Pearl
Bitcoin's design with a different puzzle: miners multiply matrices, the same math that runs AI.
"Is anyone actually paying for the AI work Pearl miners do, or is it still mining with extra steps?"
Price
PEARL in USD, daily close from CoinGecko. Thin liquidity across a handful of small venues, so expect wicks.
| Venue | Pair | 24h vol |
|---|
Hashrate
How much matmul work miners point at Pearl. Difficulty chases price, so this is the market's real vote on miner margins.
Nobody measures hashrate directly; every site estimates it. Ours: hashrate = difficulty × 248 ÷ average block time, averaged over the last 50 blocks (about 2.5 hours), the same window prlscan shows. Over such a short window one fast or slow block moves it about 2%, so expect a jumpy line.
| Pool | Fee | Blocks · share |
|---|
Early days vs Bitcoin & Zcash
Hashrate growth from launch, each chain indexed to its own day 7 = 1×, on a log scale. Units differ (SHA-256, Equihash, matmul), so compare the curve, not the number.
Useful Work Watch
The chain can check that a miner multiplied matrices. It can't check that anyone needed the answer. This panel tracks who is doing work someone paid for.
How to read it: a retail pool member runs mining software for PEARL. A provider like Together AI mines inside its own inference stack and would most likely pay itself directly, so solo blocks are a ceiling on useful mining, not proof of it. Some solo miners are just hobbyists. Source: prlscan miner labels, refreshed hourly.
Wall of Pearl
The hottest Pearl posts on X: likes and reposts, weighted toward the last few days so fresh posts rise. Perle is filtered out, and the list is curated by hand on top.
How to expand Pearl analytics
Pearlascope runs on free, public data. Every panel here can get sharper with one more source:
- Run a node.
pearldexposesgetnetworkhashpsand block data over RPC. Pool splits and miner-software share come from the coinbase tags. - Tag useful work. If you serve paid inference on Pearl, send the endpoint and pricing. It goes on the watch list with a source link.
- Send a post. Found a sharp Pearl thread? Send the link and it goes on the wall.
- Open data. Everything shown is in
data.json. Pull it, chart it, prove us wrong.
DMs open at @dudubitcoin.
The PEARL Thesis
The best idea in crypto × AI this year, priced like a GPU coin. The whole bet is whether FP8 turns it into Bitcoin for AI before the dilution eats the holders.
FP8 slips or doesn't change the math. Gaming cards keep mining random matrices and dumping ~$1.4M a day. Their power bill (~$0.10–0.23 a coin) is the only floor: ~$0.25B.
The fork lands, the AI-server cost per coin drops to gaming-card levels, Coinbase lists it. Real AI blocks start to show up but stay a minority. Pearl trades at ~10% of Zcash: ~$2B.
A named provider passes mining income on as cheaper inference, rivals on OpenRouter have to mine to keep up, security follows. 1% of Bitcoin: ~$15B. Blue sky, compute settles in PEARL: $50B+.
Prices use the supply 12 months out (+82% from today), not fully diluted. Case triggers follow Blockworks Research's three steps ("Pearl: Bitcoin for AI?", Nov 2026). Blue sky from Brian Breslow's "Heresy". The base case is mine. I hold PEARL.
Two professors in Jerusalem rebuilt Bitcoin so the mining is the math AI runs on. The idea is real, the code ships every week, and the lab just raised $130M to prove it pays. Today almost none of the mining is useful, gaming cards do 82% of it, and holders eat 81% more supply over the next year. I'm holding through that because the fix has a name (FP8), it's in an open pull request updated this week, and the market is pricing it like it never lands.
An AI server that mines gives up paid AI work: that's its real cost. Today that’s roughly 10x a gaming card’s power bill, so AI servers don’t mine. FP8 (with the Pearl-GEMM rewrite) brings it down to the same range. Source: Blockworks Research.
No halvings: the reward shrinks a little every block (2.1B × H / (H + 650,226)). Year one is the worst year, ~297M new PEARL. New coins are worth of a day's trading, against about 0.1% for Bitcoin.
Live marks: CoinGecko. Zcash is the closest thing to a differentiated proof-of-work coin that made it. 1% of Bitcoin is the bull bar.
Read the full thesis
What Pearl is
Bitcoin miners burn about as much electricity as a mid-sized country guessing numbers. The guessing is the security. The guesses get thrown away.
Pearl keeps the Bitcoin part and swaps the puzzle. Miners multiply matrices, the operation that eats almost all of AI's compute. The chain mixes fresh random noise from the last block into both matrices, the GPU multiplies, small tiles of the result become lottery tickets for the next block, and the noise peels back off so an AI customer still gets the right answer. The winner ships a zero-knowledge proof, so a laptop can check the block without redoing an H100 job and nobody's prompt is exposed. Under the hood it's btcd: UTXOs, Taproot addresses starting prl1, a 2.1B cap.
Why I care
Every useful-work chain before this one picked between easy to check and useless, or useful and impossible to check. Pearl is the first serious attempt at both, from Omri Weinstein (matrix multiplication complexity, ex-Nvidia, Hebrew University) and Ilan Komargodski (the ZK proofs it runs on, same department). Between them their research covers both halves of the design. I'll admit the local pride: the most-talked-about AI coin of the year came out of a math department in Jerusalem.
The problem today: it's mostly a GPU lottery
The chain proves the math was done, not that anyone needed it. So the cheapest way to mine is a gaming card multiplying random numbers. Blockworks estimates fewer than 1% of blocks match a real AI model and consumer cards run ~82% of hashrate. Two reasons the real AI servers stay out:
- Format. Mining only counts INT7 math. Normal models don't run in INT7, so even Pearl's best rebuilt model can only mine on two-thirds of its math.
- Size. Only big batches fill a tile. Reading a long prompt mines; writing the reply one token at a time mostly doesn't. A summariser earns tickets on 25–40% of its math, a chatbot on ~5%. A dedicated miner: ~100%.
Net: an AI server gives up $1.33–3.15 of paid AI work for every coin it mines. A gaming card pays $0.10–0.23 in power. Nobody sane switches.
The fix: FP8
FP8 is the 8-bit format Nvidia's data-center chips already run AI in. PIP-3 moves mining to it, so standard open models (Llama, DeepSeek, partly the new MoE models) mine without being rebuilt, and the Pearl-GEMM rewrite cuts the slowdown on a serving GPU from 10–15% to 0–3%. That puts the AI server's cost per coin at $0.09–0.43, the same range as a gaming card. That's the whole game. Once serving inference is as cheap a way to mine as running a gaming card, every provider who doesn't mine is leaving money on the table.
The catch: the draft only accepts Nvidia Hopper and Blackwell. RTX 40 and 30 cards and every AMD card get cut off at the fork, and they do most of the mining today. Expect a hashrate air pocket on the day it activates. And it doesn't fix the size problem: chatbots still mine badly.
Who actually flips the switch
Mining lives inside the AI software, so it's whoever runs the software on the GPU that decides, not whoever owns the building. Bitcoin miners turned AI landlords mostly can't: their tenants run the chips. The natural first adopters are serverless providers who run both the GPUs and the software, then GPU marketplaces and neoclouds. That's a thinner pipeline than "every Bitcoin miner switches to Pearl". Which is why the lab raising $130M and selling inference itself on Liquid Inference matters: it's the cleanest test, run by the people best placed to pass it.
The bull loop
A provider that mines can pass the PEARL income on as a lower price. Routers like OpenRouter send traffic to the cheapest provider, so it wins volume. Rivals have to mine to keep up. More mining means more security, more security earns a store-of-value premium, a higher price funds deeper discounts. That's the reflexive version, and it's the one worth $15B+.
The bear case, taken seriously
Jacob from Blockworks ("The $PEARL Bear Case") makes four points. I agree with three of them.
- It isn't scarce yet. Supply grows ~81% this year against Bitcoin's ~0.8%. Bitcoin's early inflation happened at a few million dollars of market cap when nobody knew how to price it. Pearl is doing it at ~$0.5B. Fair. This is the cost of holding it now.
- Useful work doesn't create demand for the token. AI customers pay in dollars. A provider's discount is funded by selling PEARL, so more usage means more selling. Fair today. It changes only if compute gets priced or settled in PEARL, which is the lab's stated plan and nowhere near shipped.
- Cheap rentable GPUs make it attackable. Ethereum Classic, Bitcoin Gold and Vertcoin all got 51%'d by rented hardware. Pearl miners earn ~$1.4M a day against Bitcoin's ~$34M of subsidy. Kryptex alone wins close to half the blocks. Fair, and the concentration is the part I like least.
- Zcash 2016. ZEC fell 91% in four months while its market cap grew 8x, because supply grew 84x. The comparison is real but the scale isn't: Pearl's next-year growth is 81%, not 84x.
Where I disagree: the bear case prices FP8 at zero. It's an open pull request with merged pieces and a funded team whose own research is the exact thing being built. Zero is the wrong number.
The boring risks
- Every new number format is a hard fork that drops old hardware. Bitcoin hasn't changed its puzzle in 17 years; Pearl shipped four consensus forks between June and August.
- Someone builds a Pearl-only chip that out-mines GPUs without serving any AI. Nobody is known to be building one.
- The top 1% of addresses hold ~71% of supply. The lab hasn't disclosed what it mined in the low-difficulty week.
- Miners sell fast: about half of pool payouts go straight to one custodian's deposit addresses, and only ~10% is still in miners' wallets a week later.
- Liquidity sits on small exchanges, and the ticker collides with Perle.
Where I stand
I hold PEARL and I'm not selling here. I'm also not adding until FP8 gets an activation height or a real provider starts mining, because buying ahead of 81% dilution with no date is how you donate to miners. I'd rather be early and small than late and big. If FP8 merges and the AI-server cost per coin really lands at gaming-card levels, this goes from a GPU coin to the only working proof of useful work in crypto, at about 3% of Zcash's market cap.
What I'm watching
The chain, not the timeline. PR #311 merging with an activation height. The share of blocks from real models. The lab's prices on Liquid Inference against the other five providers. A Coinbase listing date. A holdings disclosure. Kryptex dropping under 50%. When the first verified paying customer on mining hardware shows up in Useful Work Watch, the case moves up a tier.
The Team
Pearl Research Labs. Researchers first, company second.
Mining profitability
Your share of the network times the coins minted per day, minus pool fee and power. Network hashrate, reward and price are live; everything else is yours to change.
Card presets and street prices: prlstats (median of whattomine, miningboard, hashrate.no; data-center cards measured). Sources disagree a lot: the RTX 4090 ranges 150–330 TH/s. Put in what your rig reports. Live inputs reset on reload.
Difficulty follows price: when PEARL pumps, more GPUs show up and your share shrinks. Emissions also drop a little every block. Treat this as today's snapshot, not a forecast.
Short answer: not to buy one. The cards that pay back fastest are RTX 40-series, and the FP8 fork as drafted cuts them off, so you're betting the fork stays more than a year away. The RTX 5090 survives FP8 but doesn't pay back at today's price. Renting always loses: an H100 rents for ~$39 a day and mines ~$14. If you already own a gaming card, mining is free money until the fork. Data-center cards (H100 and up) are left out of the chart: at $26k+ they never pay back from mining alone.
Getting PEARL
BigONE or Ourbit, PEARL/USDT, limit orders only. Small venues, so withdraw to Oyster afterwards.
WPRL on Uniswap via the third-party PearlBridge. It's an IOU on a bridge, not native PEARL.
An NVIDIA GPU, the Oyster wallet and a pool. Rent a card by the hour before you buy one.
FAQ
pearl-2), prlscan.com API (hashrate, pools), CoinMetrics Community (BTC/ZEC early hashrate), Pearl source code (emission curve from validate.go, 194s blocks), hand-curated rows with links.