The reviewed record of science sign in
Pith

arxiv: 2606.04819 · v2 · pith:6EVTE5FW · submitted 2026-06-03 · cs.CR · cs.CY· cs.DC

The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol

Abhinaba Basu This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-28 05:51 UTCgrok-4.3pith:6EVTE5FWrecord.jsonopen to challenge →

classification cs.CR cs.CYcs.DC
keywords Proof-of-Useful-WorkPearl blockchainAI inferenceempirical measurementblockchain verificationGPU miningmining economicsusefulness gap
0
0 comments X

The pith

Pearl's cuPOW protocol produces zero useful AI computation on its 24 EH/s network.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper measures a deployed Proof-of-Useful-Work blockchain that claims to run AI inference while securing the chain. It finds the entire network of roughly 320,000 GPU-equivalents performs no inference at all. The dominant mining software contains no inference code, yet the verification step accepts random matrices as valid shares across every hardware type tested. GPU rental prices rose 38 percent and utilization climbed from 57 to 94 percent after the software appeared, crowding out other workloads. The actual work reduces to ordinary integer arithmetic that runs on any platform and yields marginal or negative returns for most miners at current token prices.

Core claim

Pearl's cuPOW protocol, presented as simultaneously securing the network and executing useful AI inference, produces zero useful AI computation. Network analysis of 8,012 workers shows inference-capable hardware throughout, but the leading mining client includes no inference routines. The verification protocol accepts random matrices by design, a fact confirmed by 44 accepted shares from an open-source miner on NVIDIA, AMD, CPU, and Apple Silicon devices. Statistical distribution tests are defeated by simple adversarial sampling, and the arithmetic performed is portable commodity integer work with no hardware specificity.

What carries the argument

The verification protocol that accepts random matrices as valid proof-of-work shares.

If this is right

  • GPU rental prices increased 38 percent and utilization reached 94 percent after the mining software release.
  • Mining returns range from negative 1 percent to positive 67 percent depending on GPU tier at current token prices.
  • The performed computation consists of commodity integer arithmetic runnable on any hardware platform.
  • The network consumes an estimated 112 MW while delivering no usable AI results.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Verification that cannot distinguish random inputs from real inference outputs leaves any PoUW system open to the same gap.
  • Portable integer work removes any incentive for miners to acquire specialized AI hardware.
  • Crowding of rental markets suggests PoUW mining can displace research workloads even when token economics are marginal.
  • Protocols relying on statistical checks alone remain vulnerable to adversarial sampling that passes verification without performing the claimed task.

Load-bearing premise

The dominant mining software examined represents the majority of network participants, and acceptance of random matrices by the verification protocol demonstrates absence of actual inference computation.

What would settle it

Discovery of inference code or actual trained-model outputs inside shares submitted by the dominant mining software would falsify the zero-useful-work finding.

Figures

Figures reproduced from arXiv: 2606.04819 by Abhinaba Basu.

Figure 1
Figure 1. Figure 1: Cross-platform Pearl mining results. (a) Mining throughput across hardware platforms. Our open-source AMD MI300X miner (red) achieves 3.2× the throughput of the closed-source alpha-miner on RTX 3090 (hatched). (b) Per-iteration time breakdown on MI300X showing the full Triton-based GPU pipeline. (c) Mining profitability: revenue is 1–2 orders of magnitude below hardware cost, yielding −72% (RTX 3090) to −9… view at source ↗
read the original abstract

Pearl, a Layer-1 blockchain with high-profile AI industry endorsements, markets its Proof-of-Useful-Work (PoUW) protocol as simultaneously securing the network and performing AI inference. We present the first systematic empirical measurement of a deployed PoUW system, finding that Pearl's 24 EH/s network -- representing approximately 320,000 GPU-equivalents consuming an estimated 112 MW -- produces zero useful AI computation. Budget GPU rental prices rose 38% and utilization surged from 57% to 94% following the mining software's public release, displacing legitimate research workloads. Our measurements span five dimensions: (1) network composition analysis of 8,012 workers shows all have inference-capable hardware, yet the dominant mining software contains no inference code; (2) the verification protocol accepts random matrices by design, confirmed by 44 pool-accepted shares from our open-source miner across NVIDIA, AMD, CPU, and Apple Silicon hardware; (3) statistical distribution checks are trivially defeated by adversarial Gaussian sampling; (4) mining economics are marginal at current PRL prices ($0.76), with ROI ranging from -1% to +67% depending on GPU tier -- near breakeven for most hardware; and (5) the mining computation is commodity integer arithmetic portable to any hardware platform, offering no vendor lock-in. These findings quantify the verifiability-usefulness tension identified theoretically by Leinweber et al., providing concrete measurements of its magnitude and economic consequences in a deployed system.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper claims that Pearl's 24 EH/s PoUW network produces zero useful AI computation. This is based on empirical measurements across five dimensions: network composition of 8,012 workers with inference-capable hardware but dominant software lacking inference code; verification protocol accepting random matrices confirmed by 44 shares; statistical checks defeated by adversarial sampling; marginal ROI for mining; and portable commodity computation. It also reports a 38% rise in GPU rental prices and utilization surge from 57% to 94% after mining software release, displacing research workloads.

Significance. If the central claim holds, this manuscript provides valuable empirical evidence quantifying the usefulness gap in deployed Proof-of-Useful-Work systems. It offers direct measurements of network behavior, protocol verifiability, and economic impacts on GPU markets, grounding theoretical discussions from Leinweber et al. with concrete data from a high-profile blockchain. The multi-platform testing and economic analysis are particular strengths.

major comments (1)
  1. [network composition analysis] The assertion that the network produces zero useful AI computation hinges on the dominant mining software being representative of the majority of the 24 EH/s hash rate. The analysis is limited to 8,012 workers, with no reported data on the hash-rate fraction these workers constitute. Given that the verification protocol accepts random matrices by design, this leaves open the possibility that other software used by unexamined hash rate performs actual inference while submitting valid shares, undermining the zero-useful-computation conclusion.
minor comments (1)
  1. [Abstract] Specific numerical claims (e.g., 38% price increase, 112 MW consumption) would benefit from explicit cross-references to the sections or tables detailing the underlying data collection and calculations.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for their careful review and for highlighting an important methodological point regarding our network composition analysis. We address this comment directly below.

read point-by-point responses

  1. Referee: The assertion that the network produces zero useful AI computation hinges on the dominant mining software being representative of the majority of the 24 EH/s hash rate. The analysis is limited to 8,012 workers, with no reported data on the hash-rate fraction these workers constitute. Given that the verification protocol accepts random matrices by design, this leaves open the possibility that other software used by unexamined hash rate performs actual inference while submitting valid shares, undermining the zero-useful-computation conclusion.

    Authors: Our analysis sampled 8,012 active workers from the Pearl network; all possessed inference-capable hardware yet executed mining software containing no inference code. We acknowledge that we did not collect per-worker hash-rate data and therefore cannot report the exact fraction of the 24 EH/s represented by this sample. However, these workers were drawn from the publicly observable set of participants, and the identified software was the only publicly released client at the time. More critically, our 44 pool-accepted shares using random matrices confirm that the verification protocol accepts non-inference inputs by design. Consequently, even if unexamined hash rate employs alternative software that optionally performs inference, the protocol neither requires nor verifies such computation. The zero-useful-computation conclusion therefore rests on the absence of enforced usefulness rather than solely on software prevalence. We will add an explicit discussion of sampling limitations and this protocol-level point in the revised manuscript. revision: partial

standing simulated objections not resolved
  • Precise hash-rate fraction represented by the 8,012 sampled workers (data not collected during measurements)

Circularity Check

0 steps flagged

Empirical measurements with no derivation chain or self-referential reduction

full rationale

The paper is an empirical study relying on direct observations of mining software code, accepted shares from custom miners, network worker counts, and economic calculations from public prices. No mathematical derivation, fitted parameters renamed as predictions, or first-principles claims appear in the abstract or described structure. The central finding of zero useful computation follows from software inspection and protocol behavior tests rather than any equation or ansatz that reduces to its own inputs. Self-citations are absent from the provided text, and external citations (e.g., Leinweber et al.) are not load-bearing for the measurements. The work is self-contained against external benchmarks such as code review and share verification.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is an empirical measurement study and introduces no free parameters, axioms, or invented entities; its claims rest on direct observation of deployed software and network behavior.

pith-pipeline@v0.9.1-grok · 5812 in / 945 out tokens · 40310 ms · 2026-06-28T05:51:53.021285+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

  1. [1]

    Miller, A

    A. Miller, A. Kosba, J. Katz, and E. Shi. Nonoutsourceable scratch-off puzzles to discourage bitcoin mining coalitions. InProc. ACM CCS, 2015

    2015

  2. [2]

    Hoffmann

    F. Hoffmann. Challenges of proof-of-useful-work (PoUW).arXiv:2209.03865, 2022

    work page arXiv 2022

  3. [3]

    Pearl: A layer-1 blockchain with proof of useful work.https: //pearlresearch.ai, 2025

    Pearl Research Labs. Pearl: A layer-1 blockchain with proof of useful work.https: //pearlresearch.ai, 2025

    2025

  4. [4]

    Proofs of useful work from arbitrary matrix multipli- cation.CoRR, abs/2504.09971, 2025

    I. Komargodski and O. Weinstein. Proofs of useful work from arbitrary matrix multiplication. arXiv:2504.09971, 2025

    work page arXiv 2025

  5. [5]

    A. Lihu, J. Du, I. Barjaktarevic, P. Gerzanics, and M. Harvilla. A proof of useful work for artificial intelligence on the blockchain.arXiv:2001.09244, 2020

    work page arXiv 2001

  6. [6]

    Baldominos and Y

    A. Baldominos and Y. Saez. Coin.AI: A proof-of-useful-work scheme for blockchain-based distributed deep learning.Entropy, 2019

    2019

  7. [7]

    M. Ball, A. Rosen, M. Sabin, and P. N. Vasudevan. Proofs of useful work. IACR ePrint 2017/203, 2017

    2017

  8. [8]

    alpha-miner: GPU miner for the Pearl (PRL) network.https://github

    AlphaMine Tech. alpha-miner: GPU miner for the Pearl (PRL) network.https://github. com/AlphaMine-Tech/alpha-miner, 2026

    2026

  9. [9]

    GPU rental marketplace and pricing.https://vast.ai/pricing, 2026

    vast.ai. GPU rental marketplace and pricing.https://vast.ai/pricing, 2026

    2026

  10. [10]

    H. Jia, M. Yaghini, C. A. Choquette-Choo, N. Dullerud, A. Thudi, V. Chandrasekaran, and N. Papernot. Proof-of-learning: Definitions and practice. InIEEE S&P, 2021

    2021

  11. [11]

    D. Kang, T. Hashimoto, I. Stoica, and Y. Sun. Scaling up trustless DNN inference with zero-knowledge proofs.arXiv:2210.08674, 2022

    work page arXiv 2022

  12. [12]

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica. Efficient memory management for large language model serving with PagedAttention. InProc. SOSP, 2023

    2023

  13. [13]

    Schrijvers, J

    O. Schrijvers, J. Bonneau, D. Boneh, and T. Roughgarden. Incentive compatibility of Bitcoin mining pool reward functions. InFinancial Cryptography, 2016

    2016

  14. [14]

    Frantar, S

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh. GPTQ: Accurate post-training quanti- zation for generative pre-trained transformers. InICLR, 2023

    2023

  15. [15]

    Dettmers, M

    T. Dettmers, M. Lewis, Y. Belkada, and L. Zettlemoyer. LLM.int8(): 8-bit matrix multiplica- tion for transformers at scale. InNeurIPS, 2022

    2022

  16. [16]

    Together AI partners with Pearl Research Labs.https://www.together.ai/ blog/together-ai-partners-with-pearl-research-labs, 2026

    Together AI. Together AI partners with Pearl Research Labs.https://www.together.ai/ blog/together-ai-partners-with-pearl-research-labs, 2026. 15

    2026

  17. [17]

    GPU mining Pearl:$200/day with proof of useful work

    BearMarketMiner. GPU mining Pearl:$200/day with proof of useful work. YouTube,https: //www.youtube.com/watch?v=zNhUFdy1rQM, 2026

    2026

  18. [18]

    Will PEARL mining make a comeback? YouTube livestream,https: //www.youtube.com/watch?v=YJhZ2lVXiKc, May 31, 2026

    Red Panda Mining. Will PEARL mining make a comeback? YouTube livestream,https: //www.youtube.com/watch?v=YJhZ2lVXiKc, May 31, 2026

    2026

  19. [19]

    Philippe Tillet, H. T. Kung, and David Cox. Triton: An intermediate language and compiler for tiled neural network computations. InProc. ACM SIGPLAN Workshop on Machine Learning and Programming Languages (MAPL), 2019

    2019

  20. [20]

    Filecoin energy dashboard.https://filecoin.energy/, 2026

    Filecoin Green. Filecoin energy dashboard.https://filecoin.energy/, 2026. Network data: 3 EiB committed capacity, 36% utilization, Q3 2025

    2026

  21. [21]

    Dikshit, A

    P. Dikshit, A. Emami, J. Sedlmeir, and G. Fridgen. SoK: Is proof-of-useful-work really useful? IACR ePrint 2025/1814, 2025. 16

    2025