Pearl (PEARL) sells itself on one idea: the computation that secures the chain should be worth something beyond the chain. That is the promise of proof of useful work. But is Pearl mining useful today, in practice? Two research papers published in June 2026 looked at that question from different angles, and the protocol itself has been changing in response. This post lays out the debate as fairly as we can, as of November 5, 2026. We build a Pearl wallet, so read it with that in mind.
What “useful” is supposed to mean
Pearl’s mining builds on the paper by Komargodski and Weinstein (2025, arXiv 2504.09971). The miner performs noised matrix multiplications, and a zero-knowledge proof (built on Plonky2 and STARKy) shows the work was done correctly. The key property is that the miner chooses the matrices. If those matrices come from a real AI model, the GPU is running inference and mining at the same time, with little extra cost. The official miner is built on vLLM, an AI inference engine, to make that possible.
Notice what the proof checks: that the multiplication is correct. It does not check where the matrices came from. That gap is the center of the whole debate.
The critique: “The Usefulness Gap”
In June 2026 Abhinaba Basu published “The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl’s cuPOW Protocol” (arXiv 2606.04819). Its main claims, summarized:
- Zero useful output at the time. The paper claims the network’s hashpower produced no useful AI computation. It counted 8,012 workers on inference-capable hardware, but says the dominant mining software contained no inference code.
- Random matrices pass. The protocol accepts random matrices by design. The author reports 44 pool-accepted shares from an open-source miner, and argues that the distribution checks meant to spot fake work can be defeated.
- Commodity math. The computation was integer arithmetic that could run on any hardware, so nothing tied it to AI-grade accelerators.
- Thin margins. Mining was near breakeven, with ROI estimated from −1% to +67% depending on GPU tier.
- Side effects on the GPU market. After the miner was released, the paper reports that rental prices on budget GPUs rose 38% and utilization went from 57% to 94%.
The fair reading: this is an empirical snapshot of the network in its early months, not a proof that useful mining is impossible. But it is a serious point. If the cheapest way to mine is to multiply random numbers, most miners will do that, and the “useful” part stays theoretical.
The economics view: Rafael Pass
Also in June 2026, Rafael Pass published “The Economics of Proof-of-Useful-Work” (arXiv 2606.06700). Instead of measuring the live network, it builds an equilibrium model. Compute can go to pure mining, pure inference, or “duplex” work that does both at once. Depending on conditions, the paper describes three outcomes:
| Regime | What happens |
|---|---|
| Bitconia | Mining looks like classic proof of work: compute is spent on mining alone |
| Fortessia | Duplex work replaces pure mining: more security, the same useful output |
| Duplexia | Block rewards subsidize inference: inference gets cheaper and there is more of it |
Pass also argues that useful work does not make attacks cheap: in the model, the cost of attacking the chain stays tied to the block reward. Read together, the two papers are less contradictory than they look. Basu describes a network sitting in something like the first regime. Pass describes what has to be true for it to move into the other two.
How the protocol is responding
Pearl changes its rules through Pearl Improvement Proposals (see how PIPs work). Several are aimed, directly or indirectly, at making real AI workloads the natural way to mine:
- PIP-2 (Final): Grouped-GEMM proofs for mixture-of-experts models, so a miner running a modern MoE model does not repeat commitment and noising work for every expert. Details in MoE mining and PIP-2.
- PIP-3 (Draft): moves mining from integer math to FP8 floating point, because “the overwhelming majority of AI models are natively floating-point.” Its stated goal is to facilitate useful work with low overhead. It also ties mining to specific hardware (Hopper and Blackwell per the spec), which partly answers the “commodity integer math” critique. See Pearl’s FP8 fork explained and what FP8 is.
- Open community PRs #11 and #12: weight-provenance disclosure and weight-identity commitments. These aim at the core gap: proving that the matrices are real model weights. Neither is merged as of November 5, 2026.
So, is Pearl mining useful?
Our honest summary:
- In design: yes, it can be. The underlying research lets a GPU run real AI work and mine at the same time with small overhead.
- In early practice: the best public evidence (the Basu paper) says most hashpower was not doing useful work during the period it studied.
- Going forward: it depends on whether running real models becomes cheaper than mining random matrices. PIP-2 and PIP-3 push in that direction. Weight-provenance proposals would go further. None of that is settled yet.
Either way, the security story does not depend on the answer. Like Bitcoin, Pearl is secured by real hardware and energy, and that holds whether or not the matrices are useful. For how Pearl compares on that front, see Pearl vs Bitcoin.
What this means if you mine
Today you do not need to run an AI model to mine PEARL. Mining software and pools handle the work for you. Start with how to mine Pearl and the best GPUs for Pearl mining. The direction of the protocol is worth watching, though: if PIP-3 activates as written, only certain GPUs keep mining. Check whether your GPU will mine after the FP8 fork before buying or renting hardware.
Frequently asked questions
Does Pearl mining actually do useful AI work?
It can, but the protocol does not force it. Pearl's proof checks that a noised matrix multiplication was done correctly, not that the matrices came from a real AI model. A June 2026 arXiv paper by Abhinaba Basu argued that, at the time it studied, the network's hashpower produced no useful AI computation because the dominant mining software had no inference code.
Can Pearl miners just multiply random matrices?
According to the Basu paper, yes: the protocol accepts random matrices by design, and the author reports 44 pool-accepted shares from an open-source miner. Proposals to prove that matrices are real model weights (community PRs #11 and #12 in the PIPs repository) are open but not merged as of November 5, 2026.
If the work is useful, is Pearl easier to attack?
An economics paper by Rafael Pass (arXiv, June 2026) argues no: in its model the cost of attacking the chain stays tied to the block reward, whether or not miners also sell the computation.
What is Pearl doing to make mining more useful?
PIP-2 (Final) lets miners prove mixture-of-experts workloads efficiently, and PIP-3 (still a Draft) moves mining to FP8 floating point, which its authors say matches how most AI models actually run. Both are framed as ways to make real AI workloads cheaper to mine on.
