top of a profitable random-matrix race.

Pearl's most generous reading is that this is an early network bootstrapping problem. The whitepaper itself discusses a day-one gap where useful-work holders may not yet be ready to mine while doing their real workloads. The less generous reading is that the gap is structural: verifiability rewards what the protocol can check, and the protocol can check arithmetic correctness more easily than economic usefulness.

That is the test Pearl now faces. If Pearl can show, at network scale, that PRL emissions are attached to actual inference and training rather than random compute, Weinstein will have a credible answer to proof-of-work's energy critique. If not, Pearl risks becoming a technically elegant way to recreate the exact waste it set out to escape.