The short version
Pearl is a layer 1 blockchain launched by Pearl Research Labs on 27 April 2026. It keeps Bitcoin's ledger and monetary design but replaces hashing with matrix multiplication, the kind of arithmetic used by AI models.[1]
Its native coin, PRL, is issued through mining, with a cap of 2.1 billion. By mid-October, 16.4% of that supply had been issued. The launch was heavily front-loaded, with just over 35% of the coins issued by mid-October produced in the first six days.[2][1]
Pearl's proof-of-useful-work pitch is that the same GPUs securing its network can also serve paying AI customers. If miners can earn revenue from both activities, the network could make AI computation cheaper. But the protocol only checks that the multiplication happened. It can't tell whether anyone used the result. A June study found no useful AI output in the mining activity it examined.[3][4]
PRL perpetual futures began trading on a Hyperliquid market in mid-October. Among the traders we track, open long calls outnumber shorts 40 to 2.[5][6]
Origins, team, and design
Pearl's mainnet runs on a fork of btcd, a Golang-based Bitcoin implementation, and keeps Bitcoin's unspent transaction outputs, fee auction and most-work chain rule. Miners multiply matrices instead of computing SHA-256 hashes.[1]
The idea builds on a 2025 paper by Ilan Komargodski and Omri Weinstein, which argues that turning a multiplication into a mining proof adds almost no cost. The two later turned the research into Pearl Research Labs, bringing on Idan Sugerman as a third co-founder.[3] Weinstein, Pearl's chief executive, holds a Princeton doctorate and a faculty post at the Hebrew University of Jerusalem, with previous experience at Nvidia and VAST Data.[1] Komargodski is a cryptographer at the same university. Cornell cryptographer Rafael Pass, who wrote the network’s economic analysis, advises the project.[1][3]
Pearl also makes several changes outside the mining mechanism. It accepts only Taproot addresses, enables the OP_CAT opcode and embeds a dormant quantum-resistant backup key in wallets. Blocks timestamped more than five minutes into the future are rejected. The chain has no general smart contracts.[1]
Funding
The company has announced no venture round. Its legal structure, headcount and PRL holdings remain undisclosed.[1]
How mining works
A miner commits to two matrices, typically a model's weights and the data passing through it. The protocol combines that commitment with recent block data to generate random noise, so miners cannot precompute the work or pick matrices that make it trivially easy. The resulting product is divided into tiles. If a tile's hash falls below the network target, the miner wins the block. The noise can then be removed to recover the original result.[1][3]
Publishing the winning tile directly exposes rows of the miner’s matrices, hence, the proof is wrapped in a zero-knowledge certificate contained under 60 KB. Verifiers can check the proof without seeing the model weights or user data. Blocks target a 194-second interval, and difficulty adjusts with every block to meet this target.[1][7]
The commercial idea is called duplex mining. A GPU already serving paying inference customers could use the same computation to mine PRL, adding a second source of revenue without having to fund the entire workload through mining rewards.[3]
Pearl's official miner plugs into the vLLM inference engine and replaces its multiplication routine with Pearl's version. It officially supports Nvidia H100 and H200 data-centre cards, while community miners also run on consumer RTX GPUs.[1] Abhinaba Basu, author of the first systematic study of Pearl's live network (published on arXiv in June), found that the underlying integer arithmetic could also run on Nvidia and AMD GPUs, CPUs and Apple Silicon.[4]
Duplex mining only works if mining barely slows the paid workload. Pearl's own benchmarks, cited in a thesis by No Limit Holdings, put the end-to-end overhead at 5.08% for Llama 70B on four H200s and 3.9% for DeepSeek V3.2 on eight H200s, using integer quantisation.[3]
A floating-point specification, FP8, was published in September to bring the approach closer to the formats used by production models.[8] The mainnet parameters, however, do not specify an activation height for FP8.[7] So the only overhead figures available cover integer-quantised models, and Pearl published them itself.
$PRL
Every PRL enters circulation through block rewards. There was no token sale or founder allocation.[1] The issuance formula sets cumulative supply after h blocks at 2.1 billion multiplied by h/(h + 650,226). Half the maximum supply is therefore issued by block 650,226, roughly four years at the target block pace.[9]
At block 127,295 in mid-October, 343.8 million PRL had been issued, equal to 16.4% of the cap.[2] At the target pace, the following year would add about 304 million coins, close to doubling the supply issued so far.[9][7]
Difficulty began near its minimum, so block 40,000 arrived six days after launch instead of the roughly 90 days the target pace implies. Those early blocks produced approximately 121.7 million PRL, just over 35% of the 343.8 million issued by mid-October.[1][9]
The whitepaper's wording suggests the team mined during that period, but its holdings have not been disclosed.[1] CoinDesk reports a circulating supply of 269.2 million, below the total derived from the code. We couldn't find an explanation for the gap.[10][1]
Miners receive PRL as block rewards, and holders use it for transfers and transaction fees. Operators can also sell coins earned while serving models to subsidise inference costs. The token has no governance or staking rights. A compute marketplace settled on-chain appears in the whitepaper, but it is not live.[1]
Network
Together AI, is one public example of the commercial model. In May, it launched a Pearl-powered inference endpoint under an exclusive partnership. The endpoint was discounted by more than 25%, with the expected value of PRL emissions helping offset the discount.[11] Pearl Research Labs also operates an inference platform listing eight open models, though it has not disclosed what proportion of that traffic is mined.[1]
Network hashrate has roughly doubled since June. Basu estimated it at 24 EH/s during his study, while Kryptex reported 53.26 EH/s on 13 October.[4][12] In a 24-hour snapshot at block 127,295, two pools, Kryptex (58%) and PearlHash (23%), produced four of every five blocks.[2]
The protocol has also undergone four mining-rule changes. Mainnet parameters record a mixture-of-experts (MoE) fork at block 71,935 in June, followed by a fork at block 91,630 that stopped accepting MoE proofs. Two more forks followed in August.[7][1] Pearl's code comments do not explain why MoE support was reversed. No Limit Holdings' thesis says miners had exploited the upgrade to capture disproportionate rewards.[1][3]
One August fork added a rank penalty. Before it, miners could pick a higher noise rank and clear the winning threshold more easily for the same work.[13]
What our data shows
Until mid-October, PRL traded only on smaller spot exchanges such as BigONE and CoinEx.[1] In mid-October, perpetual futures launched on a Hyperliquid HIP-3 market deployed by Entropy, which had won the PRL ticker in an auction on 26 September.[5][14]

As of 19 October, open long calls on Pearl among the traders we track outnumbered shorts 40 to 2. Over the preceding 60 days, 46 of 48 tracked calls were long, and 29 of the 48 were opened in September and 17 in October. Of the 20 calls with a classified thesis, 19 cited a narrative instead of a specific event.[6]
All ten market events we logged in that period originated from posts on X. Half concerned new trading venues, and none reported fresh measurements of useful work.[5] Pearl also did not rank among the top 50 crypto assets by share of attention in our data.[15]
Can Pearl make useful work pay?
Supporters and critics agree on the starting point. Pearl's protocol checks whether a multiplication satisfies its proof requirements; it does not establish whether the result serves a paying customer. Useful AI computation is therefore not required by the network today.[3][4] The disagreement is about incentives.
Supporters, including No Limit Holdings, argue that a miner running paid inference has an advantage because customers already cover much of the GPU cost. A miner working only for PRL must fund the entire operation through block rewards. As mining difficulty rises, the argument goes, pure miners become less competitive and eventually leave, allowing useful computation to account for a greater share of the network.[3]
Rafael Pass, who consulted for Pearl Research Labs on this work, finds in his model that the cost of a majority attack remains tied to the block reward even when mining work has value outside the chain. Rewards could effectively subsidise inference and lower the price customers pay.[3][16]
Pass describes three stages of development: Bitconia, Fortessia and Duplexia. No Limit Holdings places Pearl in Fortessia, where commercial duplex mining exists but has yet to dominate the network.[3]
Basu examined 8,012 workers in one mining pool, representing about 21% of the network, and found no inference code in the dominant mining software.[1][4] His own miner had 44 shares accepted by a pool, using random matrices unrelated to any AI model. He also demonstrated that adversarial sampling could defeat the protocol's statistical checks.[4]
Based on his analysis, Basu estimated that in June the network used about 320,000 GPU-equivalents and drew 112 MW, with zero useful AI output.[4] His argument: if skipping the AI work makes mining cheaper, miners will skip it.[1] Pearl's FP8 specification acknowledges that miners could gain speedups by dropping the usefulness requirement, and describes statistical checks intended to limit that advantage.[8]
No Limit's own thesis lists what has to happen for useful mining to win out: overhead keeps falling, useful miners become the marginal producers, and the rule changes stop.[3] The evidence so far doesn't show all three.
The number that would settle this is the share of Pearl's compute that serves paying customers. Pearl doesn't publish it, and the only outside measurement we found, Basu's June study of one pool, came before the network's hashrate doubled.
Conclusion
Pearl's mechanism works at scale. The commercial case for it does not yet have independent data behind it. Duplex mining currently rests on Pass's incentive model and the team's own overhead benchmarks, and the one outside measurement points the other way.[3][4] Traders are positioned as if that question were already answered.
Three developments would change the picture: Pearl or a major pool publishing the share of mined compute tied to paid inference, an FP8 activation height in the mainnet parameters, and an extended period without further changes to the mining rules. Until at least one of them arrives, PRL trades on the narrative rather than on measured usefulness.
Want to integrate Elfa into your workflow? Get started with the Elfa API or contact us.
Disclaimer
This research is provided for informational and educational purposes only and does not constitute financial, investment, legal, or other professional advice, or a recommendation to buy, sell, or hold any security or financial instrument.
This research was prepared using Elfa alongside other research tools, data sources, and publicly available information. Elfa does not hold any position in the assets or securities discussed and has no financial interest in the subject of this report.
The information and analysis presented are based on sources believed to be reliable but may contain errors, omissions, estimates, or forward-looking statements. No representation or warranty is made as to the accuracy or completeness of the information.
Readers should conduct their own independent research and consult a qualified financial professional before making any investment decision. Investing involves risk, including the possible loss of principal.
Sources
- Pearl Explained: Proof of Useful Work, AI Mining and PRL Tokenomics, Datawallet, 3 Oct 2026
- Pearl network mining share, 24-hour snapshot at block 127,295, prlscan, accessed 14 Oct 2026
- Heterodoxy: Pearl is reviving Bitcoin, investor thesis by No Limit Holdings compiled by ChainCatcher, 10 Oct 2026
- The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol, Abhinaba Basu, arXiv, Jun 2026
- Market events for Pearl, 20 Aug to 19 Oct 2026, Elfa, 19 Oct 2026
- Tracked trader positioning and call lifecycles on Pearl, 20 Aug to 19 Oct 2026, Elfa, 19 Oct 2026
- Pearl node source code, chaincfg/params.go (mainnet parameters), Pearl Research Labs, GitHub, accessed 19 Oct 2026
- Pearl Floating Point Scheme Specification (whitepaper), Pearl Research Labs, Sep 2026
- Pearl node source code, blockchain/validate.go (issuance formula), Pearl Research Labs, GitHub, accessed 19 Oct 2026
- Pearl (PRL) asset page, CoinDesk, 9 Oct 2026
- Together AI and Pearl Research Labs Team Up to Reduce the Cost of AI Inference, Together AI, 15 May 2026
- About PRL (network hashrate), Kryptex Pool, updated 13 Oct 2026
- Rank-penalty softfork commit (#275), Pearl Research Labs, GitHub, 5 Aug 2026
- Entropy Spends 500 HYPE on Pearl Code, TokenPost, Sep 2026
- PRL entity resolution and crypto attention rankings, Elfa, 10 to 19 Oct 2026
- The Economics of Proof-of-Useful-Work, Rafael Pass, arXiv, 8 Jun 2026




