What is Pearl (PRL)? Proof of useful work explained

Pearl is a Layer-1 blockchain whose coin, PRL, is mined on GPUs by multiplying matrices, the core operation of AI models, instead of by SHA-256 hashing alone as in Bitcoin. Pearl Research Labs launched it on April 27, 2026. This page explains how its proof of useful work operates, what it keeps from Bitcoin, who builds it, and how far the "useful" claim holds.

Pearl in one paragraph

Pearl keeps Bitcoin's design and changes the work. Its node, pearld, is a fork of btcd, a Go implementation of Bitcoin, so transactions and Taproot scripts work largely as in Bitcoin. But a miner wins a block by multiplying two noised matrices and proving the result with a zero-knowledge proof, which shows the computation is correct without revealing the matrices. If the matrices come from a real AI model, the same GPU work serves the model and secures the chain: that is the "useful" in proof of useful work. If they are random, the chain is still secured, nothing useful comes out, and the protocol cannot tell the difference.

• Genesis block: April 27, 2026, 09:00 UTC
• Target block time: 194 seconds (3 min 14 s)
• Maximum supply: 2,100,000,000 PRL, no halvings
• Block reward: 2,277.05 PRL at block ~124,157 (October 6, 2026)
• Supply mined: ~336.7M PRL, 16.0% of the cap (October 6, 2026)
• Smart contracts: none

How proof of useful work works

The scheme comes from the paper "Proofs of Useful Work from Arbitrary Matrix Multiplication" (arXiv 2504.09971) by Ilan Komargodski and Omri Weinstein. Bitcoin's hashing has no use beyond security, while matrix multiplication is most of the computing in AI models. The paper claims a "1+o(1)" overhead: as matrices grow, the extra cost of mining becomes negligible next to the multiplication itself. Simplified, a miner on mainnet does this:

1. Commit to the matrices. The miner picks matrices A and B, for example a model layer's weights and inputs, and fingerprints them with BLAKE3 hashes tied to the current block header.
2. Derive the noise. Seeds from those fingerprints generate random low-rank noise, so the work cannot be prepared in advance.
3. Multiply. The GPU computes (A + noise) × (B + noise) tile by tile, with integers between −64 and 64 summed into 32-bit totals.
4. Check each tile like a lottery ticket. Each tile's running state is hashed with BLAKE3. A hash at or below the network target wins the block.
5. Remove the noise. Cheap low-rank corrections cancel the noise terms and recover the real product A × B, the useful output.
6. Prove it. The winning tile is proven with a STARK wrapped in two Plonky2 layers, the last with zero knowledge. The proof is about 60 KB, needs no trusted setup and keeps the matrices private, so model weights never have to be published.

Pearl's 108-byte block header has no nonce. It carries a ProofCommitment hash, while the proof travels in a separate certificate of up to 65,000 bytes, which is why a one-transaction block weighs about 58.7 KB.

Pearl Research Labs publishes one reference miner, vllm-miner, in its official repository. It mines solo to your own node while serving a language model through vLLM. As of October 6, 2026 it supports only NVIDIA H100 and H200 GPUs. Consumer cards use third-party, mostly closed-source "pearlhash" miners such as SRBMiner-MULTI, BzMiner and WildRig Multi. See how to mine Pearl and hashrate by GPU.

Pearl vs Bitcoin: what stayed and what changed

Pearl is a new chain, not a split from Bitcoin: no Bitcoin history or balances were carried over. The UTXO model (coins tracked as unspent outputs), Taproot scripts, 8-decimal units and 100-block coinbase maturity stayed. Its genesis block embeds the hash of Bitcoin block 946,842 as a timestamp.

Data as of October 6, 2026. Sources: Pearl code at github.com/pearl-research-labs/pearl (node/chaincfg/params.go, node/blockchain/validate.go, node/blockchain/chain.go, node/wire/blockheader.go, node/txscript/standard.go); Bitcoin values are its standard consensus rules.

In practice, PRL can only be sent to addresses starting with prl1p (see the wallet guide), and the reward declines smoothly, from 3,229.64 PRL at block 1 to 807.41 PRL at block 650,226 (see tokenomics). Blocks have run faster than target: launch difficulty was 1, so the average since genesis is about 113 seconds, and the 24 hours before our October 6, 2026 snapshot averaged about 165 seconds.

Who builds Pearl

Pearl Research Labs is the brand. The legal entity named in its terms and privacy policy is Pearl Research Ltd, with terms under New York law. No registered address is published.

• Omri Weinstein, co-founder and CEO. B.Sc. from Tel Aviv University, PhD from Princeton, postdoc at NYU Courant, Columbia faculty from 2017. Now associate professor at the Hebrew University of Jerusalem, on leave from Columbia. He worked at NVIDIA and at VAST Data, where a conference profile lists him as chief scientist.
• Ilan Komargodski, co-founder. Cryptographer, associate professor at the Hebrew University, PhD from the Weizmann Institute. His title at Pearl is not published.

Itamar Schen co-authored early versions of the founding paper, and Rafael Pass (Cornell Tech, Technion) wrote a June 2026 economics paper while consulting for the company. The paper was posted on April 14, 2025 (also IACR ePrint 2025/685) and published in the ESA 2026 proceedings (LIPIcs vol. 388) as an invited-talk paper. Some articles simply call it "peer-reviewed"; we use the label the proceedings themselves give it.

Funding: no funding round or investor has been disclosed as of October 6, 2026. Venture-fund lists trace back to an X post about funds following @prlnet. The code has no premine, dev fee or treasury, but the team mines. Weinstein said the low launch difficulty was partly meant to give the core team resources, and Pearl's terms (August 5, 2026, §6.6) give the company the PRL mined while running customer inference. Team holdings are undisclosed.

Partners: Together AI and Attestable

Together AI, May 15, 2026. The AI cloud company Together AI announced an "exclusive partnership" and a Gemma-4-31B-it-pearl endpoint discounted by more than 25%, a discount it said is offset by the future value of crypto emissions. It is the only named third-party paying customer we found. On October 6, 2026, Together's model page said the model is not available on its serverless API, so the endpoint's current form is unclear.

Attestable, September 14, 2026. In the post that announced the FP8 scheme, @prlnet said it is partnering with Attestable, a zero-knowledge AI-integrity company. No press release or terms have been published.

Is the work actually useful?

A Pearl node checks that the multiplication is correct and the winning hash meets the target. It does not check where the matrices came from, and the reference mining function in Pearl's own repository uses random matrices. The protocol allows useful work; it does not require it.

The critique

"The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol" (arXiv 2606.04819) is a single-author preprint by Abhinaba Basu, marked "under review" (v1 June 3, v2 June 5, 2026). Its author:

• got 44 pool-accepted mainnet shares with random matrices (3 on an RTX 3090, 41 on an AMD MI300X);
• classified 8,012 AlphaPool workers (approximately, by hashrate) and found no inference-framework strings in the pool's closed-source alpha-miner, while calling his string analysis suggestive, not conclusive;
• estimated the late-May network (24 EH/s) at about 320,000 RTX 3090-class GPUs drawing about 112 MW for "zero useful AI computation".

Limits: it is not peer reviewed, it dates the launch to about March 2026 (block 1 is April 27), it gives two different return figures for the same card, and its code was not public on October 6, 2026. Hashrate may not be comparable across later forks, so its estimates should not be scaled to the ~59 to ~71 EH/s of October 6, 2026. Basu himself calls the acceptance of random matrices a design property, not a vulnerability.

Pearl's side

• Pearl's own INT whitepaper concedes that Pearl "attracts compute that does not necessarily have useful work" and proposes a future compute marketplace.
• The official miner only mines inside a vLLM model server, where the mining work is the model's own computation. Outside Pearl, the third-party BzMiner offers an optional mode that serves Pearl's pearl-ai models while mining on consumer cards.
• Pearl's network page says the Together endpoint is live as a serverless endpoint, "funded by the PRL mined on the same forward pass". On October 6, 2026, Together's own model page said the model is not available on its serverless API. Basu calls the discount financial arbitrage, with inference running separately.
• Rafael Pass's "The Economics of Proof-of-Useful-Work" (arXiv 2606.06700, June 4, 2026), cited in Pearl's FAQ, was written while consulting for Pearl Research Labs. It addresses 51% attack costs, not the random-matrix finding.

We found no direct rebuttal from Pearl as of October 6, 2026, and no one has published the share of hashrate doing paid AI work. The fair summary: useful work on Pearl is possible, not required, and unproven at scale.

Post-quantum features (opt-in)

A large quantum computer could break the elliptic-curve signatures Bitcoin and Pearl use by default. Pearl adds two defenses:

• XMSS signatures. The opcode OP_CHECKXMSSSIG verifies hash-based XMSS signatures in tapscript, the script path of a Taproot output. Signatures are 2,340 bytes, and one key can sign at most 32 times.
• P2MR outputs (BIP 360), a Taproot variant with no Schnorr key path. Consensus accepts them; addresses would start with prl1z, and none has been observed on chain.

The Oyster wallet adds an XMSS path only when asked, through the optional pq flag of getnewaddress, and the address still keeps its Schnorr key path. The Desktop Wallet creates plain Taproot addresses. So default PRL addresses are not quantum-safe. Pearl's April 2026 whitepaper called the opcode disabled; released node code executes it, but we found no evidence it has been used on mainnet.

Timeline

Data as of October 6, 2026, ~17:00 UTC. Sources: arXiv, github.com/pearl-research-labs/pearl, explorer.pearlresearch.ai, Together AI blog, x.com/prlnet, CoinGecko.

Next is FP8 (draft proposal PIP-3, open pull request #311), which would move mining to 8-bit floating point. As of October 6, 2026, its reference code covers H100/H200, B200/B300, RTX 50 and RTX PRO 6000, not RTX 30 or 40. An FP4 target of Q1 2027 comes only from an attendee's notes of a September 15 AMA. See hard forks.

Risks in brief

• Usefulness is not enforced, and the share of paid AI work is undisclosed.
• Front-loaded launch and dilution. The first 40,000 blocks minted ~36% of all PRL mined by October 6, 2026, and the team's share is undisclosed. At 194-second blocks, supply grows from ~336.7M to ~642.6M PRL in a year. See tokenomics.
• Frequent forks. Four in two months; the salted-seed release came ~90 minutes before activation. See hard forks.
• Possible GPU exclusion. As of October 6, 2026, the FP8 reference code (PR #311, no activation height) has no mining path for RTX 30 or RTX 40 cards. See hard forks.
• Pool concentration. Kryptex and PearlHash found ~75% to 80% of recent blocks (October 6, 2026). See pools.
• Thin markets. On October 6, 2026, CoinGecko listed native PRL only on SafeTrade and BigONE. We recommend no venue.
• Falling revenue. Gross revenue per TH/s fell ~85% in US dollars, from about $0.152 per day on May 29 to $0.0229 on October 6, 2026 (at $1.14 per PRL; difficulty-derived, before fees). See profitability.
• Look-alikes. Perle and fake "Pearl" tokens share the name or ticker. See scams.

As of October 6, 2026, Pearl has published no security advisories, and we found no reports of mainnet exploits or deep reorganizations. One consensus issue is open: GitHub issue #332 (September 22, 2026) argues that blocks around the rank-penalty fork height (96,251) may not pass today's rule when fully validated. Checkpoints up to block 110,000 hide this from new nodes, and a user reported a mainnet sync stuck at block 96,251 in August (issue #285). No incidents found is not proof of security.

Frequently asked questions